2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
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:
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
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
33 #include <sys/types.h>
36 #include <sys/mkdev.h>
38 #ifdef MAJOR_IN_SYSMACROS
39 #include <sys/sysmacros.h>
43 #include <sys/inotify.h>
46 #include "util/debug.h"
47 #include "util/disk_cache.h"
49 #include "radv_debug.h"
50 #include "radv_private.h"
51 #include "radv_shader.h"
54 typedef void *drmDevicePtr;
59 #include "drm-uapi/amdgpu_drm.h"
60 #include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
62 #include "util/build_id.h"
63 #include "util/debug.h"
64 #include "util/driconf.h"
65 #include "util/mesa-sha1.h"
66 #include "util/os_time.h"
67 #include "util/timespec.h"
68 #include "util/u_atomic.h"
69 #include "winsys/null/radv_null_winsys_public.h"
72 #include "vk_format.h"
74 #include "vk_sync_dummy.h"
75 #include "vulkan/vk_icd.h"
78 #include "ac_llvm_util.h"
81 /* The number of IBs per submit isn't infinite, it depends on the IP type
82 * (ie. some initial setup needed for a submit) and the number of IBs (4 DW).
83 * This limit is arbitrary but should be safe for now. Ideally, we should get
84 * this limit from the KMD.
86 #define RADV_MAX_IBS_PER_SUBMIT 192
88 /* The "RAW" clocks on Linux are called "FAST" on FreeBSD */
89 #if !defined(CLOCK_MONOTONIC_RAW) && defined(CLOCK_MONOTONIC_FAST)
90 #define CLOCK_MONOTONIC_RAW CLOCK_MONOTONIC_FAST
93 static VkResult radv_queue_submit(struct vk_queue *vqueue, struct vk_queue_submit *submission);
96 radv_get_current_time(void)
98 return os_time_get_nano();
102 radv_device_get_cache_uuid(struct radv_physical_device *pdevice, void *uuid)
104 enum radeon_family family = pdevice->rad_info.family;
105 struct mesa_sha1 ctx;
106 unsigned char sha1[20];
107 unsigned ptr_size = sizeof(void *);
109 memset(uuid, 0, VK_UUID_SIZE);
110 _mesa_sha1_init(&ctx);
112 if (!disk_cache_get_function_identifier(radv_device_get_cache_uuid, &ctx)
113 #ifdef LLVM_AVAILABLE
114 || (pdevice->use_llvm &&
115 !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo, &ctx))
120 _mesa_sha1_update(&ctx, &family, sizeof(family));
121 _mesa_sha1_update(&ctx, &ptr_size, sizeof(ptr_size));
122 _mesa_sha1_final(&ctx, sha1);
124 memcpy(uuid, sha1, VK_UUID_SIZE);
129 radv_get_driver_uuid(void *uuid)
131 ac_compute_driver_uuid(uuid, VK_UUID_SIZE);
135 radv_get_device_uuid(struct radeon_info *info, void *uuid)
137 ac_compute_device_uuid(info, uuid, VK_UUID_SIZE);
141 radv_get_adjusted_vram_size(struct radv_physical_device *device)
143 int ov = driQueryOptioni(&device->instance->dri_options, "override_vram_size");
145 return MIN2(device->rad_info.vram_size, (uint64_t)ov << 20);
146 return device->rad_info.vram_size;
150 radv_get_visible_vram_size(struct radv_physical_device *device)
152 return MIN2(radv_get_adjusted_vram_size(device), device->rad_info.vram_vis_size);
156 radv_get_vram_size(struct radv_physical_device *device)
158 uint64_t total_size = radv_get_adjusted_vram_size(device);
159 return total_size - MIN2(total_size, device->rad_info.vram_vis_size);
163 RADV_HEAP_VRAM = 1 << 0,
164 RADV_HEAP_GTT = 1 << 1,
165 RADV_HEAP_VRAM_VIS = 1 << 2,
166 RADV_HEAP_MAX = 1 << 3,
170 radv_physical_device_init_mem_types(struct radv_physical_device *device)
172 uint64_t visible_vram_size = radv_get_visible_vram_size(device);
173 uint64_t vram_size = radv_get_vram_size(device);
174 uint64_t gtt_size = device->rad_info.gart_size;
175 int vram_index = -1, visible_vram_index = -1, gart_index = -1;
177 device->memory_properties.memoryHeapCount = 0;
180 if (!device->rad_info.has_dedicated_vram) {
181 /* On APUs, the carveout is usually too small for games that request a minimum VRAM size
182 * greater than it. To workaround this, we compute the total available memory size (GTT +
183 * visible VRAM size) and report 2/3 as VRAM and 1/3 as GTT.
185 const uint64_t total_size = gtt_size + visible_vram_size;
186 visible_vram_size = align64((total_size * 2) / 3, device->rad_info.gart_page_size);
187 gtt_size = total_size - visible_vram_size;
191 /* Only get a VRAM heap if it is significant, not if it is a 16 MiB
192 * remainder above visible VRAM. */
193 if (vram_size > 0 && vram_size * 9 >= visible_vram_size) {
194 vram_index = device->memory_properties.memoryHeapCount++;
195 device->heaps |= RADV_HEAP_VRAM;
196 device->memory_properties.memoryHeaps[vram_index] = (VkMemoryHeap){
198 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
203 gart_index = device->memory_properties.memoryHeapCount++;
204 device->heaps |= RADV_HEAP_GTT;
205 device->memory_properties.memoryHeaps[gart_index] = (VkMemoryHeap){
211 if (visible_vram_size) {
212 visible_vram_index = device->memory_properties.memoryHeapCount++;
213 device->heaps |= RADV_HEAP_VRAM_VIS;
214 device->memory_properties.memoryHeaps[visible_vram_index] = (VkMemoryHeap){
215 .size = visible_vram_size,
216 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
220 unsigned type_count = 0;
222 if (vram_index >= 0 || visible_vram_index >= 0) {
223 device->memory_domains[type_count] = RADEON_DOMAIN_VRAM;
224 device->memory_flags[type_count] = RADEON_FLAG_NO_CPU_ACCESS;
225 device->memory_properties.memoryTypes[type_count++] = (VkMemoryType){
226 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
227 .heapIndex = vram_index >= 0 ? vram_index : visible_vram_index,
231 if (gart_index >= 0) {
232 device->memory_domains[type_count] = RADEON_DOMAIN_GTT;
233 device->memory_flags[type_count] = RADEON_FLAG_GTT_WC | RADEON_FLAG_CPU_ACCESS;
234 device->memory_properties.memoryTypes[type_count++] = (VkMemoryType){
236 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
237 .heapIndex = gart_index,
240 if (visible_vram_index >= 0) {
241 device->memory_domains[type_count] = RADEON_DOMAIN_VRAM;
242 device->memory_flags[type_count] = RADEON_FLAG_CPU_ACCESS;
243 device->memory_properties.memoryTypes[type_count++] = (VkMemoryType){
244 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
245 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
246 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
247 .heapIndex = visible_vram_index,
251 if (gart_index >= 0) {
252 device->memory_domains[type_count] = RADEON_DOMAIN_GTT;
253 device->memory_flags[type_count] = RADEON_FLAG_CPU_ACCESS;
254 device->memory_properties.memoryTypes[type_count++] = (VkMemoryType){
255 .propertyFlags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
256 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT | VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
257 .heapIndex = gart_index,
260 device->memory_properties.memoryTypeCount = type_count;
262 if (device->rad_info.has_l2_uncached) {
263 for (int i = 0; i < device->memory_properties.memoryTypeCount; i++) {
264 VkMemoryType mem_type = device->memory_properties.memoryTypes[i];
266 if ((mem_type.propertyFlags &
267 (VK_MEMORY_PROPERTY_HOST_COHERENT_BIT | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)) ||
268 mem_type.propertyFlags == VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT) {
270 VkMemoryPropertyFlags property_flags = mem_type.propertyFlags |
271 VK_MEMORY_PROPERTY_DEVICE_COHERENT_BIT_AMD |
272 VK_MEMORY_PROPERTY_DEVICE_UNCACHED_BIT_AMD;
274 device->memory_domains[type_count] = device->memory_domains[i];
275 device->memory_flags[type_count] = device->memory_flags[i] | RADEON_FLAG_VA_UNCACHED;
276 device->memory_properties.memoryTypes[type_count++] = (VkMemoryType){
277 .propertyFlags = property_flags,
278 .heapIndex = mem_type.heapIndex,
282 device->memory_properties.memoryTypeCount = type_count;
287 radv_get_compiler_string(struct radv_physical_device *pdevice)
289 if (!pdevice->use_llvm) {
290 /* Some games like SotTR apply shader workarounds if the LLVM
291 * version is too old or if the LLVM version string is
292 * missing. This gives 2-5% performance with SotTR and ACO.
294 if (driQueryOptionb(&pdevice->instance->dri_options, "radv_report_llvm9_version_string")) {
295 return " (LLVM 9.0.1)";
301 #ifdef LLVM_AVAILABLE
302 return " (LLVM " MESA_LLVM_VERSION_STRING ")";
304 unreachable("LLVM is not available");
309 radv_get_int_debug_option(const char *name, int default_value)
316 result = default_value;
320 result = strtol(str, &endptr, 0);
322 /* No digits founs. */
323 result = default_value;
331 radv_thread_trace_enabled()
333 return radv_get_int_debug_option("RADV_THREAD_TRACE", -1) >= 0 ||
334 getenv("RADV_THREAD_TRACE_TRIGGER");
338 radv_spm_trace_enabled()
340 return radv_thread_trace_enabled() &&
341 debug_get_bool_option("RADV_THREAD_TRACE_CACHE_COUNTERS", false);
344 #if defined(VK_USE_PLATFORM_WAYLAND_KHR) || defined(VK_USE_PLATFORM_XCB_KHR) || \
345 defined(VK_USE_PLATFORM_XLIB_KHR) || defined(VK_USE_PLATFORM_DISPLAY_KHR)
346 #define RADV_USE_WSI_PLATFORM
350 #define RADV_API_VERSION VK_MAKE_VERSION(1, 1, VK_HEADER_VERSION)
352 #define RADV_API_VERSION VK_MAKE_VERSION(1, 3, VK_HEADER_VERSION)
355 VKAPI_ATTR VkResult VKAPI_CALL
356 radv_EnumerateInstanceVersion(uint32_t *pApiVersion)
358 *pApiVersion = RADV_API_VERSION;
362 static const struct vk_instance_extension_table radv_instance_extensions_supported = {
363 .KHR_device_group_creation = true,
364 .KHR_external_fence_capabilities = true,
365 .KHR_external_memory_capabilities = true,
366 .KHR_external_semaphore_capabilities = true,
367 .KHR_get_physical_device_properties2 = true,
368 .EXT_debug_report = true,
369 .EXT_debug_utils = true,
371 #ifdef RADV_USE_WSI_PLATFORM
372 .KHR_get_surface_capabilities2 = true,
374 .KHR_surface_protected_capabilities = true,
376 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
377 .KHR_wayland_surface = true,
379 #ifdef VK_USE_PLATFORM_XCB_KHR
380 .KHR_xcb_surface = true,
382 #ifdef VK_USE_PLATFORM_XLIB_KHR
383 .KHR_xlib_surface = true,
385 #ifdef VK_USE_PLATFORM_XLIB_XRANDR_EXT
386 .EXT_acquire_xlib_display = true,
388 #ifdef VK_USE_PLATFORM_DISPLAY_KHR
390 .KHR_get_display_properties2 = true,
391 .EXT_direct_mode_display = true,
392 .EXT_display_surface_counter = true,
393 .EXT_acquire_drm_display = true,
398 radv_physical_device_get_supported_extensions(const struct radv_physical_device *device,
399 struct vk_device_extension_table *ext)
401 *ext = (struct vk_device_extension_table){
402 .KHR_8bit_storage = true,
403 .KHR_16bit_storage = true,
404 .KHR_acceleration_structure = radv_enable_rt(device, false),
405 .KHR_bind_memory2 = true,
406 .KHR_buffer_device_address = true,
407 .KHR_copy_commands2 = true,
408 .KHR_create_renderpass2 = true,
409 .KHR_dedicated_allocation = true,
410 .KHR_deferred_host_operations = true,
411 .KHR_depth_stencil_resolve = true,
412 .KHR_descriptor_update_template = true,
413 .KHR_device_group = true,
414 .KHR_draw_indirect_count = true,
415 .KHR_driver_properties = true,
416 .KHR_dynamic_rendering = true,
417 .KHR_external_fence = true,
418 .KHR_external_fence_fd = true,
419 .KHR_external_memory = true,
420 .KHR_external_memory_fd = true,
421 .KHR_external_semaphore = true,
422 .KHR_external_semaphore_fd = true,
423 .KHR_format_feature_flags2 = true,
424 .KHR_fragment_shading_rate = device->rad_info.gfx_level >= GFX10_3,
425 .KHR_get_memory_requirements2 = true,
426 .KHR_image_format_list = true,
427 .KHR_imageless_framebuffer = true,
428 #ifdef RADV_USE_WSI_PLATFORM
429 .KHR_incremental_present = true,
431 .KHR_maintenance1 = true,
432 .KHR_maintenance2 = true,
433 .KHR_maintenance3 = true,
434 .KHR_maintenance4 = true,
435 .KHR_multiview = true,
436 .KHR_pipeline_executable_properties = true,
437 .KHR_pipeline_library = !device->use_llvm,
438 .KHR_push_descriptor = true,
439 .KHR_ray_query = radv_enable_rt(device, false),
440 .KHR_ray_tracing_maintenance1 = radv_enable_rt(device, false),
441 .KHR_ray_tracing_pipeline = radv_enable_rt(device, true),
442 .KHR_relaxed_block_layout = true,
443 .KHR_sampler_mirror_clamp_to_edge = true,
444 .KHR_sampler_ycbcr_conversion = true,
445 .KHR_separate_depth_stencil_layouts = true,
446 .KHR_shader_atomic_int64 = true,
447 .KHR_shader_clock = true,
448 .KHR_shader_draw_parameters = true,
449 .KHR_shader_float16_int8 = true,
450 .KHR_shader_float_controls = true,
451 .KHR_shader_integer_dot_product = true,
452 .KHR_shader_non_semantic_info = true,
453 .KHR_shader_subgroup_extended_types = true,
454 .KHR_shader_subgroup_uniform_control_flow = true,
455 .KHR_shader_terminate_invocation = true,
456 .KHR_spirv_1_4 = true,
457 .KHR_storage_buffer_storage_class = true,
458 #ifdef RADV_USE_WSI_PLATFORM
459 .KHR_swapchain = true,
460 .KHR_swapchain_mutable_format = true,
462 .KHR_synchronization2 = true,
463 .KHR_timeline_semaphore = true,
464 .KHR_uniform_buffer_standard_layout = true,
465 .KHR_variable_pointers = true,
466 .KHR_vulkan_memory_model = true,
467 .KHR_workgroup_memory_explicit_layout = true,
468 .KHR_zero_initialize_workgroup_memory = true,
469 .EXT_4444_formats = true,
470 .EXT_border_color_swizzle = device->rad_info.gfx_level >= GFX10,
471 .EXT_buffer_device_address = true,
472 .EXT_calibrated_timestamps = RADV_SUPPORT_CALIBRATED_TIMESTAMPS,
473 .EXT_color_write_enable = true,
474 .EXT_conditional_rendering = true,
475 .EXT_conservative_rasterization = device->rad_info.gfx_level >= GFX9,
476 .EXT_custom_border_color = true,
477 .EXT_debug_marker = radv_thread_trace_enabled(),
478 .EXT_depth_clip_control = true,
479 .EXT_depth_clip_enable = true,
480 .EXT_depth_range_unrestricted = true,
481 .EXT_descriptor_indexing = true,
482 .EXT_discard_rectangles = true,
483 #ifdef VK_USE_PLATFORM_DISPLAY_KHR
484 .EXT_display_control = true,
486 .EXT_extended_dynamic_state = true,
487 .EXT_extended_dynamic_state2 = true,
488 .EXT_external_memory_dma_buf = true,
489 .EXT_external_memory_host = device->rad_info.has_userptr,
490 .EXT_global_priority = true,
491 .EXT_global_priority_query = true,
492 .EXT_host_query_reset = true,
493 .EXT_image_2d_view_of_3d = true,
494 .EXT_image_drm_format_modifier = device->rad_info.gfx_level >= GFX9,
495 .EXT_image_robustness = true,
496 .EXT_image_view_min_lod = true,
497 .EXT_index_type_uint8 = device->rad_info.gfx_level >= GFX8,
498 .EXT_inline_uniform_block = true,
499 .EXT_line_rasterization = true,
500 .EXT_memory_budget = true,
501 .EXT_memory_priority = true,
502 .EXT_multi_draw = true,
503 .EXT_non_seamless_cube_map = true,
504 .EXT_pci_bus_info = true,
506 .EXT_physical_device_drm = true,
508 .EXT_pipeline_creation_cache_control = true,
509 .EXT_pipeline_creation_feedback = true,
510 .EXT_post_depth_coverage = device->rad_info.gfx_level >= GFX10,
511 .EXT_primitive_topology_list_restart = true,
512 .EXT_primitives_generated_query = true,
513 .EXT_private_data = true,
514 .EXT_provoking_vertex = true,
515 .EXT_queue_family_foreign = true,
516 .EXT_robustness2 = true,
517 .EXT_sample_locations = device->rad_info.gfx_level < GFX10,
518 .EXT_sampler_filter_minmax = true,
519 .EXT_scalar_block_layout = device->rad_info.gfx_level >= GFX7,
520 .EXT_separate_stencil_usage = true,
521 .EXT_shader_atomic_float = true,
522 #ifdef LLVM_AVAILABLE
523 .EXT_shader_atomic_float2 = !device->use_llvm || LLVM_VERSION_MAJOR >= 14,
525 .EXT_shader_atomic_float2 = true,
527 .EXT_shader_demote_to_helper_invocation = true,
528 .EXT_shader_image_atomic_int64 = true,
529 .EXT_shader_module_identifier = true,
530 .EXT_shader_stencil_export = true,
531 .EXT_shader_subgroup_ballot = true,
532 .EXT_shader_subgroup_vote = true,
533 .EXT_shader_viewport_index_layer = true,
534 .EXT_subgroup_size_control = true,
535 .EXT_texel_buffer_alignment = true,
536 .EXT_transform_feedback = device->rad_info.gfx_level < GFX11,
537 .EXT_vertex_attribute_divisor = true,
538 .EXT_vertex_input_dynamic_state = !device->use_llvm,
539 .EXT_ycbcr_image_arrays = true,
540 .AMD_buffer_marker = true,
541 .AMD_device_coherent_memory = true,
542 .AMD_draw_indirect_count = true,
543 .AMD_gcn_shader = true,
544 .AMD_gpu_shader_half_float = device->rad_info.has_packed_math_16bit,
545 .AMD_gpu_shader_int16 = device->rad_info.has_packed_math_16bit,
546 .AMD_memory_overallocation_behavior = true,
547 .AMD_mixed_attachment_samples = true,
548 .AMD_rasterization_order = device->rad_info.has_out_of_order_rast,
549 .AMD_shader_ballot = true,
550 .AMD_shader_core_properties = true,
551 .AMD_shader_core_properties2 = true,
552 .AMD_shader_explicit_vertex_parameter = true,
553 .AMD_shader_fragment_mask = device->rad_info.gfx_level < GFX11,
554 .AMD_shader_image_load_store_lod = true,
555 .AMD_shader_trinary_minmax = true,
556 .AMD_texture_gather_bias_lod = true,
558 .ANDROID_external_memory_android_hardware_buffer = RADV_SUPPORT_ANDROID_HARDWARE_BUFFER,
559 .ANDROID_native_buffer = true,
561 .GOOGLE_decorate_string = true,
562 .GOOGLE_hlsl_functionality1 = true,
563 .GOOGLE_user_type = true,
564 .INTEL_shader_integer_functions2 = true,
565 .NV_compute_shader_derivatives = true,
566 .NV_mesh_shader = device->use_ngg && device->rad_info.gfx_level >= GFX10_3 &&
567 device->instance->perftest_flags & RADV_PERFTEST_NV_MS && !device->use_llvm,
568 /* Undocumented extension purely for vkd3d-proton. This check is to prevent anyone else from
571 .VALVE_descriptor_set_host_mapping =
572 device->vk.instance->app_info.engine_name &&
573 strcmp(device->vk.instance->app_info.engine_name, "vkd3d") == 0,
574 .VALVE_mutable_descriptor_type = true,
579 radv_is_conformant(const struct radv_physical_device *pdevice)
581 return pdevice->rad_info.gfx_level >= GFX8;
585 radv_physical_device_init_queue_table(struct radv_physical_device *pdevice)
588 pdevice->vk_queue_to_radv[idx] = RADV_QUEUE_GENERAL;
591 for (unsigned i = 1; i < RADV_MAX_QUEUE_FAMILIES; i++)
592 pdevice->vk_queue_to_radv[i] = RADV_MAX_QUEUE_FAMILIES + 1;
594 if (pdevice->rad_info.ip[AMD_IP_COMPUTE].num_queues > 0 &&
595 !(pdevice->instance->debug_flags & RADV_DEBUG_NO_COMPUTE_QUEUE)) {
596 pdevice->vk_queue_to_radv[idx] = RADV_QUEUE_COMPUTE;
599 pdevice->num_queues = idx;
603 radv_physical_device_try_create(struct radv_instance *instance, drmDevicePtr drm_device,
604 struct radv_physical_device **device_out)
611 assert(drm_device == NULL);
614 const char *path = drm_device->nodes[DRM_NODE_RENDER];
615 drmVersionPtr version;
617 fd = open(path, O_RDWR | O_CLOEXEC);
619 return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
620 "Could not open device %s: %m", path);
623 version = drmGetVersion(fd);
627 return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
628 "Could not get the kernel driver version for device %s: %m", path);
631 if (strcmp(version->name, "amdgpu")) {
632 drmFreeVersion(version);
635 return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
636 "Device '%s' is not using the AMDGPU kernel driver: %m", path);
638 drmFreeVersion(version);
640 if (instance->debug_flags & RADV_DEBUG_STARTUP)
641 fprintf(stderr, "radv: info: Found compatible device '%s'.\n", path);
645 struct radv_physical_device *device = vk_zalloc2(&instance->vk.alloc, NULL, sizeof(*device), 8,
646 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
648 result = vk_error(instance, VK_ERROR_OUT_OF_HOST_MEMORY);
652 struct vk_physical_device_dispatch_table dispatch_table;
653 vk_physical_device_dispatch_table_from_entrypoints(&dispatch_table,
654 &radv_physical_device_entrypoints, true);
655 vk_physical_device_dispatch_table_from_entrypoints(&dispatch_table,
656 &wsi_physical_device_entrypoints, false);
658 result = vk_physical_device_init(&device->vk, &instance->vk, NULL, &dispatch_table);
659 if (result != VK_SUCCESS) {
663 device->instance = instance;
666 device->ws = radv_null_winsys_create();
669 bool reserve_vmid = radv_thread_trace_enabled();
671 device->ws = radv_amdgpu_winsys_create(fd, instance->debug_flags, instance->perftest_flags,
674 device->ws = radv_null_winsys_create();
679 result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED, "failed to initialize winsys");
683 device->vk.supported_sync_types = device->ws->get_sync_types(device->ws);
686 if (drm_device && instance->vk.enabled_extensions.KHR_display) {
687 master_fd = open(drm_device->nodes[DRM_NODE_PRIMARY], O_RDWR | O_CLOEXEC);
688 if (master_fd >= 0) {
689 uint32_t accel_working = 0;
690 struct drm_amdgpu_info request = {.return_pointer = (uintptr_t)&accel_working,
691 .return_size = sizeof(accel_working),
692 .query = AMDGPU_INFO_ACCEL_WORKING};
694 if (drmCommandWrite(master_fd, DRM_AMDGPU_INFO, &request, sizeof(struct drm_amdgpu_info)) <
704 device->master_fd = master_fd;
705 device->local_fd = fd;
706 device->ws->query_info(device->ws, &device->rad_info);
708 device->use_llvm = instance->debug_flags & RADV_DEBUG_LLVM;
709 #ifndef LLVM_AVAILABLE
710 if (device->use_llvm) {
711 fprintf(stderr, "ERROR: LLVM compiler backend selected for radv, but LLVM support was not "
712 "enabled at build time.\n");
718 device->emulate_etc2 = !radv_device_supports_etc(device);
720 device->emulate_etc2 = !radv_device_supports_etc(device) &&
721 driQueryOptionb(&device->instance->dri_options, "radv_require_etc2");
724 snprintf(device->name, sizeof(device->name), "AMD RADV %s%s", device->rad_info.name,
725 radv_get_compiler_string(device));
727 const char *marketing_name = device->ws->get_chip_name(device->ws);
728 snprintf(device->marketing_name, sizeof(device->name), "%s (RADV %s%s)",
729 marketing_name, device->rad_info.name, radv_get_compiler_string(device));
731 #ifdef ENABLE_SHADER_CACHE
732 if (radv_device_get_cache_uuid(device, device->cache_uuid)) {
733 result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED, "cannot generate UUID");
737 /* The gpu id is already embedded in the uuid so we just pass "radv"
738 * when creating the cache.
740 char buf[VK_UUID_SIZE * 2 + 1];
741 disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
742 device->disk_cache = disk_cache_create(device->name, buf, 0);
745 if (!radv_is_conformant(device))
746 vk_warn_non_conformant_implementation("radv");
748 radv_get_driver_uuid(&device->driver_uuid);
749 radv_get_device_uuid(&device->rad_info, &device->device_uuid);
751 device->out_of_order_rast_allowed =
752 device->rad_info.has_out_of_order_rast &&
753 !(device->instance->debug_flags & RADV_DEBUG_NO_OUT_OF_ORDER);
755 device->dcc_msaa_allowed = (device->instance->perftest_flags & RADV_PERFTEST_DCC_MSAA);
757 device->use_ngg = (device->rad_info.gfx_level >= GFX10 &&
758 device->rad_info.family != CHIP_NAVI14 &&
759 !(device->instance->debug_flags & RADV_DEBUG_NO_NGG)) ||
760 device->rad_info.gfx_level >= GFX11;
762 device->use_ngg_culling = device->use_ngg && device->rad_info.max_render_backends > 1 &&
763 (device->rad_info.gfx_level >= GFX10_3 ||
764 (device->instance->perftest_flags & RADV_PERFTEST_NGGC)) &&
765 !(device->instance->debug_flags & RADV_DEBUG_NO_NGGC);
767 device->use_ngg_streamout = false;
769 /* Determine the number of threads per wave for all stages. */
770 device->cs_wave_size = 64;
771 device->ps_wave_size = 64;
772 device->ge_wave_size = 64;
773 device->rt_wave_size = 64;
775 if (device->rad_info.gfx_level >= GFX10) {
776 if (device->instance->perftest_flags & RADV_PERFTEST_CS_WAVE_32)
777 device->cs_wave_size = 32;
779 /* For pixel shaders, wave64 is recommanded. */
780 if (device->instance->perftest_flags & RADV_PERFTEST_PS_WAVE_32)
781 device->ps_wave_size = 32;
783 if (device->instance->perftest_flags & RADV_PERFTEST_GE_WAVE_32)
784 device->ge_wave_size = 32;
786 if (!(device->instance->perftest_flags & RADV_PERFTEST_RT_WAVE_64))
787 device->rt_wave_size = 32;
790 radv_physical_device_init_mem_types(device);
792 radv_physical_device_get_supported_extensions(device, &device->vk.supported_extensions);
794 radv_get_nir_options(device);
798 struct stat primary_stat = {0}, render_stat = {0};
800 device->available_nodes = drm_device->available_nodes;
801 device->bus_info = *drm_device->businfo.pci;
803 if ((drm_device->available_nodes & (1 << DRM_NODE_PRIMARY)) &&
804 stat(drm_device->nodes[DRM_NODE_PRIMARY], &primary_stat) != 0) {
805 result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
806 "failed to stat DRM primary node %s",
807 drm_device->nodes[DRM_NODE_PRIMARY]);
808 goto fail_perfcounters;
810 device->primary_devid = primary_stat.st_rdev;
812 if ((drm_device->available_nodes & (1 << DRM_NODE_RENDER)) &&
813 stat(drm_device->nodes[DRM_NODE_RENDER], &render_stat) != 0) {
814 result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
815 "failed to stat DRM render node %s",
816 drm_device->nodes[DRM_NODE_RENDER]);
817 goto fail_perfcounters;
819 device->render_devid = render_stat.st_rdev;
823 if ((device->instance->debug_flags & RADV_DEBUG_INFO))
824 ac_print_gpu_info(&device->rad_info, stdout);
826 radv_physical_device_init_queue_table(device);
828 /* We don't check the error code, but later check if it is initialized. */
829 ac_init_perfcounters(&device->rad_info, false, false, &device->ac_perfcounters);
831 /* The WSI is structured as a layer on top of the driver, so this has
832 * to be the last part of initialization (at least until we get other
835 result = radv_init_wsi(device);
836 if (result != VK_SUCCESS) {
837 vk_error(instance, result);
838 goto fail_perfcounters;
841 device->gs_table_depth =
842 ac_get_gs_table_depth(device->rad_info.gfx_level, device->rad_info.family);
844 ac_get_hs_info(&device->rad_info, &device->hs);
845 ac_get_task_info(&device->rad_info, &device->task_info);
847 *device_out = device;
852 ac_destroy_perfcounters(&device->ac_perfcounters);
853 disk_cache_destroy(device->disk_cache);
854 #ifdef ENABLE_SHADER_CACHE
857 device->ws->destroy(device->ws);
859 vk_physical_device_finish(&device->vk);
861 vk_free(&instance->vk.alloc, device);
871 radv_physical_device_destroy(struct radv_physical_device *device)
873 radv_finish_wsi(device);
874 ac_destroy_perfcounters(&device->ac_perfcounters);
875 device->ws->destroy(device->ws);
876 disk_cache_destroy(device->disk_cache);
877 if (device->local_fd != -1)
878 close(device->local_fd);
879 if (device->master_fd != -1)
880 close(device->master_fd);
881 vk_physical_device_finish(&device->vk);
882 vk_free(&device->instance->vk.alloc, device);
885 static const struct debug_control radv_debug_options[] = {
886 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS},
887 {"nodcc", RADV_DEBUG_NO_DCC},
888 {"shaders", RADV_DEBUG_DUMP_SHADERS},
889 {"nocache", RADV_DEBUG_NO_CACHE},
890 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS},
891 {"nohiz", RADV_DEBUG_NO_HIZ},
892 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE},
893 {"allbos", RADV_DEBUG_ALL_BOS},
894 {"noibs", RADV_DEBUG_NO_IBS},
895 {"spirv", RADV_DEBUG_DUMP_SPIRV},
896 {"vmfaults", RADV_DEBUG_VM_FAULTS},
897 {"zerovram", RADV_DEBUG_ZERO_VRAM},
898 {"syncshaders", RADV_DEBUG_SYNC_SHADERS},
899 {"preoptir", RADV_DEBUG_PREOPTIR},
900 {"nodynamicbounds", RADV_DEBUG_NO_DYNAMIC_BOUNDS},
901 {"nooutoforder", RADV_DEBUG_NO_OUT_OF_ORDER},
902 {"info", RADV_DEBUG_INFO},
903 {"startup", RADV_DEBUG_STARTUP},
904 {"checkir", RADV_DEBUG_CHECKIR},
905 {"nobinning", RADV_DEBUG_NOBINNING},
906 {"nongg", RADV_DEBUG_NO_NGG},
907 {"metashaders", RADV_DEBUG_DUMP_META_SHADERS},
908 {"nomemorycache", RADV_DEBUG_NO_MEMORY_CACHE},
909 {"discardtodemote", RADV_DEBUG_DISCARD_TO_DEMOTE},
910 {"llvm", RADV_DEBUG_LLVM},
911 {"forcecompress", RADV_DEBUG_FORCE_COMPRESS},
912 {"hang", RADV_DEBUG_HANG},
913 {"img", RADV_DEBUG_IMG},
914 {"noumr", RADV_DEBUG_NO_UMR},
915 {"invariantgeom", RADV_DEBUG_INVARIANT_GEOM},
916 {"splitfma", RADV_DEBUG_SPLIT_FMA},
917 {"nodisplaydcc", RADV_DEBUG_NO_DISPLAY_DCC},
918 {"notccompatcmask", RADV_DEBUG_NO_TC_COMPAT_CMASK},
919 {"novrsflatshading", RADV_DEBUG_NO_VRS_FLAT_SHADING},
920 {"noatocdithering", RADV_DEBUG_NO_ATOC_DITHERING},
921 {"nonggc", RADV_DEBUG_NO_NGGC},
922 {"prologs", RADV_DEBUG_DUMP_PROLOGS},
923 {"nodma", RADV_DEBUG_NO_DMA_BLIT},
927 radv_get_debug_option_name(int id)
929 assert(id < ARRAY_SIZE(radv_debug_options) - 1);
930 return radv_debug_options[id].string;
933 static const struct debug_control radv_perftest_options[] = {{"localbos", RADV_PERFTEST_LOCAL_BOS},
934 {"dccmsaa", RADV_PERFTEST_DCC_MSAA},
935 {"bolist", RADV_PERFTEST_BO_LIST},
936 {"cswave32", RADV_PERFTEST_CS_WAVE_32},
937 {"pswave32", RADV_PERFTEST_PS_WAVE_32},
938 {"gewave32", RADV_PERFTEST_GE_WAVE_32},
939 {"nosam", RADV_PERFTEST_NO_SAM},
940 {"sam", RADV_PERFTEST_SAM},
941 {"rt", RADV_PERFTEST_RT},
942 {"nggc", RADV_PERFTEST_NGGC},
943 {"emulate_rt", RADV_PERFTEST_EMULATE_RT},
944 {"nv_ms", RADV_PERFTEST_NV_MS},
945 {"rtwave64", RADV_PERFTEST_RT_WAVE_64},
949 radv_get_perftest_option_name(int id)
951 assert(id < ARRAY_SIZE(radv_perftest_options) - 1);
952 return radv_perftest_options[id].string;
956 static const driOptionDescription radv_dri_options[] = {
957 DRI_CONF_SECTION_PERFORMANCE
958 DRI_CONF_ADAPTIVE_SYNC(true)
959 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
960 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT(false)
961 DRI_CONF_VK_X11_ENSURE_MIN_IMAGE_COUNT(false)
962 DRI_CONF_VK_XWAYLAND_WAIT_READY(true)
963 DRI_CONF_RADV_REPORT_LLVM9_VERSION_STRING(false)
964 DRI_CONF_RADV_ENABLE_MRT_OUTPUT_NAN_FIXUP(false)
965 DRI_CONF_RADV_DISABLE_SHRINK_IMAGE_STORE(false)
966 DRI_CONF_RADV_NO_DYNAMIC_BOUNDS(false)
967 DRI_CONF_RADV_ABSOLUTE_DEPTH_BIAS(false)
968 DRI_CONF_RADV_OVERRIDE_UNIFORM_OFFSET_ALIGNMENT(0)
971 DRI_CONF_SECTION_DEBUG
972 DRI_CONF_OVERRIDE_VRAM_SIZE()
973 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST(false)
974 DRI_CONF_RADV_ZERO_VRAM(false)
975 DRI_CONF_RADV_LOWER_DISCARD_TO_DEMOTE(false)
976 DRI_CONF_RADV_INVARIANT_GEOM(false)
977 DRI_CONF_RADV_SPLIT_FMA(false)
978 DRI_CONF_RADV_DISABLE_TC_COMPAT_HTILE_GENERAL(false)
979 DRI_CONF_RADV_DISABLE_DCC(false)
980 DRI_CONF_RADV_REPORT_APU_AS_DGPU(false)
981 DRI_CONF_RADV_REQUIRE_ETC2(false)
982 DRI_CONF_RADV_DISABLE_ANISO_SINGLE_LEVEL(false)
983 DRI_CONF_RADV_DISABLE_SINKING_LOAD_INPUT_FS(false)
989 radv_init_dri_options(struct radv_instance *instance)
991 driParseOptionInfo(&instance->available_dri_options, radv_dri_options,
992 ARRAY_SIZE(radv_dri_options));
993 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options, 0, "radv", NULL, NULL,
994 instance->vk.app_info.app_name, instance->vk.app_info.app_version,
995 instance->vk.app_info.engine_name, instance->vk.app_info.engine_version);
997 instance->enable_mrt_output_nan_fixup =
998 driQueryOptionb(&instance->dri_options, "radv_enable_mrt_output_nan_fixup");
1000 instance->disable_shrink_image_store =
1001 driQueryOptionb(&instance->dri_options, "radv_disable_shrink_image_store");
1003 instance->absolute_depth_bias =
1004 driQueryOptionb(&instance->dri_options, "radv_absolute_depth_bias");
1006 instance->disable_tc_compat_htile_in_general =
1007 driQueryOptionb(&instance->dri_options, "radv_disable_tc_compat_htile_general");
1009 if (driQueryOptionb(&instance->dri_options, "radv_no_dynamic_bounds"))
1010 instance->debug_flags |= RADV_DEBUG_NO_DYNAMIC_BOUNDS;
1012 if (driQueryOptionb(&instance->dri_options, "radv_lower_discard_to_demote"))
1013 instance->debug_flags |= RADV_DEBUG_DISCARD_TO_DEMOTE;
1015 if (driQueryOptionb(&instance->dri_options, "radv_invariant_geom"))
1016 instance->debug_flags |= RADV_DEBUG_INVARIANT_GEOM;
1018 if (driQueryOptionb(&instance->dri_options, "radv_split_fma"))
1019 instance->debug_flags |= RADV_DEBUG_SPLIT_FMA;
1021 if (driQueryOptionb(&instance->dri_options, "radv_disable_dcc"))
1022 instance->debug_flags |= RADV_DEBUG_NO_DCC;
1024 instance->zero_vram =
1025 driQueryOptionb(&instance->dri_options, "radv_zero_vram");
1027 instance->report_apu_as_dgpu =
1028 driQueryOptionb(&instance->dri_options, "radv_report_apu_as_dgpu");
1030 instance->disable_aniso_single_level =
1031 driQueryOptionb(&instance->dri_options, "radv_disable_aniso_single_level");
1033 instance->disable_sinking_load_input_fs =
1034 driQueryOptionb(&instance->dri_options, "radv_disable_sinking_load_input_fs");
1037 VKAPI_ATTR VkResult VKAPI_CALL
1038 radv_CreateInstance(const VkInstanceCreateInfo *pCreateInfo,
1039 const VkAllocationCallbacks *pAllocator, VkInstance *pInstance)
1041 struct radv_instance *instance;
1045 pAllocator = vk_default_allocator();
1047 instance = vk_zalloc(pAllocator, sizeof(*instance), 8, VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
1049 return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY);
1051 struct vk_instance_dispatch_table dispatch_table;
1052 vk_instance_dispatch_table_from_entrypoints(&dispatch_table, &radv_instance_entrypoints, true);
1053 vk_instance_dispatch_table_from_entrypoints(&dispatch_table, &wsi_instance_entrypoints, false);
1054 struct vk_instance_extension_table extensions_supported = radv_instance_extensions_supported;
1056 result = vk_instance_init(&instance->vk, &extensions_supported, &dispatch_table,
1057 pCreateInfo, pAllocator);
1058 if (result != VK_SUCCESS) {
1059 vk_free(pAllocator, instance);
1060 return vk_error(instance, result);
1063 instance->debug_flags = parse_debug_string(getenv("RADV_DEBUG"), radv_debug_options);
1064 instance->perftest_flags = parse_debug_string(getenv("RADV_PERFTEST"), radv_perftest_options);
1066 if (instance->debug_flags & RADV_DEBUG_STARTUP)
1067 fprintf(stderr, "radv: info: Created an instance.\n");
1069 instance->physical_devices_enumerated = false;
1070 list_inithead(&instance->physical_devices);
1072 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
1074 radv_init_dri_options(instance);
1076 *pInstance = radv_instance_to_handle(instance);
1081 VKAPI_ATTR void VKAPI_CALL
1082 radv_DestroyInstance(VkInstance _instance, const VkAllocationCallbacks *pAllocator)
1084 RADV_FROM_HANDLE(radv_instance, instance, _instance);
1089 list_for_each_entry_safe(struct radv_physical_device, pdevice, &instance->physical_devices, link)
1091 radv_physical_device_destroy(pdevice);
1094 VG(VALGRIND_DESTROY_MEMPOOL(instance));
1096 driDestroyOptionCache(&instance->dri_options);
1097 driDestroyOptionInfo(&instance->available_dri_options);
1099 vk_instance_finish(&instance->vk);
1100 vk_free(&instance->vk.alloc, instance);
1104 radv_enumerate_physical_devices(struct radv_instance *instance)
1106 if (instance->physical_devices_enumerated)
1109 instance->physical_devices_enumerated = true;
1111 VkResult result = VK_SUCCESS;
1113 if (getenv("RADV_FORCE_FAMILY")) {
1114 /* When RADV_FORCE_FAMILY is set, the driver creates a nul
1115 * device that allows to test the compiler without having an
1118 struct radv_physical_device *pdevice;
1120 result = radv_physical_device_try_create(instance, NULL, &pdevice);
1121 if (result != VK_SUCCESS)
1124 list_addtail(&pdevice->link, &instance->physical_devices);
1129 /* TODO: Check for more devices ? */
1130 drmDevicePtr devices[8];
1131 int max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
1133 if (instance->debug_flags & RADV_DEBUG_STARTUP)
1134 fprintf(stderr, "radv: info: Found %d drm nodes.\n", max_devices);
1136 if (max_devices < 1)
1137 return vk_error(instance, VK_SUCCESS);
1139 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
1140 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
1141 devices[i]->bustype == DRM_BUS_PCI &&
1142 devices[i]->deviceinfo.pci->vendor_id == ATI_VENDOR_ID) {
1144 struct radv_physical_device *pdevice;
1145 result = radv_physical_device_try_create(instance, devices[i], &pdevice);
1146 /* Incompatible DRM device, skip. */
1147 if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
1148 result = VK_SUCCESS;
1152 /* Error creating the physical device, report the error. */
1153 if (result != VK_SUCCESS)
1156 list_addtail(&pdevice->link, &instance->physical_devices);
1159 drmFreeDevices(devices, max_devices);
1162 /* If we successfully enumerated any devices, call it success */
1166 VKAPI_ATTR VkResult VKAPI_CALL
1167 radv_EnumeratePhysicalDevices(VkInstance _instance, uint32_t *pPhysicalDeviceCount,
1168 VkPhysicalDevice *pPhysicalDevices)
1170 RADV_FROM_HANDLE(radv_instance, instance, _instance);
1171 VK_OUTARRAY_MAKE_TYPED(VkPhysicalDevice, out, pPhysicalDevices, pPhysicalDeviceCount);
1173 VkResult result = radv_enumerate_physical_devices(instance);
1174 if (result != VK_SUCCESS)
1177 list_for_each_entry(struct radv_physical_device, pdevice, &instance->physical_devices, link)
1179 vk_outarray_append_typed(VkPhysicalDevice, &out, i)
1181 *i = radv_physical_device_to_handle(pdevice);
1185 return vk_outarray_status(&out);
1188 VKAPI_ATTR VkResult VKAPI_CALL
1189 radv_EnumeratePhysicalDeviceGroups(VkInstance _instance, uint32_t *pPhysicalDeviceGroupCount,
1190 VkPhysicalDeviceGroupProperties *pPhysicalDeviceGroupProperties)
1192 RADV_FROM_HANDLE(radv_instance, instance, _instance);
1193 VK_OUTARRAY_MAKE_TYPED(VkPhysicalDeviceGroupProperties, out, pPhysicalDeviceGroupProperties,
1194 pPhysicalDeviceGroupCount);
1196 VkResult result = radv_enumerate_physical_devices(instance);
1197 if (result != VK_SUCCESS)
1200 list_for_each_entry(struct radv_physical_device, pdevice, &instance->physical_devices, link)
1202 vk_outarray_append_typed(VkPhysicalDeviceGroupProperties, &out, p)
1204 p->physicalDeviceCount = 1;
1205 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
1206 p->physicalDevices[0] = radv_physical_device_to_handle(pdevice);
1207 p->subsetAllocation = false;
1211 return vk_outarray_status(&out);
1214 VKAPI_ATTR void VKAPI_CALL
1215 radv_GetPhysicalDeviceFeatures(VkPhysicalDevice physicalDevice, VkPhysicalDeviceFeatures *pFeatures)
1217 RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
1218 memset(pFeatures, 0, sizeof(*pFeatures));
1220 *pFeatures = (VkPhysicalDeviceFeatures){
1221 .robustBufferAccess = true,
1222 .fullDrawIndexUint32 = true,
1223 .imageCubeArray = true,
1224 .independentBlend = true,
1225 .geometryShader = true,
1226 .tessellationShader = true,
1227 .sampleRateShading = true,
1228 .dualSrcBlend = true,
1230 .multiDrawIndirect = true,
1231 .drawIndirectFirstInstance = true,
1233 .depthBiasClamp = true,
1234 .fillModeNonSolid = true,
1235 .depthBounds = true,
1237 .largePoints = true,
1238 .alphaToOne = false,
1239 .multiViewport = true,
1240 .samplerAnisotropy = true,
1241 .textureCompressionETC2 = radv_device_supports_etc(pdevice) || pdevice->emulate_etc2,
1242 .textureCompressionASTC_LDR = false,
1243 .textureCompressionBC = true,
1244 .occlusionQueryPrecise = true,
1245 .pipelineStatisticsQuery = true,
1246 .vertexPipelineStoresAndAtomics = true,
1247 .fragmentStoresAndAtomics = true,
1248 .shaderTessellationAndGeometryPointSize = true,
1249 .shaderImageGatherExtended = true,
1250 .shaderStorageImageExtendedFormats = true,
1251 .shaderStorageImageMultisample = true,
1252 .shaderUniformBufferArrayDynamicIndexing = true,
1253 .shaderSampledImageArrayDynamicIndexing = true,
1254 .shaderStorageBufferArrayDynamicIndexing = true,
1255 .shaderStorageImageArrayDynamicIndexing = true,
1256 .shaderStorageImageReadWithoutFormat = true,
1257 .shaderStorageImageWriteWithoutFormat = true,
1258 .shaderClipDistance = true,
1259 .shaderCullDistance = true,
1260 .shaderFloat64 = true,
1261 .shaderInt64 = true,
1262 .shaderInt16 = true,
1263 .sparseBinding = true,
1264 .sparseResidencyBuffer = pdevice->rad_info.family >= CHIP_POLARIS10,
1265 .sparseResidencyImage2D = pdevice->rad_info.family >= CHIP_POLARIS10,
1266 .sparseResidencyAliased = pdevice->rad_info.family >= CHIP_POLARIS10,
1267 .variableMultisampleRate = true,
1268 .shaderResourceMinLod = true,
1269 .shaderResourceResidency = true,
1270 .inheritedQueries = true,
1275 radv_get_physical_device_features_1_1(struct radv_physical_device *pdevice,
1276 VkPhysicalDeviceVulkan11Features *f)
1278 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES);
1280 f->storageBuffer16BitAccess = true;
1281 f->uniformAndStorageBuffer16BitAccess = true;
1282 f->storagePushConstant16 = true;
1283 f->storageInputOutput16 = pdevice->rad_info.has_packed_math_16bit;
1284 f->multiview = true;
1285 f->multiviewGeometryShader = true;
1286 f->multiviewTessellationShader = true;
1287 f->variablePointersStorageBuffer = true;
1288 f->variablePointers = true;
1289 f->protectedMemory = false;
1290 f->samplerYcbcrConversion = true;
1291 f->shaderDrawParameters = true;
1295 radv_get_physical_device_features_1_2(struct radv_physical_device *pdevice,
1296 VkPhysicalDeviceVulkan12Features *f)
1298 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES);
1300 f->samplerMirrorClampToEdge = true;
1301 f->drawIndirectCount = true;
1302 f->storageBuffer8BitAccess = true;
1303 f->uniformAndStorageBuffer8BitAccess = true;
1304 f->storagePushConstant8 = true;
1305 f->shaderBufferInt64Atomics = true;
1306 f->shaderSharedInt64Atomics = true;
1307 f->shaderFloat16 = pdevice->rad_info.has_packed_math_16bit;
1308 f->shaderInt8 = true;
1310 f->descriptorIndexing = true;
1311 f->shaderInputAttachmentArrayDynamicIndexing = true;
1312 f->shaderUniformTexelBufferArrayDynamicIndexing = true;
1313 f->shaderStorageTexelBufferArrayDynamicIndexing = true;
1314 f->shaderUniformBufferArrayNonUniformIndexing = true;
1315 f->shaderSampledImageArrayNonUniformIndexing = true;
1316 f->shaderStorageBufferArrayNonUniformIndexing = true;
1317 f->shaderStorageImageArrayNonUniformIndexing = true;
1318 f->shaderInputAttachmentArrayNonUniformIndexing = true;
1319 f->shaderUniformTexelBufferArrayNonUniformIndexing = true;
1320 f->shaderStorageTexelBufferArrayNonUniformIndexing = true;
1321 f->descriptorBindingUniformBufferUpdateAfterBind = true;
1322 f->descriptorBindingSampledImageUpdateAfterBind = true;
1323 f->descriptorBindingStorageImageUpdateAfterBind = true;
1324 f->descriptorBindingStorageBufferUpdateAfterBind = true;
1325 f->descriptorBindingUniformTexelBufferUpdateAfterBind = true;
1326 f->descriptorBindingStorageTexelBufferUpdateAfterBind = true;
1327 f->descriptorBindingUpdateUnusedWhilePending = true;
1328 f->descriptorBindingPartiallyBound = true;
1329 f->descriptorBindingVariableDescriptorCount = true;
1330 f->runtimeDescriptorArray = true;
1332 f->samplerFilterMinmax = true;
1333 f->scalarBlockLayout = pdevice->rad_info.gfx_level >= GFX7;
1334 f->imagelessFramebuffer = true;
1335 f->uniformBufferStandardLayout = true;
1336 f->shaderSubgroupExtendedTypes = true;
1337 f->separateDepthStencilLayouts = true;
1338 f->hostQueryReset = true;
1339 f->timelineSemaphore = true, f->bufferDeviceAddress = true;
1340 f->bufferDeviceAddressCaptureReplay = true;
1341 f->bufferDeviceAddressMultiDevice = false;
1342 f->vulkanMemoryModel = true;
1343 f->vulkanMemoryModelDeviceScope = true;
1344 f->vulkanMemoryModelAvailabilityVisibilityChains = false;
1345 f->shaderOutputViewportIndex = true;
1346 f->shaderOutputLayer = true;
1347 f->subgroupBroadcastDynamicId = true;
1351 radv_get_physical_device_features_1_3(struct radv_physical_device *pdevice,
1352 VkPhysicalDeviceVulkan13Features *f)
1354 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_FEATURES);
1356 f->robustImageAccess = true;
1357 f->inlineUniformBlock = true;
1358 f->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1359 f->pipelineCreationCacheControl = true;
1360 f->privateData = true;
1361 f->shaderDemoteToHelperInvocation = true;
1362 f->shaderTerminateInvocation = true;
1363 f->subgroupSizeControl = true;
1364 f->computeFullSubgroups = true;
1365 f->synchronization2 = true;
1366 f->textureCompressionASTC_HDR = false;
1367 f->shaderZeroInitializeWorkgroupMemory = true;
1368 f->dynamicRendering = true;
1369 f->shaderIntegerDotProduct = true;
1370 f->maintenance4 = true;
1373 VKAPI_ATTR void VKAPI_CALL
1374 radv_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
1375 VkPhysicalDeviceFeatures2 *pFeatures)
1377 RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
1378 radv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
1380 VkPhysicalDeviceVulkan11Features core_1_1 = {
1381 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
1383 radv_get_physical_device_features_1_1(pdevice, &core_1_1);
1385 VkPhysicalDeviceVulkan12Features core_1_2 = {
1386 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
1388 radv_get_physical_device_features_1_2(pdevice, &core_1_2);
1390 VkPhysicalDeviceVulkan13Features core_1_3 = {
1391 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_FEATURES,
1393 radv_get_physical_device_features_1_3(pdevice, &core_1_3);
1395 #define CORE_FEATURE(major, minor, feature) features->feature = core_##major##_##minor.feature
1397 vk_foreach_struct(ext, pFeatures->pNext)
1399 if (vk_get_physical_device_core_1_1_feature_ext(ext, &core_1_1))
1401 if (vk_get_physical_device_core_1_2_feature_ext(ext, &core_1_2))
1403 if (vk_get_physical_device_core_1_3_feature_ext(ext, &core_1_3))
1406 switch (ext->sType) {
1407 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1408 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1409 (VkPhysicalDeviceConditionalRenderingFeaturesEXT *)ext;
1410 features->conditionalRendering = true;
1411 features->inheritedConditionalRendering = false;
1414 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1415 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1416 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1417 features->vertexAttributeInstanceRateDivisor = true;
1418 features->vertexAttributeInstanceRateZeroDivisor = true;
1421 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1422 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1423 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1424 features->transformFeedback = pdevice->rad_info.gfx_level < GFX11;
1425 features->geometryStreams = !pdevice->use_ngg_streamout && pdevice->rad_info.gfx_level < GFX11;
1428 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES: {
1429 VkPhysicalDeviceScalarBlockLayoutFeatures *features =
1430 (VkPhysicalDeviceScalarBlockLayoutFeatures *)ext;
1431 CORE_FEATURE(1, 2, scalarBlockLayout);
1434 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PRIORITY_FEATURES_EXT: {
1435 VkPhysicalDeviceMemoryPriorityFeaturesEXT *features =
1436 (VkPhysicalDeviceMemoryPriorityFeaturesEXT *)ext;
1437 features->memoryPriority = true;
1440 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1441 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features =
1442 (VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *)ext;
1443 CORE_FEATURE(1, 2, bufferDeviceAddress);
1444 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay);
1445 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice);
1448 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1449 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1450 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1451 features->depthClipEnable = true;
1454 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1455 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1456 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1457 features->computeDerivativeGroupQuads = false;
1458 features->computeDerivativeGroupLinear = true;
1461 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1462 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1463 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1464 features->ycbcrImageArrays = true;
1467 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1468 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1469 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1470 features->indexTypeUint8 = pdevice->rad_info.gfx_level >= GFX8;
1473 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1474 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1475 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1476 features->pipelineExecutableInfo = true;
1479 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1480 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1481 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1482 features->shaderSubgroupClock = true;
1483 features->shaderDeviceClock = pdevice->rad_info.gfx_level >= GFX8;
1486 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1487 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1488 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1489 features->texelBufferAlignment = true;
1492 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COHERENT_MEMORY_FEATURES_AMD: {
1493 VkPhysicalDeviceCoherentMemoryFeaturesAMD *features =
1494 (VkPhysicalDeviceCoherentMemoryFeaturesAMD *)ext;
1495 features->deviceCoherentMemory = pdevice->rad_info.has_l2_uncached;
1498 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1499 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1500 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1501 features->rectangularLines = false;
1502 features->bresenhamLines = true;
1503 features->smoothLines = false;
1504 features->stippledRectangularLines = false;
1505 /* FIXME: Some stippled Bresenham CTS fails on Vega10
1506 * but work on Raven.
1508 features->stippledBresenhamLines = pdevice->rad_info.gfx_level != GFX9;
1509 features->stippledSmoothLines = false;
1512 case VK_STRUCTURE_TYPE_DEVICE_MEMORY_OVERALLOCATION_CREATE_INFO_AMD: {
1513 VkDeviceMemoryOverallocationCreateInfoAMD *features =
1514 (VkDeviceMemoryOverallocationCreateInfoAMD *)ext;
1515 features->overallocationBehavior = true;
1518 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT: {
1519 VkPhysicalDeviceRobustness2FeaturesEXT *features =
1520 (VkPhysicalDeviceRobustness2FeaturesEXT *)ext;
1521 features->robustBufferAccess2 = true;
1522 features->robustImageAccess2 = true;
1523 features->nullDescriptor = true;
1526 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT: {
1527 VkPhysicalDeviceCustomBorderColorFeaturesEXT *features =
1528 (VkPhysicalDeviceCustomBorderColorFeaturesEXT *)ext;
1529 features->customBorderColors = true;
1530 features->customBorderColorWithoutFormat = true;
1533 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT: {
1534 VkPhysicalDeviceExtendedDynamicStateFeaturesEXT *features =
1535 (VkPhysicalDeviceExtendedDynamicStateFeaturesEXT *)ext;
1536 features->extendedDynamicState = true;
1539 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT: {
1540 VkPhysicalDeviceShaderAtomicFloatFeaturesEXT *features =
1541 (VkPhysicalDeviceShaderAtomicFloatFeaturesEXT *)ext;
1542 features->shaderBufferFloat32Atomics = true;
1543 features->shaderBufferFloat32AtomicAdd = false;
1544 features->shaderBufferFloat64Atomics = true;
1545 features->shaderBufferFloat64AtomicAdd = false;
1546 features->shaderSharedFloat32Atomics = true;
1547 features->shaderSharedFloat32AtomicAdd = pdevice->rad_info.gfx_level >= GFX8;
1548 features->shaderSharedFloat64Atomics = true;
1549 features->shaderSharedFloat64AtomicAdd = false;
1550 features->shaderImageFloat32Atomics = true;
1551 features->shaderImageFloat32AtomicAdd = false;
1552 features->sparseImageFloat32Atomics = true;
1553 features->sparseImageFloat32AtomicAdd = false;
1556 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_4444_FORMATS_FEATURES_EXT: {
1557 VkPhysicalDevice4444FormatsFeaturesEXT *features =
1558 (VkPhysicalDevice4444FormatsFeaturesEXT *)ext;
1559 features->formatA4R4G4B4 = true;
1560 features->formatA4B4G4R4 = true;
1563 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_IMAGE_ATOMIC_INT64_FEATURES_EXT: {
1564 VkPhysicalDeviceShaderImageAtomicInt64FeaturesEXT *features =
1565 (VkPhysicalDeviceShaderImageAtomicInt64FeaturesEXT *)ext;
1566 features->shaderImageInt64Atomics = true;
1567 features->sparseImageInt64Atomics = true;
1570 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MUTABLE_DESCRIPTOR_TYPE_FEATURES_VALVE: {
1571 VkPhysicalDeviceMutableDescriptorTypeFeaturesVALVE *features =
1572 (VkPhysicalDeviceMutableDescriptorTypeFeaturesVALVE *)ext;
1573 features->mutableDescriptorType = true;
1576 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADING_RATE_FEATURES_KHR: {
1577 VkPhysicalDeviceFragmentShadingRateFeaturesKHR *features =
1578 (VkPhysicalDeviceFragmentShadingRateFeaturesKHR *)ext;
1579 features->pipelineFragmentShadingRate = true;
1580 features->primitiveFragmentShadingRate = true;
1581 features->attachmentFragmentShadingRate =
1582 !(pdevice->instance->debug_flags & RADV_DEBUG_NO_HIZ) &&
1583 pdevice->rad_info.gfx_level < GFX11; /* TODO: VRS no longer uses HTILE. */
1586 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR: {
1587 VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR *features =
1588 (VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR *)ext;
1589 features->workgroupMemoryExplicitLayout = true;
1590 features->workgroupMemoryExplicitLayoutScalarBlockLayout = true;
1591 features->workgroupMemoryExplicitLayout8BitAccess = true;
1592 features->workgroupMemoryExplicitLayout16BitAccess = true;
1595 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT: {
1596 VkPhysicalDeviceProvokingVertexFeaturesEXT *features =
1597 (VkPhysicalDeviceProvokingVertexFeaturesEXT *)ext;
1598 features->provokingVertexLast = true;
1599 features->transformFeedbackPreservesProvokingVertex = true;
1602 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_2_FEATURES_EXT: {
1603 VkPhysicalDeviceExtendedDynamicState2FeaturesEXT *features =
1604 (VkPhysicalDeviceExtendedDynamicState2FeaturesEXT *)ext;
1605 features->extendedDynamicState2 = true;
1606 features->extendedDynamicState2LogicOp = true;
1607 features->extendedDynamicState2PatchControlPoints = false;
1610 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_GLOBAL_PRIORITY_QUERY_FEATURES_EXT: {
1611 VkPhysicalDeviceGlobalPriorityQueryFeaturesEXT *features =
1612 (VkPhysicalDeviceGlobalPriorityQueryFeaturesEXT *)ext;
1613 features->globalPriorityQuery = true;
1616 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ACCELERATION_STRUCTURE_FEATURES_KHR: {
1617 VkPhysicalDeviceAccelerationStructureFeaturesKHR *features =
1618 (VkPhysicalDeviceAccelerationStructureFeaturesKHR *)ext;
1619 features->accelerationStructure = true;
1620 features->accelerationStructureCaptureReplay = false;
1621 features->accelerationStructureIndirectBuild = false;
1622 features->accelerationStructureHostCommands = true;
1623 features->descriptorBindingAccelerationStructureUpdateAfterBind = true;
1626 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_UNIFORM_CONTROL_FLOW_FEATURES_KHR: {
1627 VkPhysicalDeviceShaderSubgroupUniformControlFlowFeaturesKHR *features =
1628 (VkPhysicalDeviceShaderSubgroupUniformControlFlowFeaturesKHR *)ext;
1629 features->shaderSubgroupUniformControlFlow = true;
1632 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTI_DRAW_FEATURES_EXT: {
1633 VkPhysicalDeviceMultiDrawFeaturesEXT *features = (VkPhysicalDeviceMultiDrawFeaturesEXT *)ext;
1634 features->multiDraw = true;
1637 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COLOR_WRITE_ENABLE_FEATURES_EXT: {
1638 VkPhysicalDeviceColorWriteEnableFeaturesEXT *features =
1639 (VkPhysicalDeviceColorWriteEnableFeaturesEXT *)ext;
1640 features->colorWriteEnable = true;
1643 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_2_FEATURES_EXT: {
1644 VkPhysicalDeviceShaderAtomicFloat2FeaturesEXT *features =
1645 (VkPhysicalDeviceShaderAtomicFloat2FeaturesEXT *)ext;
1646 bool has_shader_buffer_float_minmax = radv_has_shader_buffer_float_minmax(pdevice);
1647 bool has_shader_image_float_minmax =
1648 pdevice->rad_info.gfx_level != GFX8 && pdevice->rad_info.gfx_level != GFX9;
1649 features->shaderBufferFloat16Atomics = false;
1650 features->shaderBufferFloat16AtomicAdd = false;
1651 features->shaderBufferFloat16AtomicMinMax = false;
1652 features->shaderBufferFloat32AtomicMinMax = has_shader_buffer_float_minmax;
1653 features->shaderBufferFloat64AtomicMinMax = has_shader_buffer_float_minmax;
1654 features->shaderSharedFloat16Atomics = false;
1655 features->shaderSharedFloat16AtomicAdd = false;
1656 features->shaderSharedFloat16AtomicMinMax = false;
1657 features->shaderSharedFloat32AtomicMinMax = true;
1658 features->shaderSharedFloat64AtomicMinMax = true;
1659 features->shaderImageFloat32AtomicMinMax = has_shader_image_float_minmax;
1660 features->sparseImageFloat32AtomicMinMax = has_shader_image_float_minmax;
1663 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRIMITIVE_TOPOLOGY_LIST_RESTART_FEATURES_EXT: {
1664 VkPhysicalDevicePrimitiveTopologyListRestartFeaturesEXT *features =
1665 (VkPhysicalDevicePrimitiveTopologyListRestartFeaturesEXT *)ext;
1666 features->primitiveTopologyListRestart = true;
1667 features->primitiveTopologyPatchListRestart = false;
1670 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_RAY_QUERY_FEATURES_KHR: {
1671 VkPhysicalDeviceRayQueryFeaturesKHR *features =
1672 (VkPhysicalDeviceRayQueryFeaturesKHR *)ext;
1673 features->rayQuery = true;
1676 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_RAY_TRACING_PIPELINE_FEATURES_KHR: {
1677 VkPhysicalDeviceRayTracingPipelineFeaturesKHR *features =
1678 (VkPhysicalDeviceRayTracingPipelineFeaturesKHR *)ext;
1679 features->rayTracingPipeline = true;
1680 features->rayTracingPipelineShaderGroupHandleCaptureReplay = false;
1681 features->rayTracingPipelineShaderGroupHandleCaptureReplayMixed = false;
1682 features->rayTracingPipelineTraceRaysIndirect = true;
1683 features->rayTraversalPrimitiveCulling = true;
1686 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_RAY_TRACING_MAINTENANCE_1_FEATURES_KHR: {
1687 VkPhysicalDeviceRayTracingMaintenance1FeaturesKHR *features =
1688 (VkPhysicalDeviceRayTracingMaintenance1FeaturesKHR *)ext;
1689 features->rayTracingMaintenance1 = true;
1690 features->rayTracingPipelineTraceRaysIndirect2 = radv_enable_rt(pdevice, true);
1693 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_FEATURES: {
1694 VkPhysicalDeviceMaintenance4Features *features =
1695 (VkPhysicalDeviceMaintenance4Features *)ext;
1696 features->maintenance4 = true;
1699 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT: {
1700 VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT *features =
1701 (VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT *)ext;
1702 features->vertexInputDynamicState = true;
1705 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_VIEW_MIN_LOD_FEATURES_EXT: {
1706 VkPhysicalDeviceImageViewMinLodFeaturesEXT *features =
1707 (VkPhysicalDeviceImageViewMinLodFeaturesEXT *)ext;
1708 features->minLod = true;
1711 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SYNCHRONIZATION_2_FEATURES: {
1712 VkPhysicalDeviceSynchronization2Features *features =
1713 (VkPhysicalDeviceSynchronization2Features *)ext;
1714 features->synchronization2 = true;
1717 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DYNAMIC_RENDERING_FEATURES: {
1718 VkPhysicalDeviceDynamicRenderingFeatures *features =
1719 (VkPhysicalDeviceDynamicRenderingFeatures *)ext;
1720 features->dynamicRendering = true;
1723 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MESH_SHADER_FEATURES_NV: {
1724 VkPhysicalDeviceMeshShaderFeaturesNV *features =
1725 (VkPhysicalDeviceMeshShaderFeaturesNV *)ext;
1726 features->meshShader = true;
1727 features->taskShader = false; /* TODO */
1730 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXTURE_COMPRESSION_ASTC_HDR_FEATURES: {
1731 VkPhysicalDeviceTextureCompressionASTCHDRFeatures *features =
1732 (VkPhysicalDeviceTextureCompressionASTCHDRFeatures *)ext;
1733 features->textureCompressionASTC_HDR = false;
1736 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_SET_HOST_MAPPING_FEATURES_VALVE: {
1737 VkPhysicalDeviceDescriptorSetHostMappingFeaturesVALVE *features =
1738 (VkPhysicalDeviceDescriptorSetHostMappingFeaturesVALVE *)ext;
1739 features->descriptorSetHostMapping = true;
1742 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_CONTROL_FEATURES_EXT: {
1743 VkPhysicalDeviceDepthClipControlFeaturesEXT *features =
1744 (VkPhysicalDeviceDepthClipControlFeaturesEXT *)ext;
1745 features->depthClipControl = true;
1748 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_2D_VIEW_OF_3D_FEATURES_EXT: {
1749 VkPhysicalDeviceImage2DViewOf3DFeaturesEXT *features =
1750 (VkPhysicalDeviceImage2DViewOf3DFeaturesEXT *)ext;
1751 features->image2DViewOf3D = true;
1752 features->sampler2DViewOf3D = false;
1755 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_INTEGER_FUNCTIONS_2_FEATURES_INTEL: {
1756 VkPhysicalDeviceShaderIntegerFunctions2FeaturesINTEL *features =
1757 (VkPhysicalDeviceShaderIntegerFunctions2FeaturesINTEL *)ext;
1758 features->shaderIntegerFunctions2 = true;
1761 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRIMITIVES_GENERATED_QUERY_FEATURES_EXT: {
1762 VkPhysicalDevicePrimitivesGeneratedQueryFeaturesEXT *features =
1763 (VkPhysicalDevicePrimitivesGeneratedQueryFeaturesEXT *)ext;
1764 features->primitivesGeneratedQuery = true;
1765 features->primitivesGeneratedQueryWithRasterizerDiscard = true;
1766 features->primitivesGeneratedQueryWithNonZeroStreams = true;
1769 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_NON_SEAMLESS_CUBE_MAP_FEATURES_EXT : {
1770 VkPhysicalDeviceNonSeamlessCubeMapFeaturesEXT *features =
1771 (VkPhysicalDeviceNonSeamlessCubeMapFeaturesEXT *)ext;
1772 features->nonSeamlessCubeMap = true;
1775 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BORDER_COLOR_SWIZZLE_FEATURES_EXT: {
1776 VkPhysicalDeviceBorderColorSwizzleFeaturesEXT *features =
1777 (VkPhysicalDeviceBorderColorSwizzleFeaturesEXT *)ext;
1778 features->borderColorSwizzle = true;
1779 features->borderColorSwizzleFromImage = true;
1782 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MODULE_IDENTIFIER_FEATURES_EXT: {
1783 VkPhysicalDeviceShaderModuleIdentifierFeaturesEXT *features =
1784 (VkPhysicalDeviceShaderModuleIdentifierFeaturesEXT *)ext;
1785 features->shaderModuleIdentifier = true;
1795 radv_max_descriptor_set_size()
1797 /* make sure that the entire descriptor set is addressable with a signed
1798 * 32-bit int. So the sum of all limits scaled by descriptor size has to
1799 * be at most 2 GiB. the combined image & samples object count as one of
1800 * both. This limit is for the pipeline layout, not for the set layout, but
1801 * there is no set limit, so we just set a pipeline limit. I don't think
1802 * any app is going to hit this soon. */
1803 return ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS -
1804 MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_INLINE_UNIFORM_BLOCK_COUNT) /
1805 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1806 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1807 32 /* sampler, largest when combined with image */ + 64 /* sampled image */ +
1808 64 /* storage image */);
1812 radv_uniform_buffer_offset_alignment(const struct radv_physical_device *pdevice)
1814 uint32_t uniform_offset_alignment =
1815 driQueryOptioni(&pdevice->instance->dri_options, "radv_override_uniform_offset_alignment");
1816 if (!util_is_power_of_two_or_zero(uniform_offset_alignment)) {
1818 "ERROR: invalid radv_override_uniform_offset_alignment setting %d:"
1819 "not a power of two\n",
1820 uniform_offset_alignment);
1821 uniform_offset_alignment = 0;
1824 /* Take at least the hardware limit. */
1825 return MAX2(uniform_offset_alignment, 4);
1828 VKAPI_ATTR void VKAPI_CALL
1829 radv_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,
1830 VkPhysicalDeviceProperties *pProperties)
1832 RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
1833 VkSampleCountFlags sample_counts = 0xf;
1835 size_t max_descriptor_set_size = radv_max_descriptor_set_size();
1837 VkPhysicalDeviceLimits limits = {
1838 .maxImageDimension1D = (1 << 14),
1839 .maxImageDimension2D = (1 << 14),
1840 .maxImageDimension3D = (1 << 11),
1841 .maxImageDimensionCube = (1 << 14),
1842 .maxImageArrayLayers = (1 << 11),
1843 .maxTexelBufferElements = UINT32_MAX,
1844 .maxUniformBufferRange = UINT32_MAX,
1845 .maxStorageBufferRange = UINT32_MAX,
1846 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1847 .maxMemoryAllocationCount = UINT32_MAX,
1848 .maxSamplerAllocationCount = 64 * 1024,
1849 .bufferImageGranularity = 1,
1850 .sparseAddressSpaceSize = RADV_MAX_MEMORY_ALLOCATION_SIZE, /* buffer max size */
1851 .maxBoundDescriptorSets = MAX_SETS,
1852 .maxPerStageDescriptorSamplers = max_descriptor_set_size,
1853 .maxPerStageDescriptorUniformBuffers = max_descriptor_set_size,
1854 .maxPerStageDescriptorStorageBuffers = max_descriptor_set_size,
1855 .maxPerStageDescriptorSampledImages = max_descriptor_set_size,
1856 .maxPerStageDescriptorStorageImages = max_descriptor_set_size,
1857 .maxPerStageDescriptorInputAttachments = max_descriptor_set_size,
1858 .maxPerStageResources = max_descriptor_set_size,
1859 .maxDescriptorSetSamplers = max_descriptor_set_size,
1860 .maxDescriptorSetUniformBuffers = max_descriptor_set_size,
1861 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS,
1862 .maxDescriptorSetStorageBuffers = max_descriptor_set_size,
1863 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS,
1864 .maxDescriptorSetSampledImages = max_descriptor_set_size,
1865 .maxDescriptorSetStorageImages = max_descriptor_set_size,
1866 .maxDescriptorSetInputAttachments = max_descriptor_set_size,
1867 .maxVertexInputAttributes = MAX_VERTEX_ATTRIBS,
1868 .maxVertexInputBindings = MAX_VBS,
1869 .maxVertexInputAttributeOffset = UINT32_MAX,
1870 .maxVertexInputBindingStride = 2048,
1871 .maxVertexOutputComponents = 128,
1872 .maxTessellationGenerationLevel = 64,
1873 .maxTessellationPatchSize = 32,
1874 .maxTessellationControlPerVertexInputComponents = 128,
1875 .maxTessellationControlPerVertexOutputComponents = 128,
1876 .maxTessellationControlPerPatchOutputComponents = 120,
1877 .maxTessellationControlTotalOutputComponents = 4096,
1878 .maxTessellationEvaluationInputComponents = 128,
1879 .maxTessellationEvaluationOutputComponents = 128,
1880 .maxGeometryShaderInvocations = 127,
1881 .maxGeometryInputComponents = 64,
1882 .maxGeometryOutputComponents = 128,
1883 .maxGeometryOutputVertices = 256,
1884 .maxGeometryTotalOutputComponents = 1024,
1885 .maxFragmentInputComponents = 128,
1886 .maxFragmentOutputAttachments = 8,
1887 .maxFragmentDualSrcAttachments = 1,
1888 .maxFragmentCombinedOutputResources = max_descriptor_set_size,
1889 .maxComputeSharedMemorySize = pdevice->rad_info.gfx_level >= GFX7 ? 65536 : 32768,
1890 .maxComputeWorkGroupCount = {65535, 65535, 65535},
1891 .maxComputeWorkGroupInvocations = 1024,
1892 .maxComputeWorkGroupSize = {1024, 1024, 1024},
1893 .subPixelPrecisionBits = 8,
1894 .subTexelPrecisionBits = 8,
1895 .mipmapPrecisionBits = 8,
1896 .maxDrawIndexedIndexValue = UINT32_MAX,
1897 .maxDrawIndirectCount = UINT32_MAX,
1898 .maxSamplerLodBias = 16,
1899 .maxSamplerAnisotropy = 16,
1900 .maxViewports = MAX_VIEWPORTS,
1901 .maxViewportDimensions = {(1 << 14), (1 << 14)},
1902 .viewportBoundsRange = {INT16_MIN, INT16_MAX},
1903 .viewportSubPixelBits = 8,
1904 .minMemoryMapAlignment = 4096, /* A page */
1905 .minTexelBufferOffsetAlignment = 4,
1906 .minUniformBufferOffsetAlignment = radv_uniform_buffer_offset_alignment(pdevice),
1907 .minStorageBufferOffsetAlignment = 4,
1908 .minTexelOffset = -32,
1909 .maxTexelOffset = 31,
1910 .minTexelGatherOffset = -32,
1911 .maxTexelGatherOffset = 31,
1912 .minInterpolationOffset = -2,
1913 .maxInterpolationOffset = 2,
1914 .subPixelInterpolationOffsetBits = 8,
1915 .maxFramebufferWidth = MAX_FRAMEBUFFER_WIDTH,
1916 .maxFramebufferHeight = MAX_FRAMEBUFFER_HEIGHT,
1917 .maxFramebufferLayers = (1 << 10),
1918 .framebufferColorSampleCounts = sample_counts,
1919 .framebufferDepthSampleCounts = sample_counts,
1920 .framebufferStencilSampleCounts = sample_counts,
1921 .framebufferNoAttachmentsSampleCounts = sample_counts,
1922 .maxColorAttachments = MAX_RTS,
1923 .sampledImageColorSampleCounts = sample_counts,
1924 .sampledImageIntegerSampleCounts = sample_counts,
1925 .sampledImageDepthSampleCounts = sample_counts,
1926 .sampledImageStencilSampleCounts = sample_counts,
1927 .storageImageSampleCounts = sample_counts,
1928 .maxSampleMaskWords = 1,
1929 .timestampComputeAndGraphics = true,
1930 .timestampPeriod = 1000000.0 / pdevice->rad_info.clock_crystal_freq,
1931 .maxClipDistances = 8,
1932 .maxCullDistances = 8,
1933 .maxCombinedClipAndCullDistances = 8,
1934 .discreteQueuePriorities = 2,
1935 .pointSizeRange = {0.0, 8191.875},
1936 .lineWidthRange = {0.0, 8191.875},
1937 .pointSizeGranularity = (1.0 / 8.0),
1938 .lineWidthGranularity = (1.0 / 8.0),
1939 .strictLines = false, /* FINISHME */
1940 .standardSampleLocations = true,
1941 .optimalBufferCopyOffsetAlignment = 1,
1942 .optimalBufferCopyRowPitchAlignment = 1,
1943 .nonCoherentAtomSize = 64,
1946 VkPhysicalDeviceType device_type;
1948 if (pdevice->rad_info.has_dedicated_vram || pdevice->instance->report_apu_as_dgpu) {
1949 device_type = VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU;
1951 device_type = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU;
1954 *pProperties = (VkPhysicalDeviceProperties){
1955 .apiVersion = RADV_API_VERSION,
1956 .driverVersion = vk_get_driver_version(),
1957 .vendorID = ATI_VENDOR_ID,
1958 .deviceID = pdevice->rad_info.pci_id,
1959 .deviceType = device_type,
1963 .residencyNonResidentStrict = pdevice->rad_info.family >= CHIP_POLARIS10,
1964 .residencyStandard2DBlockShape = pdevice->rad_info.family >= CHIP_POLARIS10,
1968 strcpy(pProperties->deviceName, pdevice->marketing_name);
1969 memcpy(pProperties->pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
1973 radv_get_physical_device_properties_1_1(struct radv_physical_device *pdevice,
1974 VkPhysicalDeviceVulkan11Properties *p)
1976 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES);
1978 memcpy(p->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1979 memcpy(p->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1980 memset(p->deviceLUID, 0, VK_LUID_SIZE);
1981 /* The LUID is for Windows. */
1982 p->deviceLUIDValid = false;
1983 p->deviceNodeMask = 0;
1985 p->subgroupSize = RADV_SUBGROUP_SIZE;
1986 p->subgroupSupportedStages = VK_SHADER_STAGE_ALL_GRAPHICS | VK_SHADER_STAGE_COMPUTE_BIT;
1987 p->subgroupSupportedOperations =
1988 VK_SUBGROUP_FEATURE_BASIC_BIT | VK_SUBGROUP_FEATURE_VOTE_BIT |
1989 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT | VK_SUBGROUP_FEATURE_BALLOT_BIT |
1990 VK_SUBGROUP_FEATURE_CLUSTERED_BIT | VK_SUBGROUP_FEATURE_QUAD_BIT |
1991 VK_SUBGROUP_FEATURE_SHUFFLE_BIT | VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT;
1992 p->subgroupQuadOperationsInAllStages = true;
1994 p->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
1995 p->maxMultiviewViewCount = MAX_VIEWS;
1996 p->maxMultiviewInstanceIndex = INT_MAX;
1997 p->protectedNoFault = false;
1998 p->maxPerSetDescriptors = RADV_MAX_PER_SET_DESCRIPTORS;
1999 p->maxMemoryAllocationSize = RADV_MAX_MEMORY_ALLOCATION_SIZE;
2003 radv_get_physical_device_properties_1_2(struct radv_physical_device *pdevice,
2004 VkPhysicalDeviceVulkan12Properties *p)
2006 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES);
2008 p->driverID = VK_DRIVER_ID_MESA_RADV;
2009 snprintf(p->driverName, VK_MAX_DRIVER_NAME_SIZE, "radv");
2010 snprintf(p->driverInfo, VK_MAX_DRIVER_INFO_SIZE, "Mesa " PACKAGE_VERSION MESA_GIT_SHA1 "%s",
2011 radv_get_compiler_string(pdevice));
2013 if (radv_is_conformant(pdevice)) {
2014 if (pdevice->rad_info.gfx_level >= GFX10_3) {
2015 p->conformanceVersion = (VkConformanceVersion){
2022 p->conformanceVersion = (VkConformanceVersion){
2030 p->conformanceVersion = (VkConformanceVersion){
2038 /* On AMD hardware, denormals and rounding modes for fp16/fp64 are
2039 * controlled by the same config register.
2041 if (pdevice->rad_info.has_packed_math_16bit) {
2042 p->denormBehaviorIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY;
2043 p->roundingModeIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY;
2045 p->denormBehaviorIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL;
2046 p->roundingModeIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL;
2049 /* With LLVM, do not allow both preserving and flushing denorms because
2050 * different shaders in the same pipeline can have different settings and
2051 * this won't work for merged shaders. To make it work, this requires LLVM
2052 * support for changing the register. The same logic applies for the
2053 * rounding modes because they are configured with the same config
2056 p->shaderDenormFlushToZeroFloat32 = true;
2057 p->shaderDenormPreserveFloat32 = !pdevice->use_llvm;
2058 p->shaderRoundingModeRTEFloat32 = true;
2059 p->shaderRoundingModeRTZFloat32 = !pdevice->use_llvm;
2060 p->shaderSignedZeroInfNanPreserveFloat32 = true;
2062 p->shaderDenormFlushToZeroFloat16 =
2063 pdevice->rad_info.has_packed_math_16bit && !pdevice->use_llvm;
2064 p->shaderDenormPreserveFloat16 = pdevice->rad_info.has_packed_math_16bit;
2065 p->shaderRoundingModeRTEFloat16 = pdevice->rad_info.has_packed_math_16bit;
2066 p->shaderRoundingModeRTZFloat16 = pdevice->rad_info.has_packed_math_16bit && !pdevice->use_llvm;
2067 p->shaderSignedZeroInfNanPreserveFloat16 = pdevice->rad_info.has_packed_math_16bit;
2069 p->shaderDenormFlushToZeroFloat64 = pdevice->rad_info.gfx_level >= GFX8 && !pdevice->use_llvm;
2070 p->shaderDenormPreserveFloat64 = pdevice->rad_info.gfx_level >= GFX8;
2071 p->shaderRoundingModeRTEFloat64 = pdevice->rad_info.gfx_level >= GFX8;
2072 p->shaderRoundingModeRTZFloat64 = pdevice->rad_info.gfx_level >= GFX8 && !pdevice->use_llvm;
2073 p->shaderSignedZeroInfNanPreserveFloat64 = pdevice->rad_info.gfx_level >= GFX8;
2075 p->maxUpdateAfterBindDescriptorsInAllPools = UINT32_MAX / 64;
2076 p->shaderUniformBufferArrayNonUniformIndexingNative = false;
2077 p->shaderSampledImageArrayNonUniformIndexingNative = false;
2078 p->shaderStorageBufferArrayNonUniformIndexingNative = false;
2079 p->shaderStorageImageArrayNonUniformIndexingNative = false;
2080 p->shaderInputAttachmentArrayNonUniformIndexingNative = false;
2081 p->robustBufferAccessUpdateAfterBind = true;
2082 p->quadDivergentImplicitLod = false;
2084 size_t max_descriptor_set_size =
2085 ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS -
2086 MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_INLINE_UNIFORM_BLOCK_COUNT) /
2087 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
2088 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
2089 32 /* sampler, largest when combined with image */ + 64 /* sampled image */ +
2090 64 /* storage image */);
2091 p->maxPerStageDescriptorUpdateAfterBindSamplers = max_descriptor_set_size;
2092 p->maxPerStageDescriptorUpdateAfterBindUniformBuffers = max_descriptor_set_size;
2093 p->maxPerStageDescriptorUpdateAfterBindStorageBuffers = max_descriptor_set_size;
2094 p->maxPerStageDescriptorUpdateAfterBindSampledImages = max_descriptor_set_size;
2095 p->maxPerStageDescriptorUpdateAfterBindStorageImages = max_descriptor_set_size;
2096 p->maxPerStageDescriptorUpdateAfterBindInputAttachments = max_descriptor_set_size;
2097 p->maxPerStageUpdateAfterBindResources = max_descriptor_set_size;
2098 p->maxDescriptorSetUpdateAfterBindSamplers = max_descriptor_set_size;
2099 p->maxDescriptorSetUpdateAfterBindUniformBuffers = max_descriptor_set_size;
2100 p->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS;
2101 p->maxDescriptorSetUpdateAfterBindStorageBuffers = max_descriptor_set_size;
2102 p->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS;
2103 p->maxDescriptorSetUpdateAfterBindSampledImages = max_descriptor_set_size;
2104 p->maxDescriptorSetUpdateAfterBindStorageImages = max_descriptor_set_size;
2105 p->maxDescriptorSetUpdateAfterBindInputAttachments = max_descriptor_set_size;
2107 /* We support all of the depth resolve modes */
2108 p->supportedDepthResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT |
2109 VK_RESOLVE_MODE_AVERAGE_BIT | VK_RESOLVE_MODE_MIN_BIT |
2110 VK_RESOLVE_MODE_MAX_BIT;
2112 /* Average doesn't make sense for stencil so we don't support that */
2113 p->supportedStencilResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT |
2114 VK_RESOLVE_MODE_MIN_BIT | VK_RESOLVE_MODE_MAX_BIT;
2116 p->independentResolveNone = true;
2117 p->independentResolve = true;
2119 /* GFX6-8 only support single channel min/max filter. */
2120 p->filterMinmaxImageComponentMapping = pdevice->rad_info.gfx_level >= GFX9;
2121 p->filterMinmaxSingleComponentFormats = true;
2123 p->maxTimelineSemaphoreValueDifference = UINT64_MAX;
2125 p->framebufferIntegerColorSampleCounts = VK_SAMPLE_COUNT_1_BIT;
2129 radv_get_physical_device_properties_1_3(struct radv_physical_device *pdevice,
2130 VkPhysicalDeviceVulkan13Properties *p)
2132 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_PROPERTIES);
2134 p->minSubgroupSize = 64;
2135 p->maxSubgroupSize = 64;
2136 p->maxComputeWorkgroupSubgroups = UINT32_MAX;
2137 p->requiredSubgroupSizeStages = 0;
2138 if (pdevice->rad_info.gfx_level >= GFX10) {
2139 /* Only GFX10+ supports wave32. */
2140 p->minSubgroupSize = 32;
2141 p->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
2144 p->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
2145 p->maxPerStageDescriptorInlineUniformBlocks = MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_SETS;
2146 p->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks = MAX_INLINE_UNIFORM_BLOCK_SIZE * MAX_SETS;
2147 p->maxDescriptorSetInlineUniformBlocks = MAX_INLINE_UNIFORM_BLOCK_COUNT;
2148 p->maxDescriptorSetUpdateAfterBindInlineUniformBlocks = MAX_INLINE_UNIFORM_BLOCK_COUNT;
2149 p->maxInlineUniformTotalSize = UINT16_MAX;
2151 bool accel = pdevice->rad_info.has_accelerated_dot_product;
2152 p->integerDotProduct8BitUnsignedAccelerated = accel;
2153 p->integerDotProduct8BitSignedAccelerated = accel;
2154 p->integerDotProduct8BitMixedSignednessAccelerated = false;
2155 p->integerDotProduct4x8BitPackedUnsignedAccelerated = accel;
2156 p->integerDotProduct4x8BitPackedSignedAccelerated = accel;
2157 p->integerDotProduct4x8BitPackedMixedSignednessAccelerated = false;
2158 p->integerDotProduct16BitUnsignedAccelerated = accel;
2159 p->integerDotProduct16BitSignedAccelerated = accel;
2160 p->integerDotProduct16BitMixedSignednessAccelerated = false;
2161 p->integerDotProduct32BitUnsignedAccelerated = false;
2162 p->integerDotProduct32BitSignedAccelerated = false;
2163 p->integerDotProduct32BitMixedSignednessAccelerated = false;
2164 p->integerDotProduct64BitUnsignedAccelerated = false;
2165 p->integerDotProduct64BitSignedAccelerated = false;
2166 p->integerDotProduct64BitMixedSignednessAccelerated = false;
2167 p->integerDotProductAccumulatingSaturating8BitUnsignedAccelerated = accel;
2168 p->integerDotProductAccumulatingSaturating8BitSignedAccelerated = accel;
2169 p->integerDotProductAccumulatingSaturating8BitMixedSignednessAccelerated = false;
2170 p->integerDotProductAccumulatingSaturating4x8BitPackedUnsignedAccelerated = accel;
2171 p->integerDotProductAccumulatingSaturating4x8BitPackedSignedAccelerated = accel;
2172 p->integerDotProductAccumulatingSaturating4x8BitPackedMixedSignednessAccelerated = false;
2173 p->integerDotProductAccumulatingSaturating16BitUnsignedAccelerated = accel;
2174 p->integerDotProductAccumulatingSaturating16BitSignedAccelerated = accel;
2175 p->integerDotProductAccumulatingSaturating16BitMixedSignednessAccelerated = false;
2176 p->integerDotProductAccumulatingSaturating32BitUnsignedAccelerated = false;
2177 p->integerDotProductAccumulatingSaturating32BitSignedAccelerated = false;
2178 p->integerDotProductAccumulatingSaturating32BitMixedSignednessAccelerated = false;
2179 p->integerDotProductAccumulatingSaturating64BitUnsignedAccelerated = false;
2180 p->integerDotProductAccumulatingSaturating64BitSignedAccelerated = false;
2181 p->integerDotProductAccumulatingSaturating64BitMixedSignednessAccelerated = false;
2183 p->storageTexelBufferOffsetAlignmentBytes = 4;
2184 p->storageTexelBufferOffsetSingleTexelAlignment = true;
2185 p->uniformTexelBufferOffsetAlignmentBytes = 4;
2186 p->uniformTexelBufferOffsetSingleTexelAlignment = true;
2188 p->maxBufferSize = RADV_MAX_MEMORY_ALLOCATION_SIZE;
2191 VKAPI_ATTR void VKAPI_CALL
2192 radv_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
2193 VkPhysicalDeviceProperties2 *pProperties)
2195 RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
2196 radv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
2198 VkPhysicalDeviceVulkan11Properties core_1_1 = {
2199 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES,
2201 radv_get_physical_device_properties_1_1(pdevice, &core_1_1);
2203 VkPhysicalDeviceVulkan12Properties core_1_2 = {
2204 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES,
2206 radv_get_physical_device_properties_1_2(pdevice, &core_1_2);
2208 VkPhysicalDeviceVulkan13Properties core_1_3 = {
2209 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_PROPERTIES,
2211 radv_get_physical_device_properties_1_3(pdevice, &core_1_3);
2213 vk_foreach_struct(ext, pProperties->pNext)
2215 if (vk_get_physical_device_core_1_1_property_ext(ext, &core_1_1))
2217 if (vk_get_physical_device_core_1_2_property_ext(ext, &core_1_2))
2219 if (vk_get_physical_device_core_1_3_property_ext(ext, &core_1_3))
2222 switch (ext->sType) {
2223 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
2224 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
2225 (VkPhysicalDevicePushDescriptorPropertiesKHR *)ext;
2226 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
2229 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT: {
2230 VkPhysicalDeviceDiscardRectanglePropertiesEXT *properties =
2231 (VkPhysicalDeviceDiscardRectanglePropertiesEXT *)ext;
2232 properties->maxDiscardRectangles = MAX_DISCARD_RECTANGLES;
2235 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
2236 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *properties =
2237 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *)ext;
2238 properties->minImportedHostPointerAlignment = 4096;
2241 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD: {
2242 VkPhysicalDeviceShaderCorePropertiesAMD *properties =
2243 (VkPhysicalDeviceShaderCorePropertiesAMD *)ext;
2245 /* Shader engines. */
2246 properties->shaderEngineCount = pdevice->rad_info.max_se;
2247 properties->shaderArraysPerEngineCount = pdevice->rad_info.max_sa_per_se;
2248 properties->computeUnitsPerShaderArray = pdevice->rad_info.min_good_cu_per_sa;
2249 properties->simdPerComputeUnit = pdevice->rad_info.num_simd_per_compute_unit;
2250 properties->wavefrontsPerSimd = pdevice->rad_info.max_wave64_per_simd;
2251 properties->wavefrontSize = 64;
2254 properties->sgprsPerSimd = pdevice->rad_info.num_physical_sgprs_per_simd;
2255 properties->minSgprAllocation = pdevice->rad_info.min_sgpr_alloc;
2256 properties->maxSgprAllocation = pdevice->rad_info.max_sgpr_alloc;
2257 properties->sgprAllocationGranularity = pdevice->rad_info.sgpr_alloc_granularity;
2260 properties->vgprsPerSimd = pdevice->rad_info.num_physical_wave64_vgprs_per_simd;
2261 properties->minVgprAllocation = pdevice->rad_info.min_wave64_vgpr_alloc;
2262 properties->maxVgprAllocation = pdevice->rad_info.max_vgpr_alloc;
2263 properties->vgprAllocationGranularity = pdevice->rad_info.wave64_vgpr_alloc_granularity;
2266 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_2_AMD: {
2267 VkPhysicalDeviceShaderCoreProperties2AMD *properties =
2268 (VkPhysicalDeviceShaderCoreProperties2AMD *)ext;
2270 properties->shaderCoreFeatures = 0;
2271 properties->activeComputeUnitCount = pdevice->rad_info.num_good_compute_units;
2274 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
2275 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *properties =
2276 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
2277 properties->maxVertexAttribDivisor = UINT32_MAX;
2280 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONSERVATIVE_RASTERIZATION_PROPERTIES_EXT: {
2281 VkPhysicalDeviceConservativeRasterizationPropertiesEXT *properties =
2282 (VkPhysicalDeviceConservativeRasterizationPropertiesEXT *)ext;
2283 properties->primitiveOverestimationSize = 0;
2284 properties->maxExtraPrimitiveOverestimationSize = 0;
2285 properties->extraPrimitiveOverestimationSizeGranularity = 0;
2286 properties->primitiveUnderestimation = false;
2287 properties->conservativePointAndLineRasterization = false;
2288 properties->degenerateTrianglesRasterized = true;
2289 properties->degenerateLinesRasterized = false;
2290 properties->fullyCoveredFragmentShaderInputVariable = false;
2291 properties->conservativeRasterizationPostDepthCoverage = false;
2295 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
2296 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
2297 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
2298 properties->pciDomain = pdevice->bus_info.domain;
2299 properties->pciBus = pdevice->bus_info.bus;
2300 properties->pciDevice = pdevice->bus_info.dev;
2301 properties->pciFunction = pdevice->bus_info.func;
2305 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
2306 VkPhysicalDeviceTransformFeedbackPropertiesEXT *properties =
2307 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
2308 properties->maxTransformFeedbackStreams = MAX_SO_STREAMS;
2309 properties->maxTransformFeedbackBuffers = MAX_SO_BUFFERS;
2310 properties->maxTransformFeedbackBufferSize = UINT32_MAX;
2311 properties->maxTransformFeedbackStreamDataSize = 512;
2312 properties->maxTransformFeedbackBufferDataSize = 512;
2313 properties->maxTransformFeedbackBufferDataStride = 512;
2314 properties->transformFeedbackQueries = !pdevice->use_ngg_streamout;
2315 properties->transformFeedbackStreamsLinesTriangles = !pdevice->use_ngg_streamout;
2316 properties->transformFeedbackRasterizationStreamSelect = false;
2317 properties->transformFeedbackDraw = true;
2320 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLE_LOCATIONS_PROPERTIES_EXT: {
2321 VkPhysicalDeviceSampleLocationsPropertiesEXT *properties =
2322 (VkPhysicalDeviceSampleLocationsPropertiesEXT *)ext;
2323 properties->sampleLocationSampleCounts = VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT |
2324 VK_SAMPLE_COUNT_8_BIT;
2325 properties->maxSampleLocationGridSize = (VkExtent2D){2, 2};
2326 properties->sampleLocationCoordinateRange[0] = 0.0f;
2327 properties->sampleLocationCoordinateRange[1] = 0.9375f;
2328 properties->sampleLocationSubPixelBits = 4;
2329 properties->variableSampleLocations = false;
2332 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
2333 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
2334 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
2335 props->lineSubPixelPrecisionBits = 4;
2338 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_PROPERTIES_EXT: {
2339 VkPhysicalDeviceRobustness2PropertiesEXT *properties =
2340 (VkPhysicalDeviceRobustness2PropertiesEXT *)ext;
2341 properties->robustStorageBufferAccessSizeAlignment = 4;
2342 properties->robustUniformBufferAccessSizeAlignment = 4;
2345 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_PROPERTIES_EXT: {
2346 VkPhysicalDeviceCustomBorderColorPropertiesEXT *props =
2347 (VkPhysicalDeviceCustomBorderColorPropertiesEXT *)ext;
2348 props->maxCustomBorderColorSamplers = RADV_BORDER_COLOR_COUNT;
2351 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADING_RATE_PROPERTIES_KHR: {
2352 VkPhysicalDeviceFragmentShadingRatePropertiesKHR *props =
2353 (VkPhysicalDeviceFragmentShadingRatePropertiesKHR *)ext;
2354 props->minFragmentShadingRateAttachmentTexelSize = (VkExtent2D){8, 8};
2355 props->maxFragmentShadingRateAttachmentTexelSize = (VkExtent2D){8, 8};
2356 props->maxFragmentShadingRateAttachmentTexelSizeAspectRatio = 1;
2357 props->primitiveFragmentShadingRateWithMultipleViewports = true;
2358 props->layeredShadingRateAttachments = false; /* TODO */
2359 props->fragmentShadingRateNonTrivialCombinerOps = true;
2360 props->maxFragmentSize = (VkExtent2D){2, 2};
2361 props->maxFragmentSizeAspectRatio = 2;
2362 props->maxFragmentShadingRateCoverageSamples = 32;
2363 props->maxFragmentShadingRateRasterizationSamples = VK_SAMPLE_COUNT_8_BIT;
2364 props->fragmentShadingRateWithShaderDepthStencilWrites = false;
2365 props->fragmentShadingRateWithSampleMask = true;
2366 props->fragmentShadingRateWithShaderSampleMask = false;
2367 props->fragmentShadingRateWithConservativeRasterization = true;
2368 props->fragmentShadingRateWithFragmentShaderInterlock = false;
2369 props->fragmentShadingRateWithCustomSampleLocations = false;
2370 props->fragmentShadingRateStrictMultiplyCombiner = true;
2373 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_PROPERTIES_EXT: {
2374 VkPhysicalDeviceProvokingVertexPropertiesEXT *props =
2375 (VkPhysicalDeviceProvokingVertexPropertiesEXT *)ext;
2376 props->provokingVertexModePerPipeline = true;
2377 props->transformFeedbackPreservesTriangleFanProvokingVertex = true;
2380 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ACCELERATION_STRUCTURE_PROPERTIES_KHR: {
2381 VkPhysicalDeviceAccelerationStructurePropertiesKHR *props =
2382 (VkPhysicalDeviceAccelerationStructurePropertiesKHR *)ext;
2383 props->maxGeometryCount = (1 << 24) - 1;
2384 props->maxInstanceCount = (1 << 24) - 1;
2385 props->maxPrimitiveCount = (1 << 29) - 1;
2386 props->maxPerStageDescriptorAccelerationStructures =
2387 pProperties->properties.limits.maxPerStageDescriptorStorageBuffers;
2388 props->maxPerStageDescriptorUpdateAfterBindAccelerationStructures =
2389 pProperties->properties.limits.maxPerStageDescriptorStorageBuffers;
2390 props->maxDescriptorSetAccelerationStructures =
2391 pProperties->properties.limits.maxDescriptorSetStorageBuffers;
2392 props->maxDescriptorSetUpdateAfterBindAccelerationStructures =
2393 pProperties->properties.limits.maxDescriptorSetStorageBuffers;
2394 props->minAccelerationStructureScratchOffsetAlignment = 128;
2398 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRM_PROPERTIES_EXT: {
2399 VkPhysicalDeviceDrmPropertiesEXT *props = (VkPhysicalDeviceDrmPropertiesEXT *)ext;
2400 if (pdevice->available_nodes & (1 << DRM_NODE_PRIMARY)) {
2401 props->hasPrimary = true;
2402 props->primaryMajor = (int64_t)major(pdevice->primary_devid);
2403 props->primaryMinor = (int64_t)minor(pdevice->primary_devid);
2405 props->hasPrimary = false;
2407 if (pdevice->available_nodes & (1 << DRM_NODE_RENDER)) {
2408 props->hasRender = true;
2409 props->renderMajor = (int64_t)major(pdevice->render_devid);
2410 props->renderMinor = (int64_t)minor(pdevice->render_devid);
2412 props->hasRender = false;
2417 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTI_DRAW_PROPERTIES_EXT: {
2418 VkPhysicalDeviceMultiDrawPropertiesEXT *props = (VkPhysicalDeviceMultiDrawPropertiesEXT *)ext;
2419 props->maxMultiDrawCount = 2048;
2422 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_RAY_TRACING_PIPELINE_PROPERTIES_KHR: {
2423 VkPhysicalDeviceRayTracingPipelinePropertiesKHR *props =
2424 (VkPhysicalDeviceRayTracingPipelinePropertiesKHR *)ext;
2425 props->shaderGroupHandleSize = RADV_RT_HANDLE_SIZE;
2426 props->maxRayRecursionDepth = 31; /* Minimum allowed for DXR. */
2427 props->maxShaderGroupStride = 16384; /* dummy */
2428 props->shaderGroupBaseAlignment = 16;
2429 props->shaderGroupHandleCaptureReplaySize = 16;
2430 props->maxRayDispatchInvocationCount = 1024 * 1024 * 64;
2431 props->shaderGroupHandleAlignment = 16;
2432 props->maxRayHitAttributeSize = RADV_MAX_HIT_ATTRIB_SIZE;
2435 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_PROPERTIES: {
2436 VkPhysicalDeviceMaintenance4Properties *properties =
2437 (VkPhysicalDeviceMaintenance4Properties *)ext;
2438 properties->maxBufferSize = RADV_MAX_MEMORY_ALLOCATION_SIZE;
2441 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MESH_SHADER_PROPERTIES_NV: {
2442 VkPhysicalDeviceMeshShaderPropertiesNV *properties =
2443 (VkPhysicalDeviceMeshShaderPropertiesNV *)ext;
2445 /* Task shader limitations:
2446 * Same as compute, because TS are compiled to CS.
2448 properties->maxDrawMeshTasksCount = 65535;
2449 properties->maxTaskTotalMemorySize = 65536;
2450 properties->maxTaskWorkGroupInvocations = 1024;
2451 properties->maxTaskWorkGroupSize[0] = 1024;
2452 properties->maxTaskWorkGroupSize[1] = 1024;
2453 properties->maxTaskWorkGroupSize[2] = 1024;
2454 properties->maxTaskOutputCount = 65535;
2456 /* Mesh shader limitations:
2457 * Same as NGG, because MS are compiled to NGG.
2459 properties->maxMeshMultiviewViewCount = MAX_VIEWS;
2460 properties->maxMeshOutputPrimitives = 256;
2461 properties->maxMeshOutputVertices = 256;
2462 properties->maxMeshTotalMemorySize = 31 * 1024; /* Reserve 1K for prim indices, etc. */
2463 properties->maxMeshWorkGroupInvocations = 256;
2464 properties->maxMeshWorkGroupSize[0] = 256;
2465 properties->maxMeshWorkGroupSize[1] = 256;
2466 properties->maxMeshWorkGroupSize[2] = 256;
2467 properties->meshOutputPerPrimitiveGranularity = 1;
2468 properties->meshOutputPerVertexGranularity = 1;
2472 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MODULE_IDENTIFIER_PROPERTIES_EXT: {
2473 VkPhysicalDeviceShaderModuleIdentifierPropertiesEXT *properties =
2474 (VkPhysicalDeviceShaderModuleIdentifierPropertiesEXT *)ext;
2475 STATIC_ASSERT(sizeof(vk_shaderModuleIdentifierAlgorithmUUID) ==
2476 sizeof(properties->shaderModuleIdentifierAlgorithmUUID));
2477 memcpy(properties->shaderModuleIdentifierAlgorithmUUID,
2478 vk_shaderModuleIdentifierAlgorithmUUID,
2479 sizeof(properties->shaderModuleIdentifierAlgorithmUUID));
2489 radv_get_physical_device_queue_family_properties(struct radv_physical_device *pdevice,
2491 VkQueueFamilyProperties **pQueueFamilyProperties)
2493 int num_queue_families = 1;
2495 if (pdevice->rad_info.ip[AMD_IP_COMPUTE].num_queues > 0 &&
2496 !(pdevice->instance->debug_flags & RADV_DEBUG_NO_COMPUTE_QUEUE))
2497 num_queue_families++;
2499 if (pQueueFamilyProperties == NULL) {
2500 *pCount = num_queue_families;
2509 *pQueueFamilyProperties[idx] = (VkQueueFamilyProperties){
2510 .queueFlags = VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT |
2511 VK_QUEUE_SPARSE_BINDING_BIT,
2513 .timestampValidBits = 64,
2514 .minImageTransferGranularity = (VkExtent3D){1, 1, 1},
2519 if (pdevice->rad_info.ip[AMD_IP_COMPUTE].num_queues > 0 &&
2520 !(pdevice->instance->debug_flags & RADV_DEBUG_NO_COMPUTE_QUEUE)) {
2521 if (*pCount > idx) {
2522 *pQueueFamilyProperties[idx] = (VkQueueFamilyProperties){
2524 VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT | VK_QUEUE_SPARSE_BINDING_BIT,
2525 .queueCount = pdevice->rad_info.ip[AMD_IP_COMPUTE].num_queues,
2526 .timestampValidBits = 64,
2527 .minImageTransferGranularity = (VkExtent3D){1, 1, 1},
2535 static const VkQueueGlobalPriorityEXT radv_global_queue_priorities[] = {
2536 VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT,
2537 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT,
2538 VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT,
2539 VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT,
2542 VKAPI_ATTR void VKAPI_CALL
2543 radv_GetPhysicalDeviceQueueFamilyProperties2(VkPhysicalDevice physicalDevice, uint32_t *pCount,
2544 VkQueueFamilyProperties2 *pQueueFamilyProperties)
2546 RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
2547 if (!pQueueFamilyProperties) {
2548 radv_get_physical_device_queue_family_properties(pdevice, pCount, NULL);
2551 VkQueueFamilyProperties *properties[] = {
2552 &pQueueFamilyProperties[0].queueFamilyProperties,
2553 &pQueueFamilyProperties[1].queueFamilyProperties,
2554 &pQueueFamilyProperties[2].queueFamilyProperties,
2556 radv_get_physical_device_queue_family_properties(pdevice, pCount, properties);
2557 assert(*pCount <= 3);
2559 for (uint32_t i = 0; i < *pCount; i++) {
2560 vk_foreach_struct(ext, pQueueFamilyProperties[i].pNext)
2562 switch (ext->sType) {
2563 case VK_STRUCTURE_TYPE_QUEUE_FAMILY_GLOBAL_PRIORITY_PROPERTIES_EXT: {
2564 VkQueueFamilyGlobalPriorityPropertiesEXT *prop =
2565 (VkQueueFamilyGlobalPriorityPropertiesEXT *)ext;
2566 STATIC_ASSERT(ARRAY_SIZE(radv_global_queue_priorities) <= VK_MAX_GLOBAL_PRIORITY_SIZE_EXT);
2567 prop->priorityCount = ARRAY_SIZE(radv_global_queue_priorities);
2568 memcpy(&prop->priorities, radv_global_queue_priorities, sizeof(radv_global_queue_priorities));
2579 radv_get_memory_budget_properties(VkPhysicalDevice physicalDevice,
2580 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
2582 RADV_FROM_HANDLE(radv_physical_device, device, physicalDevice);
2583 VkPhysicalDeviceMemoryProperties *memory_properties = &device->memory_properties;
2585 /* For all memory heaps, the computation of budget is as follow:
2586 * heap_budget = heap_size - global_heap_usage + app_heap_usage
2588 * The Vulkan spec 1.1.97 says that the budget should include any
2589 * currently allocated device memory.
2591 * Note that the application heap usages are not really accurate (eg.
2592 * in presence of shared buffers).
2594 if (!device->rad_info.has_dedicated_vram) {
2595 /* On APUs, the driver exposes fake heaps to the application because usually the carveout is
2596 * too small for games but the budgets need to be redistributed accordingly.
2599 assert(device->heaps == (RADV_HEAP_GTT | RADV_HEAP_VRAM_VIS));
2600 assert(device->memory_properties.memoryHeaps[0].flags == 0); /* GTT */
2601 assert(device->memory_properties.memoryHeaps[1].flags == VK_MEMORY_HEAP_DEVICE_LOCAL_BIT);
2602 uint8_t gtt_heap_idx = 0, vram_vis_heap_idx = 1;
2604 /* Get the visible VRAM/GTT heap sizes and internal usages. */
2605 uint64_t gtt_heap_size = device->memory_properties.memoryHeaps[gtt_heap_idx].size;
2606 uint64_t vram_vis_heap_size = device->memory_properties.memoryHeaps[vram_vis_heap_idx].size;
2608 uint64_t vram_vis_internal_usage = device->ws->query_value(device->ws, RADEON_ALLOCATED_VRAM_VIS) +
2609 device->ws->query_value(device->ws, RADEON_ALLOCATED_VRAM);
2610 uint64_t gtt_internal_usage = device->ws->query_value(device->ws, RADEON_ALLOCATED_GTT);
2612 /* Compute the total heap size, internal and system usage. */
2613 uint64_t total_heap_size = vram_vis_heap_size + gtt_heap_size;
2614 uint64_t total_internal_usage = vram_vis_internal_usage + gtt_internal_usage;
2615 uint64_t total_system_usage = device->ws->query_value(device->ws, RADEON_VRAM_VIS_USAGE) +
2616 device->ws->query_value(device->ws, RADEON_GTT_USAGE);
2618 uint64_t total_usage = MAX2(total_internal_usage, total_system_usage);
2620 /* Compute the total free space that can be allocated for this process accross all heaps. */
2621 uint64_t total_free_space = total_heap_size - MIN2(total_heap_size, total_usage);
2623 /* Compute the remaining visible VRAM size for this process. */
2624 uint64_t vram_vis_free_space = vram_vis_heap_size - MIN2(vram_vis_heap_size, vram_vis_internal_usage);
2626 /* Distribute the total free space (2/3rd as VRAM and 1/3rd as GTT) to match the heap sizes,
2627 * and align down to the page size to be conservative.
2629 vram_vis_free_space = ROUND_DOWN_TO(MIN2((total_free_space * 2) / 3, vram_vis_free_space),
2630 device->rad_info.gart_page_size);
2631 uint64_t gtt_free_space = total_free_space - vram_vis_free_space;
2633 memoryBudget->heapBudget[vram_vis_heap_idx] = vram_vis_free_space + vram_vis_internal_usage;
2634 memoryBudget->heapUsage[vram_vis_heap_idx] = vram_vis_internal_usage;
2635 memoryBudget->heapBudget[gtt_heap_idx] = gtt_free_space + gtt_internal_usage;
2636 memoryBudget->heapUsage[gtt_heap_idx] = gtt_internal_usage;
2638 unsigned mask = device->heaps;
2641 uint64_t internal_usage = 0, system_usage = 0;
2642 unsigned type = 1u << u_bit_scan(&mask);
2645 case RADV_HEAP_VRAM:
2646 internal_usage = device->ws->query_value(device->ws, RADEON_ALLOCATED_VRAM);
2647 system_usage = device->ws->query_value(device->ws, RADEON_VRAM_USAGE);
2649 case RADV_HEAP_VRAM_VIS:
2650 internal_usage = device->ws->query_value(device->ws, RADEON_ALLOCATED_VRAM_VIS);
2651 if (!(device->heaps & RADV_HEAP_VRAM))
2652 internal_usage += device->ws->query_value(device->ws, RADEON_ALLOCATED_VRAM);
2653 system_usage = device->ws->query_value(device->ws, RADEON_VRAM_VIS_USAGE);
2656 internal_usage = device->ws->query_value(device->ws, RADEON_ALLOCATED_GTT);
2657 system_usage = device->ws->query_value(device->ws, RADEON_GTT_USAGE);
2661 uint64_t total_usage = MAX2(internal_usage, system_usage);
2663 uint64_t free_space = device->memory_properties.memoryHeaps[heap].size -
2664 MIN2(device->memory_properties.memoryHeaps[heap].size, total_usage);
2665 memoryBudget->heapBudget[heap] = free_space + internal_usage;
2666 memoryBudget->heapUsage[heap] = internal_usage;
2670 assert(heap == memory_properties->memoryHeapCount);
2673 /* The heapBudget and heapUsage values must be zero for array elements
2674 * greater than or equal to
2675 * VkPhysicalDeviceMemoryProperties::memoryHeapCount.
2677 for (uint32_t i = memory_properties->memoryHeapCount; i < VK_MAX_MEMORY_HEAPS; i++) {
2678 memoryBudget->heapBudget[i] = 0;
2679 memoryBudget->heapUsage[i] = 0;
2683 VKAPI_ATTR void VKAPI_CALL
2684 radv_GetPhysicalDeviceMemoryProperties2(VkPhysicalDevice physicalDevice,
2685 VkPhysicalDeviceMemoryProperties2 *pMemoryProperties)
2687 RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
2689 pMemoryProperties->memoryProperties = pdevice->memory_properties;
2691 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memory_budget =
2692 vk_find_struct(pMemoryProperties->pNext, PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT);
2694 radv_get_memory_budget_properties(physicalDevice, memory_budget);
2697 VKAPI_ATTR VkResult VKAPI_CALL
2698 radv_GetMemoryHostPointerPropertiesEXT(
2699 VkDevice _device, VkExternalMemoryHandleTypeFlagBits handleType, const void *pHostPointer,
2700 VkMemoryHostPointerPropertiesEXT *pMemoryHostPointerProperties)
2702 RADV_FROM_HANDLE(radv_device, device, _device);
2704 switch (handleType) {
2705 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
2706 const struct radv_physical_device *physical_device = device->physical_device;
2707 uint32_t memoryTypeBits = 0;
2708 for (int i = 0; i < physical_device->memory_properties.memoryTypeCount; i++) {
2709 if (physical_device->memory_domains[i] == RADEON_DOMAIN_GTT &&
2710 !(physical_device->memory_flags[i] & RADEON_FLAG_GTT_WC)) {
2711 memoryTypeBits = (1 << i);
2715 pMemoryHostPointerProperties->memoryTypeBits = memoryTypeBits;
2719 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
2723 static enum radeon_ctx_priority
2724 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT *pObj)
2726 /* Default to MEDIUM when a specific global priority isn't requested */
2728 return RADEON_CTX_PRIORITY_MEDIUM;
2730 switch (pObj->globalPriority) {
2731 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2732 return RADEON_CTX_PRIORITY_REALTIME;
2733 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2734 return RADEON_CTX_PRIORITY_HIGH;
2735 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2736 return RADEON_CTX_PRIORITY_MEDIUM;
2737 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2738 return RADEON_CTX_PRIORITY_LOW;
2740 unreachable("Illegal global priority value");
2741 return RADEON_CTX_PRIORITY_INVALID;
2746 radv_queue_init(struct radv_device *device, struct radv_queue *queue, int idx,
2747 const VkDeviceQueueCreateInfo *create_info,
2748 const VkDeviceQueueGlobalPriorityCreateInfoEXT *global_priority)
2750 queue->device = device;
2751 queue->priority = radv_get_queue_global_priority(global_priority);
2752 queue->hw_ctx = device->hw_ctx[queue->priority];
2753 queue->state.qf = vk_queue_to_radv(device->physical_device, create_info->queueFamilyIndex);
2755 VkResult result = vk_queue_init(&queue->vk, &device->vk, create_info, idx);
2756 if (result != VK_SUCCESS)
2759 queue->vk.driver_submit = radv_queue_submit;
2765 radv_queue_state_finish(struct radv_queue_state *queue, struct radeon_winsys *ws)
2767 if (queue->initial_full_flush_preamble_cs)
2768 ws->cs_destroy(queue->initial_full_flush_preamble_cs);
2769 if (queue->initial_preamble_cs)
2770 ws->cs_destroy(queue->initial_preamble_cs);
2771 if (queue->continue_preamble_cs)
2772 ws->cs_destroy(queue->continue_preamble_cs);
2773 if (queue->descriptor_bo)
2774 ws->buffer_destroy(ws, queue->descriptor_bo);
2775 if (queue->scratch_bo)
2776 ws->buffer_destroy(ws, queue->scratch_bo);
2777 if (queue->esgs_ring_bo)
2778 ws->buffer_destroy(ws, queue->esgs_ring_bo);
2779 if (queue->gsvs_ring_bo)
2780 ws->buffer_destroy(ws, queue->gsvs_ring_bo);
2781 if (queue->tess_rings_bo)
2782 ws->buffer_destroy(ws, queue->tess_rings_bo);
2783 if (queue->task_rings_bo)
2784 ws->buffer_destroy(ws, queue->task_rings_bo);
2786 ws->buffer_destroy(ws, queue->gds_bo);
2787 if (queue->gds_oa_bo)
2788 ws->buffer_destroy(ws, queue->gds_oa_bo);
2789 if (queue->compute_scratch_bo)
2790 ws->buffer_destroy(ws, queue->compute_scratch_bo);
2794 radv_queue_finish(struct radv_queue *queue)
2796 radv_queue_state_finish(&queue->state, queue->device->ws);
2797 vk_queue_finish(&queue->vk);
2801 radv_device_init_border_color(struct radv_device *device)
2805 result = device->ws->buffer_create(
2806 device->ws, RADV_BORDER_COLOR_BUFFER_SIZE, 4096, RADEON_DOMAIN_VRAM,
2807 RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_READ_ONLY | RADEON_FLAG_NO_INTERPROCESS_SHARING,
2808 RADV_BO_PRIORITY_SHADER, 0, &device->border_color_data.bo);
2810 if (result != VK_SUCCESS)
2811 return vk_error(device, result);
2813 result = device->ws->buffer_make_resident(device->ws, device->border_color_data.bo, true);
2814 if (result != VK_SUCCESS)
2815 return vk_error(device, result);
2817 device->border_color_data.colors_gpu_ptr = device->ws->buffer_map(device->border_color_data.bo);
2818 if (!device->border_color_data.colors_gpu_ptr)
2819 return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
2820 mtx_init(&device->border_color_data.mutex, mtx_plain);
2826 radv_device_finish_border_color(struct radv_device *device)
2828 if (device->border_color_data.bo) {
2829 device->ws->buffer_make_resident(device->ws, device->border_color_data.bo, false);
2830 device->ws->buffer_destroy(device->ws, device->border_color_data.bo);
2832 mtx_destroy(&device->border_color_data.mutex);
2837 radv_device_init_vs_prologs(struct radv_device *device)
2839 u_rwlock_init(&device->vs_prologs_lock);
2840 device->vs_prologs = _mesa_hash_table_create(NULL, &radv_hash_vs_prolog, &radv_cmp_vs_prolog);
2841 if (!device->vs_prologs)
2842 return vk_error(device->physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
2844 /* don't pre-compile prologs if we want to print them */
2845 if (device->instance->debug_flags & RADV_DEBUG_DUMP_PROLOGS)
2848 struct radv_vs_input_state state;
2849 state.nontrivial_divisors = 0;
2850 memset(state.offsets, 0, sizeof(state.offsets));
2851 state.alpha_adjust_lo = 0;
2852 state.alpha_adjust_hi = 0;
2853 memset(state.formats, 0, sizeof(state.formats));
2855 struct radv_vs_prolog_key key;
2857 key.misaligned_mask = 0;
2859 key.is_ngg = device->physical_device->use_ngg;
2860 key.next_stage = MESA_SHADER_VERTEX;
2861 key.wave32 = device->physical_device->ge_wave_size == 32;
2863 for (unsigned i = 1; i <= MAX_VERTEX_ATTRIBS; i++) {
2864 state.attribute_mask = BITFIELD_MASK(i);
2865 state.instance_rate_inputs = 0;
2867 key.num_attributes = i;
2869 device->simple_vs_prologs[i - 1] = radv_create_vs_prolog(device, &key);
2870 if (!device->simple_vs_prologs[i - 1])
2871 return vk_error(device->physical_device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
2875 for (unsigned num_attributes = 1; num_attributes <= 16; num_attributes++) {
2876 state.attribute_mask = BITFIELD_MASK(num_attributes);
2878 for (unsigned i = 0; i < num_attributes; i++)
2879 state.divisors[i] = 1;
2881 for (unsigned count = 1; count <= num_attributes; count++) {
2882 for (unsigned start = 0; start <= (num_attributes - count); start++) {
2883 state.instance_rate_inputs = u_bit_consecutive(start, count);
2885 key.num_attributes = num_attributes;
2887 struct radv_shader_part *prolog = radv_create_vs_prolog(device, &key);
2889 return vk_error(device->physical_device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
2892 radv_instance_rate_prolog_index(num_attributes, state.instance_rate_inputs));
2893 device->instance_rate_vs_prologs[idx++] = prolog;
2897 assert(idx == ARRAY_SIZE(device->instance_rate_vs_prologs));
2903 radv_device_finish_vs_prologs(struct radv_device *device)
2905 if (device->vs_prologs) {
2906 hash_table_foreach(device->vs_prologs, entry)
2908 free((void *)entry->key);
2909 radv_shader_part_destroy(device, entry->data);
2911 _mesa_hash_table_destroy(device->vs_prologs, NULL);
2914 for (unsigned i = 0; i < ARRAY_SIZE(device->simple_vs_prologs); i++)
2915 radv_shader_part_destroy(device, device->simple_vs_prologs[i]);
2917 for (unsigned i = 0; i < ARRAY_SIZE(device->instance_rate_vs_prologs); i++)
2918 radv_shader_part_destroy(device, device->instance_rate_vs_prologs[i]);
2922 radv_device_init_vrs_state(struct radv_device *device)
2924 /* FIXME: 4k depth buffers should be large enough for now but we might want to adjust this
2925 * dynamically at some point.
2927 uint32_t width = 4096, height = 4096;
2933 VkImageCreateInfo image_create_info = {
2934 .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
2935 .imageType = VK_IMAGE_TYPE_2D,
2936 .format = VK_FORMAT_D16_UNORM,
2937 .extent = {width, height, 1},
2940 .samples = VK_SAMPLE_COUNT_1_BIT,
2941 .tiling = VK_IMAGE_TILING_OPTIMAL,
2942 .usage = VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT,
2943 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
2944 .queueFamilyIndexCount = 0,
2945 .pQueueFamilyIndices = NULL,
2946 .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
2949 result = radv_CreateImage(radv_device_to_handle(device), &image_create_info,
2950 &device->meta_state.alloc, &image);
2951 if (result != VK_SUCCESS)
2954 VkBufferCreateInfo buffer_create_info = {
2955 .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
2956 .size = radv_image_from_handle(image)->planes[0].surface.meta_size,
2957 .usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
2958 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
2961 result = radv_CreateBuffer(radv_device_to_handle(device), &buffer_create_info,
2962 &device->meta_state.alloc, &buffer);
2963 if (result != VK_SUCCESS)
2966 VkBufferMemoryRequirementsInfo2 info = {
2967 .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2,
2970 VkMemoryRequirements2 mem_req = {
2971 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
2973 radv_GetBufferMemoryRequirements2(radv_device_to_handle(device), &info, &mem_req);
2975 VkMemoryAllocateInfo alloc_info = {
2976 .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
2977 .allocationSize = mem_req.memoryRequirements.size,
2980 result = radv_AllocateMemory(radv_device_to_handle(device), &alloc_info,
2981 &device->meta_state.alloc, &mem);
2982 if (result != VK_SUCCESS)
2985 VkBindBufferMemoryInfo bind_info = {
2986 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
2992 result = radv_BindBufferMemory2(radv_device_to_handle(device), 1, &bind_info);
2993 if (result != VK_SUCCESS)
2996 device->vrs.image = radv_image_from_handle(image);
2997 device->vrs.buffer = radv_buffer_from_handle(buffer);
2998 device->vrs.mem = radv_device_memory_from_handle(mem);
3003 radv_FreeMemory(radv_device_to_handle(device), mem, &device->meta_state.alloc);
3005 radv_DestroyBuffer(radv_device_to_handle(device), buffer, &device->meta_state.alloc);
3007 radv_DestroyImage(radv_device_to_handle(device), image, &device->meta_state.alloc);
3013 radv_device_finish_vrs_image(struct radv_device *device)
3015 if (!device->vrs.image)
3018 radv_FreeMemory(radv_device_to_handle(device), radv_device_memory_to_handle(device->vrs.mem),
3019 &device->meta_state.alloc);
3020 radv_DestroyBuffer(radv_device_to_handle(device), radv_buffer_to_handle(device->vrs.buffer),
3021 &device->meta_state.alloc);
3022 radv_DestroyImage(radv_device_to_handle(device), radv_image_to_handle(device->vrs.image),
3023 &device->meta_state.alloc);
3026 static enum radv_force_vrs
3027 radv_parse_vrs_rates(const char *str)
3029 if (!strcmp(str, "2x2")) {
3030 return RADV_FORCE_VRS_2x2;
3031 } else if (!strcmp(str, "2x1")) {
3032 return RADV_FORCE_VRS_2x1;
3033 } else if (!strcmp(str, "1x2")) {
3034 return RADV_FORCE_VRS_1x2;
3035 } else if (!strcmp(str, "1x1")) {
3036 return RADV_FORCE_VRS_1x1;
3039 fprintf(stderr, "radv: Invalid VRS rates specified (valid values are 2x2, 2x1, 1x2 and 1x1)\n");
3040 return RADV_FORCE_VRS_1x1;
3044 radv_get_force_vrs_config_file(void)
3046 return getenv("RADV_FORCE_VRS_CONFIG_FILE");
3049 static enum radv_force_vrs
3050 radv_parse_force_vrs_config_file(const char *config_file)
3052 enum radv_force_vrs force_vrs = RADV_FORCE_VRS_1x1;
3056 f = fopen(config_file, "r");
3058 fprintf(stderr, "radv: Can't open file: '%s'.\n", config_file);
3062 if (fread(buf, sizeof(buf), 1, f) == 1) {
3064 force_vrs = radv_parse_vrs_rates(buf);
3073 #define BUF_LEN ((10 * (sizeof(struct inotify_event) + NAME_MAX + 1)))
3076 radv_notifier_thread_run(void *data)
3078 struct radv_device *device = data;
3079 struct radv_notifier *notifier = &device->notifier;
3082 while (!notifier->quit) {
3083 const char *file = radv_get_force_vrs_config_file();
3084 struct timespec tm = { .tv_nsec = 100000000 }; /* 1OOms */
3087 length = read(notifier->fd, buf, BUF_LEN);
3088 while (i < length) {
3089 struct inotify_event *event = (struct inotify_event *)&buf[i];
3091 i += sizeof(struct inotify_event) + event->len;
3092 if (event->mask & IN_MODIFY || event->mask & IN_DELETE_SELF) {
3093 /* Sleep 100ms for editors that use a temporary file and delete the original. */
3094 thrd_sleep(&tm, NULL);
3095 device->force_vrs = radv_parse_force_vrs_config_file(file);
3097 fprintf(stderr, "radv: Updated the per-vertex VRS rate to '%d'.\n", device->force_vrs);
3099 if (event->mask & IN_DELETE_SELF) {
3100 inotify_rm_watch(notifier->fd, notifier->watch);
3101 notifier->watch = inotify_add_watch(notifier->fd, file, IN_MODIFY | IN_DELETE_SELF);
3106 thrd_sleep(&tm, NULL);
3115 radv_device_init_notifier(struct radv_device *device)
3120 struct radv_notifier *notifier = &device->notifier;
3121 const char *file = radv_get_force_vrs_config_file();
3124 notifier->fd = inotify_init1(IN_NONBLOCK);
3125 if (notifier->fd < 0)
3128 notifier->watch = inotify_add_watch(notifier->fd, file, IN_MODIFY | IN_DELETE_SELF);
3129 if (notifier->watch < 0)
3132 ret = thrd_create(¬ifier->thread, radv_notifier_thread_run, device);
3139 inotify_rm_watch(notifier->fd, notifier->watch);
3141 close(notifier->fd);
3148 radv_device_finish_notifier(struct radv_device *device)
3151 struct radv_notifier *notifier = &device->notifier;
3153 if (!notifier->thread)
3156 notifier->quit = true;
3157 thrd_join(notifier->thread, NULL);
3158 inotify_rm_watch(notifier->fd, notifier->watch);
3159 close(notifier->fd);
3164 radv_device_finish_perf_counter_lock_cs(struct radv_device *device)
3166 if (!device->perf_counter_lock_cs)
3169 for (unsigned i = 0; i < 2 * PERF_CTR_MAX_PASSES; ++i) {
3170 if (device->perf_counter_lock_cs[i])
3171 device->ws->cs_destroy(device->perf_counter_lock_cs[i]);
3174 free(device->perf_counter_lock_cs);
3177 VKAPI_ATTR VkResult VKAPI_CALL
3178 radv_CreateDevice(VkPhysicalDevice physicalDevice, const VkDeviceCreateInfo *pCreateInfo,
3179 const VkAllocationCallbacks *pAllocator, VkDevice *pDevice)
3181 RADV_FROM_HANDLE(radv_physical_device, physical_device, physicalDevice);
3183 struct radv_device *device;
3185 bool keep_shader_info = false;
3186 bool robust_buffer_access = false;
3187 bool robust_buffer_access2 = false;
3188 bool overallocation_disallowed = false;
3189 bool custom_border_colors = false;
3190 bool attachment_vrs_enabled = false;
3191 bool image_float32_atomics = false;
3192 bool vs_prologs = false;
3193 bool global_bo_list = false;
3194 bool image_2d_view_of_3d = false;
3195 bool primitives_generated_query = false;
3196 bool use_perf_counters = false;
3198 /* Check enabled features */
3199 if (pCreateInfo->pEnabledFeatures) {
3200 if (pCreateInfo->pEnabledFeatures->robustBufferAccess)
3201 robust_buffer_access = true;
3204 vk_foreach_struct_const(ext, pCreateInfo->pNext)
3206 switch (ext->sType) {
3207 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2: {
3208 const VkPhysicalDeviceFeatures2 *features = (const void *)ext;
3209 if (features->features.robustBufferAccess)
3210 robust_buffer_access = true;
3213 case VK_STRUCTURE_TYPE_DEVICE_MEMORY_OVERALLOCATION_CREATE_INFO_AMD: {
3214 const VkDeviceMemoryOverallocationCreateInfoAMD *overallocation = (const void *)ext;
3215 if (overallocation->overallocationBehavior ==
3216 VK_MEMORY_OVERALLOCATION_BEHAVIOR_DISALLOWED_AMD)
3217 overallocation_disallowed = true;
3220 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT: {
3221 const VkPhysicalDeviceCustomBorderColorFeaturesEXT *border_color_features =
3223 custom_border_colors = border_color_features->customBorderColors;
3226 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADING_RATE_FEATURES_KHR: {
3227 const VkPhysicalDeviceFragmentShadingRateFeaturesKHR *vrs = (const void *)ext;
3228 attachment_vrs_enabled = vrs->attachmentFragmentShadingRate;
3231 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT: {
3232 const VkPhysicalDeviceRobustness2FeaturesEXT *features = (const void *)ext;
3233 if (features->robustBufferAccess2)
3234 robust_buffer_access2 = true;
3237 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT: {
3238 const VkPhysicalDeviceShaderAtomicFloatFeaturesEXT *features = (const void *)ext;
3239 if (features->shaderImageFloat32Atomics ||
3240 features->sparseImageFloat32Atomics)
3241 image_float32_atomics = true;
3244 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_2_FEATURES_EXT: {
3245 const VkPhysicalDeviceShaderAtomicFloat2FeaturesEXT *features = (const void *)ext;
3246 if (features->shaderImageFloat32AtomicMinMax ||
3247 features->sparseImageFloat32AtomicMinMax)
3248 image_float32_atomics = true;
3251 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT: {
3252 const VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT *features = (const void *)ext;
3253 if (features->vertexInputDynamicState)
3257 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES: {
3258 const VkPhysicalDeviceVulkan12Features *features = (const void *)ext;
3259 if (features->bufferDeviceAddress || features->descriptorIndexing)
3260 global_bo_list = true;
3263 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_2D_VIEW_OF_3D_FEATURES_EXT: {
3264 const VkPhysicalDeviceImage2DViewOf3DFeaturesEXT *features = (const void *)ext;
3265 if (features->image2DViewOf3D)
3266 image_2d_view_of_3d = true;
3269 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRIMITIVES_GENERATED_QUERY_FEATURES_EXT: {
3270 const VkPhysicalDevicePrimitivesGeneratedQueryFeaturesEXT *features = (const void *)ext;
3271 if (features->primitivesGeneratedQuery ||
3272 features->primitivesGeneratedQueryWithRasterizerDiscard ||
3273 features->primitivesGeneratedQueryWithNonZeroStreams)
3274 primitives_generated_query = true;
3277 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PERFORMANCE_QUERY_FEATURES_KHR: {
3278 const VkPhysicalDevicePerformanceQueryFeaturesKHR *features = (const void *)ext;
3279 if (features->performanceCounterQueryPools)
3280 use_perf_counters = true;
3288 device = vk_zalloc2(&physical_device->instance->vk.alloc, pAllocator, sizeof(*device), 8,
3289 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
3291 return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
3293 struct vk_device_dispatch_table dispatch_table;
3295 if (physical_device->instance->vk.app_info.app_name &&
3296 !strcmp(physical_device->instance->vk.app_info.app_name, "metroexodus")) {
3297 /* Metro Exodus (Linux native) calls vkGetSemaphoreCounterValue() with a NULL semaphore and it
3298 * crashes sometimes. Workaround this game bug by enabling an internal layer. Remove this
3299 * when the game is fixed.
3301 vk_device_dispatch_table_from_entrypoints(&dispatch_table, &metro_exodus_device_entrypoints, true);
3302 vk_device_dispatch_table_from_entrypoints(&dispatch_table, &radv_device_entrypoints, false);
3303 } else if (radv_thread_trace_enabled()) {
3304 vk_device_dispatch_table_from_entrypoints(&dispatch_table, &sqtt_device_entrypoints, true);
3305 vk_device_dispatch_table_from_entrypoints(&dispatch_table, &radv_device_entrypoints, false);
3307 vk_device_dispatch_table_from_entrypoints(&dispatch_table, &radv_device_entrypoints, true);
3309 vk_device_dispatch_table_from_entrypoints(&dispatch_table, &wsi_device_entrypoints, false);
3312 vk_device_init(&device->vk, &physical_device->vk, &dispatch_table, pCreateInfo, pAllocator);
3313 if (result != VK_SUCCESS) {
3314 vk_free(&device->vk.alloc, device);
3318 device->instance = physical_device->instance;
3319 device->physical_device = physical_device;
3320 simple_mtx_init(&device->trace_mtx, mtx_plain);
3321 simple_mtx_init(&device->pstate_mtx, mtx_plain);
3323 device->ws = physical_device->ws;
3324 vk_device_set_drm_fd(&device->vk, device->ws->get_fd(device->ws));
3326 /* With update after bind we can't attach bo's to the command buffer
3327 * from the descriptor set anymore, so we have to use a global BO list.
3329 device->use_global_bo_list = global_bo_list ||
3330 (device->instance->perftest_flags & RADV_PERFTEST_BO_LIST) ||
3331 device->vk.enabled_extensions.EXT_descriptor_indexing ||
3332 device->vk.enabled_extensions.EXT_buffer_device_address ||
3333 device->vk.enabled_extensions.KHR_buffer_device_address ||
3334 device->vk.enabled_extensions.KHR_ray_tracing_pipeline ||
3335 device->vk.enabled_extensions.KHR_acceleration_structure ||
3336 device->vk.enabled_extensions.VALVE_descriptor_set_host_mapping;
3338 device->robust_buffer_access = robust_buffer_access || robust_buffer_access2;
3339 device->robust_buffer_access2 = robust_buffer_access2;
3341 device->attachment_vrs_enabled = attachment_vrs_enabled;
3343 device->image_float32_atomics = image_float32_atomics;
3345 device->image_2d_view_of_3d = image_2d_view_of_3d;
3347 device->primitives_generated_query = primitives_generated_query;
3349 radv_init_shader_arenas(device);
3351 device->overallocation_disallowed = overallocation_disallowed;
3352 mtx_init(&device->overallocation_mutex, mtx_plain);
3354 /* Create one context per queue priority. */
3355 for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
3356 const VkDeviceQueueCreateInfo *queue_create = &pCreateInfo->pQueueCreateInfos[i];
3357 const VkDeviceQueueGlobalPriorityCreateInfoEXT *global_priority =
3358 vk_find_struct_const(queue_create->pNext, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
3359 enum radeon_ctx_priority priority = radv_get_queue_global_priority(global_priority);
3361 if (device->hw_ctx[priority])
3364 result = device->ws->ctx_create(device->ws, priority, &device->hw_ctx[priority]);
3365 if (result != VK_SUCCESS)
3369 for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
3370 const VkDeviceQueueCreateInfo *queue_create = &pCreateInfo->pQueueCreateInfos[i];
3371 uint32_t qfi = queue_create->queueFamilyIndex;
3372 const VkDeviceQueueGlobalPriorityCreateInfoEXT *global_priority =
3373 vk_find_struct_const(queue_create->pNext, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
3375 device->queues[qfi] =
3376 vk_alloc(&device->vk.alloc, queue_create->queueCount * sizeof(struct radv_queue), 8,
3377 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
3378 if (!device->queues[qfi]) {
3379 result = VK_ERROR_OUT_OF_HOST_MEMORY;
3383 memset(device->queues[qfi], 0, queue_create->queueCount * sizeof(struct radv_queue));
3385 device->queue_count[qfi] = queue_create->queueCount;
3387 for (unsigned q = 0; q < queue_create->queueCount; q++) {
3388 result = radv_queue_init(device, &device->queues[qfi][q], q, queue_create, global_priority);
3389 if (result != VK_SUCCESS)
3393 device->private_sdma_queue = VK_NULL_HANDLE;
3395 device->pbb_allowed = device->physical_device->rad_info.gfx_level >= GFX9 &&
3396 !(device->instance->debug_flags & RADV_DEBUG_NOBINNING);
3398 /* The maximum number of scratch waves. Scratch space isn't divided
3399 * evenly between CUs. The number is only a function of the number of CUs.
3400 * We can decrease the constant to decrease the scratch buffer size.
3402 * sctx->scratch_waves must be >= the maximum possible size of
3403 * 1 threadgroup, so that the hw doesn't hang from being unable
3406 * The recommended value is 4 per CU at most. Higher numbers don't
3407 * bring much benefit, but they still occupy chip resources (think
3408 * async compute). I've seen ~2% performance difference between 4 and 32.
3410 uint32_t max_threads_per_block = 2048;
3411 device->scratch_waves =
3412 MAX2(32 * physical_device->rad_info.num_good_compute_units, max_threads_per_block / 64);
3414 device->dispatch_initiator = S_00B800_COMPUTE_SHADER_EN(1);
3416 if (device->physical_device->rad_info.gfx_level >= GFX7) {
3417 /* If the KMD allows it (there is a KMD hw register for it),
3418 * allow launching waves out-of-order.
3420 device->dispatch_initiator |= S_00B800_ORDER_MODE(1);
3423 if (device->instance->debug_flags & RADV_DEBUG_HANG) {
3424 /* Enable GPU hangs detection and dump logs if a GPU hang is
3427 keep_shader_info = true;
3429 if (!radv_init_trace(device))
3433 "*****************************************************************************\n");
3435 "* WARNING: RADV_DEBUG=hang is costly and should only be used for debugging! *\n");
3437 "*****************************************************************************\n");
3439 /* Wait for idle after every draw/dispatch to identify the
3442 device->instance->debug_flags |= RADV_DEBUG_SYNC_SHADERS;
3444 radv_dump_enabled_options(device, stderr);
3447 if (radv_thread_trace_enabled()) {
3448 if (device->physical_device->rad_info.gfx_level < GFX8 ||
3449 device->physical_device->rad_info.gfx_level > GFX10_3) {
3450 fprintf(stderr, "GPU hardware not supported: refer to "
3451 "the RGP documentation for the list of "
3452 "supported GPUs!\n");
3456 if (!radv_thread_trace_init(device))
3459 fprintf(stderr, "radv: Thread trace support is enabled (initial buffer size: %u MiB, "
3460 "instruction timing: %s, cache counters: %s).\n",
3461 device->thread_trace.buffer_size / (1024 * 1024),
3462 radv_is_instruction_timing_enabled() ? "enabled" : "disabled",
3463 radv_spm_trace_enabled() ? "enabled" : "disabled");
3465 if (radv_spm_trace_enabled()) {
3466 if (device->physical_device->rad_info.gfx_level >= GFX10) {
3467 if (!radv_spm_init(device))
3470 fprintf(stderr, "radv: SPM isn't supported for this GPU (%s)!\n",
3471 device->physical_device->name);
3476 if (getenv("RADV_TRAP_HANDLER")) {
3477 /* TODO: Add support for more hardware. */
3478 assert(device->physical_device->rad_info.gfx_level == GFX8);
3480 fprintf(stderr, "**********************************************************************\n");
3481 fprintf(stderr, "* WARNING: RADV_TRAP_HANDLER is experimental and only for debugging! *\n");
3482 fprintf(stderr, "**********************************************************************\n");
3484 /* To get the disassembly of the faulty shaders, we have to
3485 * keep some shader info around.
3487 keep_shader_info = true;
3489 if (!radv_trap_handler_init(device))
3493 if (device->physical_device->rad_info.gfx_level >= GFX10_3) {
3494 if (getenv("RADV_FORCE_VRS_CONFIG_FILE")) {
3495 const char *file = radv_get_force_vrs_config_file();
3497 device->force_vrs = radv_parse_force_vrs_config_file(file);
3499 if (radv_device_init_notifier(device)) {
3500 device->force_vrs_enabled = true;
3502 fprintf(stderr, "radv: Failed to initialize the notifier for RADV_FORCE_VRS_CONFIG_FILE!\n");
3504 } else if (getenv("RADV_FORCE_VRS")) {
3505 const char *vrs_rates = getenv("RADV_FORCE_VRS");
3507 device->force_vrs = radv_parse_vrs_rates(vrs_rates);
3508 device->force_vrs_enabled = device->force_vrs != RADV_FORCE_VRS_1x1;
3512 /* PKT3_LOAD_SH_REG_INDEX is supported on GFX8+, but it hangs with compute queues until GFX10.3. */
3513 device->load_grid_size_from_user_sgpr = device->physical_device->rad_info.gfx_level >= GFX10_3;
3515 device->keep_shader_info = keep_shader_info;
3516 result = radv_device_init_meta(device);
3517 if (result != VK_SUCCESS)
3520 radv_device_init_msaa(device);
3522 /* If the border color extension is enabled, let's create the buffer we need. */
3523 if (custom_border_colors) {
3524 result = radv_device_init_border_color(device);
3525 if (result != VK_SUCCESS)
3530 result = radv_device_init_vs_prologs(device);
3531 if (result != VK_SUCCESS)
3535 if (device->physical_device->rad_info.gfx_level >= GFX7)
3536 cik_create_gfx_config(device);
3538 VkPipelineCacheCreateInfo ci;
3539 ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
3542 ci.pInitialData = NULL;
3543 ci.initialDataSize = 0;
3545 result = radv_CreatePipelineCache(radv_device_to_handle(device), &ci, NULL, &pc);
3546 if (result != VK_SUCCESS)
3549 device->mem_cache = radv_pipeline_cache_from_handle(pc);
3551 device->force_aniso = MIN2(16, radv_get_int_debug_option("RADV_TEX_ANISO", -1));
3552 if (device->force_aniso >= 0) {
3553 fprintf(stderr, "radv: Forcing anisotropy filter to %ix\n",
3554 1 << util_logbase2(device->force_aniso));
3557 if (use_perf_counters) {
3558 size_t bo_size = PERF_CTR_BO_PASS_OFFSET + sizeof(uint64_t) * PERF_CTR_MAX_PASSES;
3560 device->ws->buffer_create(device->ws, bo_size, 4096, RADEON_DOMAIN_GTT,
3561 RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING,
3562 RADV_BO_PRIORITY_UPLOAD_BUFFER, 0, &device->perf_counter_bo);
3563 if (result != VK_SUCCESS)
3566 device->perf_counter_lock_cs =
3567 calloc(sizeof(struct radeon_winsys_cs *), 2 * PERF_CTR_MAX_PASSES);
3568 if (!device->perf_counter_lock_cs) {
3569 result = VK_ERROR_OUT_OF_HOST_MEMORY;
3574 *pDevice = radv_device_to_handle(device);
3578 radv_DestroyPipelineCache(radv_device_to_handle(device), pc, NULL);
3580 radv_device_finish_meta(device);
3582 radv_thread_trace_finish(device);
3584 radv_spm_finish(device);
3586 radv_trap_handler_finish(device);
3587 radv_finish_trace(device);
3589 radv_device_finish_perf_counter_lock_cs(device);
3590 if (device->perf_counter_bo)
3591 device->ws->buffer_destroy(device->ws, device->perf_counter_bo);
3592 if (device->gfx_init)
3593 device->ws->buffer_destroy(device->ws, device->gfx_init);
3595 radv_device_finish_notifier(device);
3596 radv_device_finish_vs_prologs(device);
3597 radv_device_finish_border_color(device);
3599 for (unsigned i = 0; i < RADV_MAX_QUEUE_FAMILIES; i++) {
3600 for (unsigned q = 0; q < device->queue_count[i]; q++)
3601 radv_queue_finish(&device->queues[i][q]);
3602 if (device->queue_count[i])
3603 vk_free(&device->vk.alloc, device->queues[i]);
3606 for (unsigned i = 0; i < RADV_NUM_HW_CTX; i++) {
3607 if (device->hw_ctx[i])
3608 device->ws->ctx_destroy(device->hw_ctx[i]);
3611 simple_mtx_destroy(&device->pstate_mtx);
3612 simple_mtx_destroy(&device->trace_mtx);
3613 mtx_destroy(&device->overallocation_mutex);
3615 vk_device_finish(&device->vk);
3616 vk_free(&device->vk.alloc, device);
3620 VKAPI_ATTR void VKAPI_CALL
3621 radv_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
3623 RADV_FROM_HANDLE(radv_device, device, _device);
3628 radv_device_finish_perf_counter_lock_cs(device);
3629 if (device->perf_counter_bo)
3630 device->ws->buffer_destroy(device->ws, device->perf_counter_bo);
3632 if (device->gfx_init)
3633 device->ws->buffer_destroy(device->ws, device->gfx_init);
3635 radv_device_finish_notifier(device);
3636 radv_device_finish_vs_prologs(device);
3637 radv_device_finish_border_color(device);
3638 radv_device_finish_vrs_image(device);
3640 for (unsigned i = 0; i < RADV_MAX_QUEUE_FAMILIES; i++) {
3641 for (unsigned q = 0; q < device->queue_count[i]; q++)
3642 radv_queue_finish(&device->queues[i][q]);
3643 if (device->queue_count[i])
3644 vk_free(&device->vk.alloc, device->queues[i]);
3646 if (device->private_sdma_queue != VK_NULL_HANDLE) {
3647 radv_queue_finish(device->private_sdma_queue);
3648 vk_free(&device->vk.alloc, device->private_sdma_queue);
3651 for (unsigned i = 0; i < RADV_NUM_HW_CTX; i++) {
3652 if (device->hw_ctx[i])
3653 device->ws->ctx_destroy(device->hw_ctx[i]);
3656 mtx_destroy(&device->overallocation_mutex);
3657 simple_mtx_destroy(&device->pstate_mtx);
3658 simple_mtx_destroy(&device->trace_mtx);
3660 radv_device_finish_meta(device);
3662 VkPipelineCache pc = radv_pipeline_cache_to_handle(device->mem_cache);
3663 radv_DestroyPipelineCache(radv_device_to_handle(device), pc, NULL);
3665 radv_trap_handler_finish(device);
3666 radv_finish_trace(device);
3668 radv_destroy_shader_arenas(device);
3670 radv_thread_trace_finish(device);
3672 radv_spm_finish(device);
3674 vk_device_finish(&device->vk);
3675 vk_free(&device->vk.alloc, device);
3678 VKAPI_ATTR VkResult VKAPI_CALL
3679 radv_EnumerateInstanceLayerProperties(uint32_t *pPropertyCount, VkLayerProperties *pProperties)
3681 if (pProperties == NULL) {
3682 *pPropertyCount = 0;
3686 /* None supported at this time */
3687 return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
3690 VKAPI_ATTR VkResult VKAPI_CALL
3691 radv_EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice, uint32_t *pPropertyCount,
3692 VkLayerProperties *pProperties)
3694 if (pProperties == NULL) {
3695 *pPropertyCount = 0;
3699 /* None supported at this time */
3700 return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
3704 radv_fill_shader_rings(struct radv_device *device, uint32_t *map, bool add_sample_positions,
3705 uint32_t esgs_ring_size, struct radeon_winsys_bo *esgs_ring_bo,
3706 uint32_t gsvs_ring_size, struct radeon_winsys_bo *gsvs_ring_bo,
3707 struct radeon_winsys_bo *tess_rings_bo,
3708 struct radeon_winsys_bo *task_rings_bo,
3709 struct radeon_winsys_bo *mesh_scratch_ring_bo)
3711 uint32_t *desc = &map[4];
3714 uint64_t esgs_va = radv_buffer_get_va(esgs_ring_bo);
3716 /* stride 0, num records - size, add tid, swizzle, elsize4,
3719 desc[1] = S_008F04_BASE_ADDRESS_HI(esgs_va >> 32);
3720 desc[2] = esgs_ring_size;
3721 desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3722 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
3723 S_008F0C_INDEX_STRIDE(3) | S_008F0C_ADD_TID_ENABLE(1);
3725 if (device->physical_device->rad_info.gfx_level >= GFX11)
3726 desc[1] |= S_008F04_SWIZZLE_ENABLE_GFX11(1);
3728 desc[1] |= S_008F04_SWIZZLE_ENABLE_GFX6(1);
3730 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3731 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) |
3732 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
3733 } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
3734 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
3735 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
3737 desc[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
3738 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) | S_008F0C_ELEMENT_SIZE(1);
3741 /* GS entry for ES->GS ring */
3742 /* stride 0, num records - size, elsize0,
3745 desc[5] = S_008F04_BASE_ADDRESS_HI(esgs_va >> 32);
3746 desc[6] = esgs_ring_size;
3747 desc[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3748 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
3750 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3751 desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) |
3752 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
3753 } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
3754 desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
3755 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
3757 desc[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
3758 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
3765 uint64_t gsvs_va = radv_buffer_get_va(gsvs_ring_bo);
3767 /* VS entry for GS->VS ring */
3768 /* stride 0, num records - size, elsize0,
3771 desc[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va >> 32);
3772 desc[2] = gsvs_ring_size;
3773 desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3774 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
3776 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3777 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) |
3778 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
3779 } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
3780 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
3781 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
3783 desc[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
3784 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
3787 /* stride gsvs_itemsize, num records 64
3788 elsize 4, index stride 16 */
3789 /* shader will patch stride and desc[2] */
3791 desc[5] = S_008F04_BASE_ADDRESS_HI(gsvs_va >> 32);
3793 desc[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3794 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
3795 S_008F0C_INDEX_STRIDE(1) | S_008F0C_ADD_TID_ENABLE(true);
3797 if (device->physical_device->rad_info.gfx_level >= GFX11)
3798 desc[5] |= S_008F04_SWIZZLE_ENABLE_GFX11(1);
3800 desc[5] |= S_008F04_SWIZZLE_ENABLE_GFX6(1);
3802 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3803 desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) |
3804 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
3805 } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
3806 desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
3807 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
3809 desc[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
3810 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) | S_008F0C_ELEMENT_SIZE(1);
3816 if (tess_rings_bo) {
3817 uint64_t tess_va = radv_buffer_get_va(tess_rings_bo);
3818 uint64_t tess_offchip_va = tess_va + device->physical_device->hs.tess_offchip_ring_offset;
3821 desc[1] = S_008F04_BASE_ADDRESS_HI(tess_va >> 32);
3822 desc[2] = device->physical_device->hs.tess_factor_ring_size;
3823 desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3824 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
3826 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3827 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) |
3828 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW);
3829 } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
3830 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
3831 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) | S_008F0C_RESOURCE_LEVEL(1);
3833 desc[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
3834 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
3837 desc[4] = tess_offchip_va;
3838 desc[5] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va >> 32);
3839 desc[6] = device->physical_device->hs.tess_offchip_ring_size;
3840 desc[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3841 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
3843 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3844 desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) |
3845 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW);
3846 } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
3847 desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
3848 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) | S_008F0C_RESOURCE_LEVEL(1);
3850 desc[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
3851 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
3857 if (task_rings_bo) {
3858 uint64_t task_va = radv_buffer_get_va(task_rings_bo);
3859 uint64_t task_draw_ring_va = task_va + device->physical_device->task_info.draw_ring_offset;
3860 uint64_t task_payload_ring_va = task_va + device->physical_device->task_info.payload_ring_offset;
3862 desc[0] = task_draw_ring_va;
3863 desc[1] = S_008F04_BASE_ADDRESS_HI(task_draw_ring_va >> 32);
3864 desc[2] = device->physical_device->task_info.num_entries * AC_TASK_DRAW_ENTRY_BYTES;
3865 desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3866 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
3868 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3869 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_UINT) |
3870 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
3872 assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
3873 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_UINT) |
3874 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
3877 desc[4] = task_payload_ring_va;
3878 desc[5] = S_008F04_BASE_ADDRESS_HI(task_payload_ring_va >> 32);
3879 desc[6] = device->physical_device->task_info.num_entries * AC_TASK_PAYLOAD_ENTRY_BYTES;
3880 desc[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3881 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
3883 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3884 desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_UINT) |
3885 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
3887 assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
3888 desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_UINT) |
3889 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
3895 if (mesh_scratch_ring_bo) {
3896 uint64_t va = radv_buffer_get_va(mesh_scratch_ring_bo);
3899 desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32);
3900 desc[2] = RADV_MESH_SCRATCH_NUM_ENTRIES * RADV_MESH_SCRATCH_ENTRY_BYTES;
3901 desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
3902 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
3904 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3905 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_UINT) |
3906 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
3908 assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
3909 desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_UINT) |
3910 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
3916 if (add_sample_positions) {
3917 /* add sample positions after all rings */
3918 memcpy(desc, device->sample_locations_1x, 8);
3920 memcpy(desc, device->sample_locations_2x, 16);
3922 memcpy(desc, device->sample_locations_4x, 32);
3924 memcpy(desc, device->sample_locations_8x, 64);
3929 radv_emit_gs_ring_sizes(struct radv_device *device, struct radeon_cmdbuf *cs,
3930 struct radeon_winsys_bo *esgs_ring_bo, uint32_t esgs_ring_size,
3931 struct radeon_winsys_bo *gsvs_ring_bo, uint32_t gsvs_ring_size)
3933 if (!esgs_ring_bo && !gsvs_ring_bo)
3937 radv_cs_add_buffer(device->ws, cs, esgs_ring_bo);
3940 radv_cs_add_buffer(device->ws, cs, gsvs_ring_bo);
3942 if (device->physical_device->rad_info.gfx_level >= GFX7) {
3943 radeon_set_uconfig_reg_seq(cs, R_030900_VGT_ESGS_RING_SIZE, 2);
3944 radeon_emit(cs, esgs_ring_size >> 8);
3945 radeon_emit(cs, gsvs_ring_size >> 8);
3947 radeon_set_config_reg_seq(cs, R_0088C8_VGT_ESGS_RING_SIZE, 2);
3948 radeon_emit(cs, esgs_ring_size >> 8);
3949 radeon_emit(cs, gsvs_ring_size >> 8);
3954 radv_emit_tess_factor_ring(struct radv_device *device, struct radeon_cmdbuf *cs,
3955 struct radeon_winsys_bo *tess_rings_bo)
3958 uint32_t tf_ring_size;
3962 tf_ring_size = device->physical_device->hs.tess_factor_ring_size / 4;
3963 tf_va = radv_buffer_get_va(tess_rings_bo);
3965 radv_cs_add_buffer(device->ws, cs, tess_rings_bo);
3967 if (device->physical_device->rad_info.gfx_level >= GFX7) {
3968 if (device->physical_device->rad_info.gfx_level >= GFX11) {
3969 /* TF_RING_SIZE is per SE on GFX11. */
3970 tf_ring_size /= device->physical_device->rad_info.max_se;
3973 radeon_set_uconfig_reg(cs, R_030938_VGT_TF_RING_SIZE, S_030938_SIZE(tf_ring_size));
3974 radeon_set_uconfig_reg(cs, R_030940_VGT_TF_MEMORY_BASE, tf_va >> 8);
3976 if (device->physical_device->rad_info.gfx_level >= GFX10) {
3977 radeon_set_uconfig_reg(cs, R_030984_VGT_TF_MEMORY_BASE_HI,
3978 S_030984_BASE_HI(tf_va >> 40));
3979 } else if (device->physical_device->rad_info.gfx_level == GFX9) {
3980 radeon_set_uconfig_reg(cs, R_030944_VGT_TF_MEMORY_BASE_HI, S_030944_BASE_HI(tf_va >> 40));
3983 radeon_set_uconfig_reg(cs, R_03093C_VGT_HS_OFFCHIP_PARAM, device->physical_device->hs.hs_offchip_param);
3985 radeon_set_config_reg(cs, R_008988_VGT_TF_RING_SIZE, S_008988_SIZE(tf_ring_size));
3986 radeon_set_config_reg(cs, R_0089B8_VGT_TF_MEMORY_BASE, tf_va >> 8);
3987 radeon_set_config_reg(cs, R_0089B0_VGT_HS_OFFCHIP_PARAM, device->physical_device->hs.hs_offchip_param);
3992 radv_initialise_task_control_buffer(struct radv_device *device,
3993 struct radeon_winsys_bo *task_rings_bo)
3995 uint32_t *ptr = (uint32_t *)device->ws->buffer_map(task_rings_bo);
3997 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
3999 const uint32_t num_entries = device->physical_device->task_info.num_entries;
4000 const uint64_t task_va = radv_buffer_get_va(task_rings_bo);
4001 const uint64_t task_draw_ring_va = task_va + device->physical_device->task_info.draw_ring_offset;
4002 assert((task_draw_ring_va & 0xFFFFFF00) == (task_draw_ring_va & 0xFFFFFFFF));
4004 /* 64-bit write_ptr */
4005 ptr[0] = num_entries;
4007 /* 64-bit read_ptr */
4008 ptr[2] = num_entries;
4010 /* 64-bit dealloc_ptr */
4011 ptr[4] = num_entries;
4014 ptr[6] = num_entries;
4015 /* 64-bit draw ring address */
4016 ptr[7] = task_draw_ring_va;
4017 ptr[8] = task_draw_ring_va >> 32;
4019 device->ws->buffer_unmap(task_rings_bo);
4024 radv_emit_task_rings(struct radv_device *device, struct radeon_cmdbuf *cs,
4025 struct radeon_winsys_bo *task_rings_bo, bool compute)
4030 const uint64_t task_ctrlbuf_va = radv_buffer_get_va(task_rings_bo);
4031 assert(radv_is_aligned(task_ctrlbuf_va, 256));
4032 radv_cs_add_buffer(device->ws, cs, task_rings_bo);
4034 /* Tell the GPU where the task control buffer is. */
4035 radeon_emit(cs, PKT3(PKT3_DISPATCH_TASK_STATE_INIT, 1, 0) | PKT3_SHADER_TYPE_S(!!compute));
4036 /* bits [31:8]: control buffer address lo, bits[7:0]: reserved (set to zero) */
4037 radeon_emit(cs, task_ctrlbuf_va & 0xFFFFFF00);
4038 /* bits [31:0]: control buffer address hi */
4039 radeon_emit(cs, task_ctrlbuf_va >> 32);
4043 radv_emit_graphics_scratch(struct radv_device *device, struct radeon_cmdbuf *cs,
4044 uint32_t size_per_wave, uint32_t waves,
4045 struct radeon_winsys_bo *scratch_bo)
4047 struct radeon_info *info = &device->physical_device->rad_info;
4052 radv_cs_add_buffer(device->ws, cs, scratch_bo);
4054 if (info->gfx_level >= GFX11) {
4055 uint64_t va = radv_buffer_get_va(scratch_bo);
4057 /* WAVES is per SE for SPI_TMPRING_SIZE. */
4058 waves /= info->num_se;
4060 radeon_set_context_reg_seq(cs, R_0286E8_SPI_TMPRING_SIZE, 3);
4061 radeon_emit(cs, S_0286E8_WAVES(waves) | S_0286E8_WAVESIZE(round_up_u32(size_per_wave, 256)));
4062 radeon_emit(cs, va >> 8); /* SPI_GFX_SCRATCH_BASE_LO */
4063 radeon_emit(cs, va >> 40); /* SPI_GFX_SCRATCH_BASE_HI */
4065 radeon_set_context_reg(
4066 cs, R_0286E8_SPI_TMPRING_SIZE,
4067 S_0286E8_WAVES(waves) | S_0286E8_WAVESIZE(round_up_u32(size_per_wave, 1024)));
4072 radv_emit_compute_scratch(struct radv_device *device, struct radeon_cmdbuf *cs,
4073 uint32_t size_per_wave, uint32_t waves,
4074 struct radeon_winsys_bo *compute_scratch_bo)
4076 struct radeon_info *info = &device->physical_device->rad_info;
4077 uint64_t scratch_va;
4080 if (!compute_scratch_bo)
4083 scratch_va = radv_buffer_get_va(compute_scratch_bo);
4084 rsrc1 = S_008F04_BASE_ADDRESS_HI(scratch_va >> 32);
4086 if (device->physical_device->rad_info.gfx_level >= GFX11)
4087 rsrc1 |= S_008F04_SWIZZLE_ENABLE_GFX11(1);
4089 rsrc1 |= S_008F04_SWIZZLE_ENABLE_GFX6(1);
4091 radv_cs_add_buffer(device->ws, cs, compute_scratch_bo);
4093 if (info->gfx_level >= GFX11) {
4094 radeon_set_sh_reg_seq(cs, R_00B840_COMPUTE_DISPATCH_SCRATCH_BASE_LO, 4);
4095 radeon_emit(cs, scratch_va >> 8);
4096 radeon_emit(cs, scratch_va >> 40);
4098 radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0, 2);
4101 radeon_emit(cs, scratch_va);
4102 radeon_emit(cs, rsrc1);
4104 radeon_set_sh_reg(cs, R_00B860_COMPUTE_TMPRING_SIZE,
4105 S_00B860_WAVES(waves) |
4106 S_00B860_WAVESIZE(round_up_u32(size_per_wave, info->gfx_level >= GFX11 ? 256 : 1024)));
4110 radv_emit_compute_shader_pointers(struct radv_device *device, struct radeon_cmdbuf *cs,
4111 struct radeon_winsys_bo *descriptor_bo)
4116 uint64_t va = radv_buffer_get_va(descriptor_bo);
4117 radv_cs_add_buffer(device->ws, cs, descriptor_bo);
4119 /* Compute shader user data 0-1 have the scratch pointer (unlike GFX shaders),
4120 * so emit the descriptor pointer to user data 2-3 instead (task_ring_offsets arg).
4122 radv_emit_shader_pointer(device, cs, R_00B908_COMPUTE_USER_DATA_2, va, true);
4126 radv_emit_graphics_shader_pointers(struct radv_device *device, struct radeon_cmdbuf *cs,
4127 struct radeon_winsys_bo *descriptor_bo)
4134 va = radv_buffer_get_va(descriptor_bo);
4136 radv_cs_add_buffer(device->ws, cs, descriptor_bo);
4138 if (device->physical_device->rad_info.gfx_level >= GFX11) {
4139 uint32_t regs[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0,
4140 R_00B420_SPI_SHADER_PGM_LO_HS,
4141 R_00B220_SPI_SHADER_PGM_LO_GS};
4143 for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
4144 radv_emit_shader_pointer(device, cs, regs[i], va, true);
4146 } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
4147 uint32_t regs[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0, R_00B130_SPI_SHADER_USER_DATA_VS_0,
4148 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS,
4149 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS};
4151 for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
4152 radv_emit_shader_pointer(device, cs, regs[i], va, true);
4154 } else if (device->physical_device->rad_info.gfx_level == GFX9) {
4155 uint32_t regs[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0, R_00B130_SPI_SHADER_USER_DATA_VS_0,
4156 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS,
4157 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS};
4159 for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
4160 radv_emit_shader_pointer(device, cs, regs[i], va, true);
4163 uint32_t regs[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0, R_00B130_SPI_SHADER_USER_DATA_VS_0,
4164 R_00B230_SPI_SHADER_USER_DATA_GS_0, R_00B330_SPI_SHADER_USER_DATA_ES_0,
4165 R_00B430_SPI_SHADER_USER_DATA_HS_0, R_00B530_SPI_SHADER_USER_DATA_LS_0};
4167 for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
4168 radv_emit_shader_pointer(device, cs, regs[i], va, true);
4174 radv_init_graphics_state(struct radeon_cmdbuf *cs, struct radv_device *device)
4176 if (device->gfx_init) {
4177 uint64_t va = radv_buffer_get_va(device->gfx_init);
4179 radeon_emit(cs, PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0));
4180 radeon_emit(cs, va);
4181 radeon_emit(cs, va >> 32);
4182 radeon_emit(cs, device->gfx_init_size_dw & 0xffff);
4184 radv_cs_add_buffer(device->ws, cs, device->gfx_init);
4186 si_emit_graphics(device, cs);
4191 radv_init_compute_state(struct radeon_cmdbuf *cs, struct radv_device *device)
4193 si_emit_compute(device, cs);
4197 radv_update_preamble_cs(struct radv_queue_state *queue, struct radv_device *device,
4198 const struct radv_queue_ring_info *needs)
4200 struct radeon_winsys *ws = device->ws;
4201 struct radeon_winsys_bo *scratch_bo = queue->scratch_bo;
4202 struct radeon_winsys_bo *descriptor_bo = queue->descriptor_bo;
4203 struct radeon_winsys_bo *compute_scratch_bo = queue->compute_scratch_bo;
4204 struct radeon_winsys_bo *esgs_ring_bo = queue->esgs_ring_bo;
4205 struct radeon_winsys_bo *gsvs_ring_bo = queue->gsvs_ring_bo;
4206 struct radeon_winsys_bo *tess_rings_bo = queue->tess_rings_bo;
4207 struct radeon_winsys_bo *task_rings_bo = queue->task_rings_bo;
4208 struct radeon_winsys_bo *mesh_scratch_ring_bo = queue->mesh_scratch_ring_bo;
4209 struct radeon_winsys_bo *gds_bo = queue->gds_bo;
4210 struct radeon_winsys_bo *gds_oa_bo = queue->gds_oa_bo;
4211 struct radeon_cmdbuf *dest_cs[3] = {0};
4212 const uint32_t ring_bo_flags = RADEON_FLAG_NO_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING;
4213 VkResult result = VK_SUCCESS;
4215 const bool add_sample_positions = !queue->ring_info.sample_positions && needs->sample_positions;
4216 const uint32_t scratch_size = needs->scratch_size_per_wave * needs->scratch_waves;
4217 const uint32_t queue_scratch_size =
4218 queue->ring_info.scratch_size_per_wave * queue->ring_info.scratch_waves;
4220 if (scratch_size > queue_scratch_size) {
4221 result = ws->buffer_create(ws, scratch_size, 4096, RADEON_DOMAIN_VRAM, ring_bo_flags,
4222 RADV_BO_PRIORITY_SCRATCH, 0, &scratch_bo);
4223 if (result != VK_SUCCESS)
4227 const uint32_t compute_scratch_size =
4228 needs->compute_scratch_size_per_wave * needs->compute_scratch_waves;
4229 const uint32_t compute_queue_scratch_size =
4230 queue->ring_info.compute_scratch_size_per_wave * queue->ring_info.compute_scratch_waves;
4231 if (compute_scratch_size > compute_queue_scratch_size) {
4232 result = ws->buffer_create(ws, compute_scratch_size, 4096, RADEON_DOMAIN_VRAM, ring_bo_flags,
4233 RADV_BO_PRIORITY_SCRATCH, 0, &compute_scratch_bo);
4234 if (result != VK_SUCCESS)
4238 if (needs->esgs_ring_size > queue->ring_info.esgs_ring_size) {
4239 result = ws->buffer_create(ws, needs->esgs_ring_size, 4096, RADEON_DOMAIN_VRAM, ring_bo_flags,
4240 RADV_BO_PRIORITY_SCRATCH, 0, &esgs_ring_bo);
4241 if (result != VK_SUCCESS)
4245 if (needs->gsvs_ring_size > queue->ring_info.gsvs_ring_size) {
4246 result = ws->buffer_create(ws, needs->gsvs_ring_size, 4096, RADEON_DOMAIN_VRAM, ring_bo_flags,
4247 RADV_BO_PRIORITY_SCRATCH, 0, &gsvs_ring_bo);
4248 if (result != VK_SUCCESS)
4252 if (!queue->ring_info.tess_rings && needs->tess_rings) {
4253 result = ws->buffer_create(
4254 ws, device->physical_device->hs.tess_offchip_ring_offset + device->physical_device->hs.tess_offchip_ring_size, 256,
4255 RADEON_DOMAIN_VRAM, ring_bo_flags, RADV_BO_PRIORITY_SCRATCH, 0, &tess_rings_bo);
4256 if (result != VK_SUCCESS)
4260 if (!queue->ring_info.task_rings && needs->task_rings) {
4261 assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
4263 /* We write the control buffer from the CPU, so need to grant CPU access to the BO.
4264 * The draw ring needs to be zero-initialized otherwise the ready bits will be incorrect.
4266 uint32_t task_rings_bo_flags =
4267 RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_ZERO_VRAM;
4269 result = ws->buffer_create(ws, device->physical_device->task_info.bo_size_bytes, 256,
4270 RADEON_DOMAIN_VRAM, task_rings_bo_flags, RADV_BO_PRIORITY_SCRATCH,
4272 if (result != VK_SUCCESS)
4275 result = radv_initialise_task_control_buffer(device, task_rings_bo);
4276 if (result != VK_SUCCESS)
4280 if (!queue->ring_info.mesh_scratch_ring && needs->mesh_scratch_ring) {
4281 assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
4283 ws->buffer_create(ws, RADV_MESH_SCRATCH_NUM_ENTRIES * RADV_MESH_SCRATCH_ENTRY_BYTES, 256,
4284 RADEON_DOMAIN_VRAM, ring_bo_flags, RADV_BO_PRIORITY_SCRATCH, 0, &mesh_scratch_ring_bo);
4286 if (result != VK_SUCCESS)
4290 if (!queue->ring_info.gds && needs->gds) {
4291 assert(device->physical_device->rad_info.gfx_level >= GFX10);
4293 /* 4 streamout GDS counters.
4294 * We need 256B (64 dw) of GDS, otherwise streamout hangs.
4296 result = ws->buffer_create(ws, 256, 4, RADEON_DOMAIN_GDS, ring_bo_flags,
4297 RADV_BO_PRIORITY_SCRATCH, 0, &gds_bo);
4298 if (result != VK_SUCCESS)
4302 if (!queue->ring_info.gds_oa && needs->gds_oa) {
4303 assert(device->physical_device->rad_info.gfx_level >= GFX10);
4305 result = ws->buffer_create(ws, 4, 1, RADEON_DOMAIN_OA, ring_bo_flags,
4306 RADV_BO_PRIORITY_SCRATCH, 0, &gds_oa_bo);
4307 if (result != VK_SUCCESS)
4311 /* Re-initialize the descriptor BO when any ring BOs changed.
4313 * Additionally, make sure to create the descriptor BO for the compute queue
4314 * when it uses the task shader rings. The task rings BO is shared between the
4315 * GFX and compute queues and already initialized here.
4317 if ((queue->qf == RADV_QUEUE_COMPUTE && !descriptor_bo && task_rings_bo) ||
4318 scratch_bo != queue->scratch_bo || esgs_ring_bo != queue->esgs_ring_bo ||
4319 gsvs_ring_bo != queue->gsvs_ring_bo || tess_rings_bo != queue->tess_rings_bo ||
4320 task_rings_bo != queue->task_rings_bo || mesh_scratch_ring_bo != queue->mesh_scratch_ring_bo ||
4321 add_sample_positions) {
4323 if (gsvs_ring_bo || esgs_ring_bo || tess_rings_bo || task_rings_bo || mesh_scratch_ring_bo || add_sample_positions) {
4324 size = 160; /* 2 dword + 2 padding + 4 dword * 9 */
4325 if (add_sample_positions)
4326 size += 128; /* 64+32+16+8 = 120 bytes */
4327 } else if (scratch_bo) {
4328 size = 8; /* 2 dword */
4331 result = ws->buffer_create(
4332 ws, size, 4096, RADEON_DOMAIN_VRAM,
4333 RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_READ_ONLY,
4334 RADV_BO_PRIORITY_DESCRIPTOR, 0, &descriptor_bo);
4335 if (result != VK_SUCCESS)
4339 if (descriptor_bo != queue->descriptor_bo) {
4340 uint32_t *map = (uint32_t *)ws->buffer_map(descriptor_bo);
4345 uint64_t scratch_va = radv_buffer_get_va(scratch_bo);
4346 uint32_t rsrc1 = S_008F04_BASE_ADDRESS_HI(scratch_va >> 32);
4348 if (device->physical_device->rad_info.gfx_level >= GFX11)
4349 rsrc1 |= S_008F04_SWIZZLE_ENABLE_GFX11(1);
4351 rsrc1 |= S_008F04_SWIZZLE_ENABLE_GFX6(1);
4353 map[0] = scratch_va;
4357 if (esgs_ring_bo || gsvs_ring_bo || tess_rings_bo || task_rings_bo || mesh_scratch_ring_bo || add_sample_positions)
4358 radv_fill_shader_rings(device, map, add_sample_positions, needs->esgs_ring_size,
4359 esgs_ring_bo, needs->gsvs_ring_size, gsvs_ring_bo, tess_rings_bo,
4360 task_rings_bo, mesh_scratch_ring_bo);
4362 ws->buffer_unmap(descriptor_bo);
4365 for (int i = 0; i < 3; ++i) {
4366 /* Don't create continue preamble when it's not necessary. */
4368 /* We only need the continue preamble when we can't use indirect buffers. */
4369 if (!(device->instance->debug_flags & RADV_DEBUG_NO_IBS) &&
4370 device->physical_device->rad_info.gfx_level >= GFX7)
4372 /* Continue preamble is unnecessary when no shader rings are used. */
4373 if (!needs->scratch_size_per_wave && !needs->compute_scratch_size_per_wave &&
4374 !needs->esgs_ring_size && !needs->gsvs_ring_size && !needs->tess_rings &&
4375 !needs->task_rings && !needs->mesh_scratch_ring && !needs->gds && !needs->gds_oa && !needs->sample_positions)
4379 enum rgp_flush_bits sqtt_flush_bits = 0;
4380 struct radeon_cmdbuf *cs = NULL;
4381 cs = ws->cs_create(ws, radv_queue_family_to_ring(device->physical_device, queue->qf));
4383 result = VK_ERROR_OUT_OF_HOST_MEMORY;
4390 radv_cs_add_buffer(ws, cs, scratch_bo);
4392 /* Emit initial configuration. */
4393 switch (queue->qf) {
4394 case RADV_QUEUE_GENERAL:
4395 radv_init_graphics_state(cs, device);
4397 if (esgs_ring_bo || gsvs_ring_bo || tess_rings_bo || task_rings_bo) {
4398 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
4399 radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
4401 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
4402 radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
4405 radv_emit_gs_ring_sizes(device, cs, esgs_ring_bo, needs->esgs_ring_size, gsvs_ring_bo,
4406 needs->gsvs_ring_size);
4407 radv_emit_tess_factor_ring(device, cs, tess_rings_bo);
4408 radv_emit_task_rings(device, cs, task_rings_bo, false);
4409 radv_emit_graphics_shader_pointers(device, cs, descriptor_bo);
4410 radv_emit_compute_scratch(device, cs, needs->compute_scratch_size_per_wave,
4411 needs->compute_scratch_waves, compute_scratch_bo);
4412 radv_emit_graphics_scratch(device, cs, needs->scratch_size_per_wave, needs->scratch_waves,
4415 case RADV_QUEUE_COMPUTE:
4416 radv_init_compute_state(cs, device);
4418 if (task_rings_bo) {
4419 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
4420 radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
4423 radv_emit_task_rings(device, cs, task_rings_bo, true);
4424 radv_emit_compute_shader_pointers(device, cs, descriptor_bo);
4425 radv_emit_compute_scratch(device, cs, needs->compute_scratch_size_per_wave,
4426 needs->compute_scratch_waves, compute_scratch_bo);
4433 radv_cs_add_buffer(ws, cs, gds_bo);
4435 radv_cs_add_buffer(ws, cs, gds_oa_bo);
4438 /* The two initial preambles have a cache flush at the beginning. */
4439 const enum amd_gfx_level gfx_level = device->physical_device->rad_info.gfx_level;
4440 const bool is_mec = queue->qf == RADV_QUEUE_COMPUTE && gfx_level >= GFX7;
4441 enum radv_cmd_flush_bits flush_bits = RADV_CMD_FLAG_INV_ICACHE | RADV_CMD_FLAG_INV_SCACHE |
4442 RADV_CMD_FLAG_INV_VCACHE | RADV_CMD_FLAG_INV_L2 |
4443 RADV_CMD_FLAG_START_PIPELINE_STATS;
4446 /* The full flush preamble should also wait for previous shader work to finish. */
4447 flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH;
4448 if (queue->qf == RADV_QUEUE_GENERAL)
4449 flush_bits |= RADV_CMD_FLAG_PS_PARTIAL_FLUSH;
4452 si_cs_emit_cache_flush(cs, gfx_level, NULL, 0, is_mec, flush_bits, &sqtt_flush_bits, 0);
4455 result = ws->cs_finalize(cs);
4456 if (result != VK_SUCCESS)
4460 if (queue->initial_full_flush_preamble_cs)
4461 ws->cs_destroy(queue->initial_full_flush_preamble_cs);
4463 if (queue->initial_preamble_cs)
4464 ws->cs_destroy(queue->initial_preamble_cs);
4466 if (queue->continue_preamble_cs)
4467 ws->cs_destroy(queue->continue_preamble_cs);
4469 queue->initial_full_flush_preamble_cs = dest_cs[0];
4470 queue->initial_preamble_cs = dest_cs[1];
4471 queue->continue_preamble_cs = dest_cs[2];
4473 if (scratch_bo != queue->scratch_bo) {
4474 if (queue->scratch_bo)
4475 ws->buffer_destroy(ws, queue->scratch_bo);
4476 queue->scratch_bo = scratch_bo;
4479 if (compute_scratch_bo != queue->compute_scratch_bo) {
4480 if (queue->compute_scratch_bo)
4481 ws->buffer_destroy(ws, queue->compute_scratch_bo);
4482 queue->compute_scratch_bo = compute_scratch_bo;
4485 if (esgs_ring_bo != queue->esgs_ring_bo) {
4486 if (queue->esgs_ring_bo)
4487 ws->buffer_destroy(ws, queue->esgs_ring_bo);
4488 queue->esgs_ring_bo = esgs_ring_bo;
4491 if (gsvs_ring_bo != queue->gsvs_ring_bo) {
4492 if (queue->gsvs_ring_bo)
4493 ws->buffer_destroy(ws, queue->gsvs_ring_bo);
4494 queue->gsvs_ring_bo = gsvs_ring_bo;
4497 if (descriptor_bo != queue->descriptor_bo) {
4498 if (queue->descriptor_bo)
4499 ws->buffer_destroy(ws, queue->descriptor_bo);
4500 queue->descriptor_bo = descriptor_bo;
4503 queue->tess_rings_bo = tess_rings_bo;
4504 queue->task_rings_bo = task_rings_bo;
4505 queue->mesh_scratch_ring_bo = mesh_scratch_ring_bo;
4506 queue->gds_bo = gds_bo;
4507 queue->gds_oa_bo = gds_oa_bo;
4508 queue->ring_info = *needs;
4511 for (int i = 0; i < ARRAY_SIZE(dest_cs); ++i)
4513 ws->cs_destroy(dest_cs[i]);
4514 if (descriptor_bo && descriptor_bo != queue->descriptor_bo)
4515 ws->buffer_destroy(ws, descriptor_bo);
4516 if (scratch_bo && scratch_bo != queue->scratch_bo)
4517 ws->buffer_destroy(ws, scratch_bo);
4518 if (compute_scratch_bo && compute_scratch_bo != queue->compute_scratch_bo)
4519 ws->buffer_destroy(ws, compute_scratch_bo);
4520 if (esgs_ring_bo && esgs_ring_bo != queue->esgs_ring_bo)
4521 ws->buffer_destroy(ws, esgs_ring_bo);
4522 if (gsvs_ring_bo && gsvs_ring_bo != queue->gsvs_ring_bo)
4523 ws->buffer_destroy(ws, gsvs_ring_bo);
4524 if (tess_rings_bo && tess_rings_bo != queue->tess_rings_bo)
4525 ws->buffer_destroy(ws, tess_rings_bo);
4526 if (task_rings_bo && task_rings_bo != queue->task_rings_bo)
4527 ws->buffer_destroy(ws, task_rings_bo);
4528 if (gds_bo && gds_bo != queue->gds_bo)
4529 ws->buffer_destroy(ws, gds_bo);
4530 if (gds_oa_bo && gds_oa_bo != queue->gds_oa_bo)
4531 ws->buffer_destroy(ws, gds_oa_bo);
4533 return vk_error(queue, result);
4536 static struct radeon_cmdbuf *
4537 radv_create_perf_counter_lock_cs(struct radv_device *device, unsigned pass, bool unlock)
4539 struct radeon_cmdbuf **cs_ref = &device->perf_counter_lock_cs[pass * 2 + (unlock ? 1 : 0)];
4540 struct radeon_cmdbuf *cs;
4545 cs = device->ws->cs_create(device->ws, AMD_IP_GFX);
4549 ASSERTED unsigned cdw = radeon_check_space(device->ws, cs, 21);
4552 uint64_t mutex_va = radv_buffer_get_va(device->perf_counter_bo) + PERF_CTR_BO_LOCK_OFFSET;
4553 radeon_emit(cs, PKT3(PKT3_ATOMIC_MEM, 7, 0));
4554 radeon_emit(cs, ATOMIC_OP(TC_OP_ATOMIC_CMPSWAP_32) | ATOMIC_COMMAND(ATOMIC_COMMAND_LOOP));
4555 radeon_emit(cs, mutex_va); /* addr lo */
4556 radeon_emit(cs, mutex_va >> 32); /* addr hi */
4557 radeon_emit(cs, 1); /* data lo */
4558 radeon_emit(cs, 0); /* data hi */
4559 radeon_emit(cs, 0); /* compare data lo */
4560 radeon_emit(cs, 0); /* compare data hi */
4561 radeon_emit(cs, 10); /* loop interval */
4564 uint64_t va = radv_buffer_get_va(device->perf_counter_bo) + PERF_CTR_BO_PASS_OFFSET;
4565 uint64_t unset_va = va + (unlock ? 8 * pass : 0);
4566 uint64_t set_va = va + (unlock ? 0 : 8 * pass);
4568 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
4569 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_IMM) | COPY_DATA_DST_SEL(COPY_DATA_DST_MEM) |
4570 COPY_DATA_COUNT_SEL | COPY_DATA_WR_CONFIRM);
4571 radeon_emit(cs, 0); /* immediate */
4573 radeon_emit(cs, unset_va);
4574 radeon_emit(cs, unset_va >> 32);
4576 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
4577 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_IMM) | COPY_DATA_DST_SEL(COPY_DATA_DST_MEM) |
4578 COPY_DATA_COUNT_SEL | COPY_DATA_WR_CONFIRM);
4579 radeon_emit(cs, 1); /* immediate */
4581 radeon_emit(cs, set_va);
4582 radeon_emit(cs, set_va >> 32);
4585 uint64_t mutex_va = radv_buffer_get_va(device->perf_counter_bo) + PERF_CTR_BO_LOCK_OFFSET;
4587 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
4588 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_IMM) | COPY_DATA_DST_SEL(COPY_DATA_DST_MEM) |
4589 COPY_DATA_COUNT_SEL | COPY_DATA_WR_CONFIRM);
4590 radeon_emit(cs, 0); /* immediate */
4592 radeon_emit(cs, mutex_va);
4593 radeon_emit(cs, mutex_va >> 32);
4596 assert(cs->cdw <= cdw);
4598 VkResult result = device->ws->cs_finalize(cs);
4599 if (result != VK_SUCCESS) {
4600 device->ws->cs_destroy(cs);
4604 /* All the casts are to avoid MSVC errors around pointer truncation in a non-taken
4607 if (p_atomic_cmpxchg((uintptr_t*)cs_ref, 0, (uintptr_t)cs) != 0) {
4608 device->ws->cs_destroy(cs);
4615 radv_sparse_buffer_bind_memory(struct radv_device *device, const VkSparseBufferMemoryBindInfo *bind)
4617 RADV_FROM_HANDLE(radv_buffer, buffer, bind->buffer);
4620 for (uint32_t i = 0; i < bind->bindCount; ++i) {
4621 struct radv_device_memory *mem = NULL;
4623 if (bind->pBinds[i].memory != VK_NULL_HANDLE)
4624 mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
4626 result = device->ws->buffer_virtual_bind(device->ws, buffer->bo,
4627 bind->pBinds[i].resourceOffset, bind->pBinds[i].size,
4628 mem ? mem->bo : NULL, bind->pBinds[i].memoryOffset);
4629 if (result != VK_SUCCESS)
4637 radv_sparse_image_opaque_bind_memory(struct radv_device *device,
4638 const VkSparseImageOpaqueMemoryBindInfo *bind)
4640 RADV_FROM_HANDLE(radv_image, image, bind->image);
4643 for (uint32_t i = 0; i < bind->bindCount; ++i) {
4644 struct radv_device_memory *mem = NULL;
4646 if (bind->pBinds[i].memory != VK_NULL_HANDLE)
4647 mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
4649 result = device->ws->buffer_virtual_bind(device->ws, image->bo,
4650 bind->pBinds[i].resourceOffset, bind->pBinds[i].size,
4651 mem ? mem->bo : NULL, bind->pBinds[i].memoryOffset);
4652 if (result != VK_SUCCESS)
4660 radv_sparse_image_bind_memory(struct radv_device *device, const VkSparseImageMemoryBindInfo *bind)
4662 RADV_FROM_HANDLE(radv_image, image, bind->image);
4663 struct radeon_surf *surface = &image->planes[0].surface;
4664 uint32_t bs = vk_format_get_blocksize(image->vk.format);
4667 for (uint32_t i = 0; i < bind->bindCount; ++i) {
4668 struct radv_device_memory *mem = NULL;
4669 uint32_t offset, pitch;
4670 uint32_t mem_offset = bind->pBinds[i].memoryOffset;
4671 const uint32_t layer = bind->pBinds[i].subresource.arrayLayer;
4672 const uint32_t level = bind->pBinds[i].subresource.mipLevel;
4674 VkExtent3D bind_extent = bind->pBinds[i].extent;
4676 DIV_ROUND_UP(bind_extent.width, vk_format_get_blockwidth(image->vk.format));
4677 bind_extent.height =
4678 DIV_ROUND_UP(bind_extent.height, vk_format_get_blockheight(image->vk.format));
4680 VkOffset3D bind_offset = bind->pBinds[i].offset;
4681 bind_offset.x /= vk_format_get_blockwidth(image->vk.format);
4682 bind_offset.y /= vk_format_get_blockheight(image->vk.format);
4684 if (bind->pBinds[i].memory != VK_NULL_HANDLE)
4685 mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
4687 if (device->physical_device->rad_info.gfx_level >= GFX9) {
4688 offset = surface->u.gfx9.surf_slice_size * layer + surface->u.gfx9.prt_level_offset[level];
4689 pitch = surface->u.gfx9.prt_level_pitch[level];
4691 offset = (uint64_t)surface->u.legacy.level[level].offset_256B * 256 +
4692 surface->u.legacy.level[level].slice_size_dw * 4 * layer;
4693 pitch = surface->u.legacy.level[level].nblk_x;
4696 offset += (bind_offset.y * pitch * bs) + (bind_offset.x * surface->prt_tile_height * bs);
4698 uint32_t aligned_extent_width = ALIGN(bind_extent.width, surface->prt_tile_width);
4700 bool whole_subres = bind_offset.x == 0 && aligned_extent_width == pitch;
4703 uint32_t aligned_extent_height = ALIGN(bind_extent.height, surface->prt_tile_height);
4705 uint32_t size = aligned_extent_width * aligned_extent_height * bs;
4706 result = device->ws->buffer_virtual_bind(device->ws, image->bo, offset, size,
4707 mem ? mem->bo : NULL, mem_offset);
4708 if (result != VK_SUCCESS)
4711 uint32_t img_increment = pitch * bs;
4712 uint32_t mem_increment = aligned_extent_width * bs;
4713 uint32_t size = mem_increment * surface->prt_tile_height;
4714 for (unsigned y = 0; y < bind_extent.height; y += surface->prt_tile_height) {
4715 result = device->ws->buffer_virtual_bind(
4716 device->ws, image->bo, offset + img_increment * y, size, mem ? mem->bo : NULL,
4717 mem_offset + mem_increment * y);
4718 if (result != VK_SUCCESS)
4728 radv_update_preambles(struct radv_queue_state *queue, struct radv_device *device,
4729 struct vk_command_buffer *const *cmd_buffers, uint32_t cmd_buffer_count,
4730 bool *uses_perf_counters)
4732 if (queue->qf == RADV_QUEUE_TRANSFER)
4735 /* Figure out the needs of the current submission.
4736 * Start by copying the queue's current info.
4737 * This is done because we only allow two possible behaviours for these buffers:
4738 * - Grow when the newly needed amount is larger than what we had
4739 * - Allocate the max size and reuse it, but don't free it until the queue is destroyed
4741 struct radv_queue_ring_info needs = queue->ring_info;
4742 *uses_perf_counters = false;
4743 for (uint32_t j = 0; j < cmd_buffer_count; j++) {
4744 struct radv_cmd_buffer *cmd_buffer = container_of(cmd_buffers[j], struct radv_cmd_buffer, vk);
4746 needs.scratch_size_per_wave =
4747 MAX2(needs.scratch_size_per_wave, cmd_buffer->scratch_size_per_wave_needed);
4748 needs.scratch_waves = MAX2(needs.scratch_waves, cmd_buffer->scratch_waves_wanted);
4749 needs.compute_scratch_size_per_wave = MAX2(needs.compute_scratch_size_per_wave,
4750 cmd_buffer->compute_scratch_size_per_wave_needed);
4751 needs.compute_scratch_waves =
4752 MAX2(needs.compute_scratch_waves, cmd_buffer->compute_scratch_waves_wanted);
4753 needs.esgs_ring_size = MAX2(needs.esgs_ring_size, cmd_buffer->esgs_ring_size_needed);
4754 needs.gsvs_ring_size = MAX2(needs.gsvs_ring_size, cmd_buffer->gsvs_ring_size_needed);
4755 needs.tess_rings |= cmd_buffer->tess_rings_needed;
4756 needs.task_rings |= cmd_buffer->task_rings_needed;
4757 needs.mesh_scratch_ring |= cmd_buffer->mesh_scratch_ring_needed;
4758 needs.gds |= cmd_buffer->gds_needed;
4759 needs.gds_oa |= cmd_buffer->gds_oa_needed;
4760 needs.sample_positions |= cmd_buffer->sample_positions_needed;
4761 *uses_perf_counters |= cmd_buffer->state.uses_perf_counters;
4764 /* Sanitize scratch size information. */
4765 needs.scratch_waves = needs.scratch_size_per_wave
4766 ? MIN2(needs.scratch_waves, UINT32_MAX / needs.scratch_size_per_wave)
4768 needs.compute_scratch_waves =
4769 needs.compute_scratch_size_per_wave
4770 ? MIN2(needs.compute_scratch_waves, UINT32_MAX / needs.compute_scratch_size_per_wave)
4773 /* Return early if we already match these needs.
4774 * Note that it's not possible for any of the needed values to be less
4775 * than what the queue already had, because we only ever increase the allocated size.
4777 if (queue->initial_full_flush_preamble_cs &&
4778 queue->ring_info.scratch_size_per_wave == needs.scratch_size_per_wave &&
4779 queue->ring_info.scratch_waves == needs.scratch_waves &&
4780 queue->ring_info.compute_scratch_size_per_wave == needs.compute_scratch_size_per_wave &&
4781 queue->ring_info.compute_scratch_waves == needs.compute_scratch_waves &&
4782 queue->ring_info.esgs_ring_size == needs.esgs_ring_size &&
4783 queue->ring_info.gsvs_ring_size == needs.gsvs_ring_size &&
4784 queue->ring_info.tess_rings == needs.tess_rings &&
4785 queue->ring_info.task_rings == needs.task_rings &&
4786 queue->ring_info.mesh_scratch_ring == needs.mesh_scratch_ring &&
4787 queue->ring_info.gds == needs.gds &&
4788 queue->ring_info.gds_oa == needs.gds_oa &&
4789 queue->ring_info.sample_positions == needs.sample_positions)
4792 return radv_update_preamble_cs(queue, device, &needs);
4795 struct radv_deferred_queue_submission {
4796 struct radv_queue *queue;
4797 VkCommandBuffer *cmd_buffers;
4798 uint32_t cmd_buffer_count;
4800 /* Sparse bindings that happen on a queue. */
4801 VkSparseBufferMemoryBindInfo *buffer_binds;
4802 uint32_t buffer_bind_count;
4803 VkSparseImageOpaqueMemoryBindInfo *image_opaque_binds;
4804 uint32_t image_opaque_bind_count;
4805 VkSparseImageMemoryBindInfo *image_binds;
4806 uint32_t image_bind_count;
4809 VkPipelineStageFlags2 wait_dst_stage_mask;
4810 struct radv_semaphore_part **wait_semaphores;
4811 uint32_t wait_semaphore_count;
4812 struct radv_semaphore_part **signal_semaphores;
4813 uint32_t signal_semaphore_count;
4816 uint64_t *wait_values;
4817 uint64_t *signal_values;
4819 struct radv_semaphore_part *temporary_semaphore_parts;
4820 uint32_t temporary_semaphore_part_count;
4822 struct list_head queue_pending_list;
4823 uint32_t submission_wait_count;
4825 struct list_head processing_list;
4829 radv_queue_submit_bind_sparse_memory(struct radv_device *device, struct vk_queue_submit *submission)
4831 for (uint32_t i = 0; i < submission->buffer_bind_count; ++i) {
4832 VkResult result = radv_sparse_buffer_bind_memory(device, submission->buffer_binds + i);
4833 if (result != VK_SUCCESS)
4837 for (uint32_t i = 0; i < submission->image_opaque_bind_count; ++i) {
4839 radv_sparse_image_opaque_bind_memory(device, submission->image_opaque_binds + i);
4840 if (result != VK_SUCCESS)
4844 for (uint32_t i = 0; i < submission->image_bind_count; ++i) {
4845 VkResult result = radv_sparse_image_bind_memory(device, submission->image_binds + i);
4846 if (result != VK_SUCCESS)
4854 radv_queue_submit_empty(struct radv_queue *queue, struct vk_queue_submit *submission)
4856 struct radeon_winsys_ctx *ctx = queue->hw_ctx;
4857 struct radv_winsys_submit_info submit = {
4858 .ip_type = radv_queue_ring(queue),
4859 .queue_index = queue->vk.index_in_family,
4862 return queue->device->ws->cs_submit(ctx, 1, &submit, submission->wait_count, submission->waits,
4863 submission->signal_count, submission->signals, false);
4867 radv_queue_submit_normal(struct radv_queue *queue, struct vk_queue_submit *submission)
4869 struct radeon_winsys_ctx *ctx = queue->hw_ctx;
4870 uint32_t max_cs_submission = queue->device->trace_bo ? 1 : RADV_MAX_IBS_PER_SUBMIT;
4871 bool can_patch = true;
4874 bool uses_perf_counters = false;
4875 unsigned cmd_buffer_count = submission->command_buffer_count;
4877 result = radv_update_preambles(&queue->state, queue->device, submission->command_buffers,
4878 submission->command_buffer_count, &uses_perf_counters);
4879 if (result != VK_SUCCESS)
4882 if (queue->device->trace_bo)
4883 simple_mtx_lock(&queue->device->trace_mtx);
4885 if (uses_perf_counters)
4886 cmd_buffer_count += 2;
4888 struct radeon_cmdbuf **cs_array = malloc(sizeof(struct radeon_cmdbuf *) * cmd_buffer_count);
4892 for (uint32_t j = 0; j < submission->command_buffer_count; j++) {
4893 struct radv_cmd_buffer *cmd_buffer = (struct radv_cmd_buffer *)submission->command_buffers[j];
4894 assert(cmd_buffer->vk.level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
4896 cs_array[j + (uses_perf_counters ? 1 : 0)] = cmd_buffer->cs;
4897 if ((cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT))
4900 cmd_buffer->status = RADV_CMD_BUFFER_STATUS_PENDING;
4903 if (uses_perf_counters) {
4905 radv_create_perf_counter_lock_cs(queue->device, submission->perf_pass_index, false);
4906 cs_array[cmd_buffer_count - 1] =
4907 radv_create_perf_counter_lock_cs(queue->device, submission->perf_pass_index, true);
4909 if (!cs_array[0] || !cs_array[cmd_buffer_count - 1]) {
4910 result = VK_ERROR_OUT_OF_HOST_MEMORY;
4915 /* For fences on the same queue/vm amdgpu doesn't wait till all processing is finished
4916 * before starting the next cmdbuffer, so we need to do it here. */
4917 bool need_wait = submission->wait_count > 0;
4919 struct radv_winsys_submit_info submit = {
4920 .ip_type = radv_queue_ring(queue),
4921 .queue_index = queue->vk.index_in_family,
4922 .cs_array = cs_array,
4924 .initial_preamble_cs =
4925 need_wait ? queue->state.initial_full_flush_preamble_cs : queue->state.initial_preamble_cs,
4926 .continue_preamble_cs = queue->state.continue_preamble_cs,
4929 for (uint32_t j = 0; j < cmd_buffer_count; j += advance) {
4930 advance = MIN2(max_cs_submission, cmd_buffer_count - j);
4931 bool last_submit = j + advance == cmd_buffer_count;
4933 if (queue->device->trace_bo)
4934 *queue->device->trace_id_ptr = 0;
4936 submit.cs_count = advance;
4938 result = queue->device->ws->cs_submit(
4939 ctx, 1, &submit, j == 0 ? submission->wait_count : 0, submission->waits,
4940 last_submit ? submission->signal_count : 0, submission->signals, can_patch);
4942 if (result != VK_SUCCESS)
4945 if (queue->device->trace_bo) {
4946 radv_check_gpu_hangs(queue, cs_array[j]);
4949 if (queue->device->tma_bo) {
4950 radv_check_trap_handler(queue);
4953 submit.cs_array += advance;
4954 submit.initial_preamble_cs = queue->state.initial_preamble_cs;
4959 if (queue->device->trace_bo)
4960 simple_mtx_unlock(&queue->device->trace_mtx);
4966 radv_queue_submit(struct vk_queue *vqueue, struct vk_queue_submit *submission)
4968 struct radv_queue *queue = (struct radv_queue *)vqueue;
4971 result = radv_queue_submit_bind_sparse_memory(queue->device, submission);
4972 if (result != VK_SUCCESS)
4975 if (!submission->command_buffer_count && !submission->wait_count && !submission->signal_count)
4978 if (!submission->command_buffer_count) {
4979 result = radv_queue_submit_empty(queue, submission);
4981 result = radv_queue_submit_normal(queue, submission);
4985 if (result != VK_SUCCESS && result != VK_ERROR_DEVICE_LOST) {
4986 /* When something bad happened during the submission, such as
4987 * an out of memory issue, it might be hard to recover from
4988 * this inconsistent state. To avoid this sort of problem, we
4989 * assume that we are in a really bad situation and return
4990 * VK_ERROR_DEVICE_LOST to ensure the clients do not attempt
4991 * to submit the same job again to this device.
4993 result = vk_device_set_lost(&queue->device->vk, "vkQueueSubmit() failed");
4999 radv_queue_internal_submit(struct radv_queue *queue, struct radeon_cmdbuf *cs)
5001 struct radeon_winsys_ctx *ctx = queue->hw_ctx;
5002 struct radv_winsys_submit_info submit = {
5003 .ip_type = radv_queue_ring(queue),
5004 .queue_index = queue->vk.index_in_family,
5009 VkResult result = queue->device->ws->cs_submit(ctx, 1, &submit, 0, NULL, 0, NULL, false);
5010 if (result != VK_SUCCESS)
5016 VKAPI_ATTR VkResult VKAPI_CALL
5017 radv_EnumerateInstanceExtensionProperties(const char *pLayerName, uint32_t *pPropertyCount,
5018 VkExtensionProperties *pProperties)
5021 return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
5023 return vk_enumerate_instance_extension_properties(&radv_instance_extensions_supported,
5024 pPropertyCount, pProperties);
5027 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
5028 radv_GetInstanceProcAddr(VkInstance _instance, const char *pName)
5030 RADV_FROM_HANDLE(radv_instance, instance, _instance);
5032 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
5033 * when we have to return valid function pointers, NULL, or it's left
5034 * undefined. See the table for exact details.
5039 #define LOOKUP_RADV_ENTRYPOINT(entrypoint) \
5040 if (strcmp(pName, "vk" #entrypoint) == 0) \
5041 return (PFN_vkVoidFunction)radv_##entrypoint
5043 LOOKUP_RADV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
5044 LOOKUP_RADV_ENTRYPOINT(EnumerateInstanceLayerProperties);
5045 LOOKUP_RADV_ENTRYPOINT(EnumerateInstanceVersion);
5046 LOOKUP_RADV_ENTRYPOINT(CreateInstance);
5048 /* GetInstanceProcAddr() can also be called with a NULL instance.
5049 * See https://gitlab.khronos.org/vulkan/vulkan/issues/2057
5051 LOOKUP_RADV_ENTRYPOINT(GetInstanceProcAddr);
5053 #undef LOOKUP_RADV_ENTRYPOINT
5055 if (instance == NULL)
5058 return vk_instance_get_proc_addr(&instance->vk, &radv_instance_entrypoints, pName);
5061 /* Windows will use a dll definition file to avoid build errors. */
5067 /* The loader wants us to expose a second GetInstanceProcAddr function
5068 * to work around certain LD_PRELOAD issues seen in apps.
5071 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
5072 vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName)
5074 return radv_GetInstanceProcAddr(instance, pName);
5078 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
5079 vk_icdGetPhysicalDeviceProcAddr(VkInstance _instance, const char *pName)
5081 RADV_FROM_HANDLE(radv_instance, instance, _instance);
5082 return vk_instance_get_physical_device_proc_addr(&instance->vk, pName);
5086 radv_get_memory_fd(struct radv_device *device, struct radv_device_memory *memory, int *pFD)
5088 /* Only set BO metadata for the first plane */
5089 if (memory->image && memory->image->offset == 0) {
5090 struct radeon_bo_metadata metadata;
5091 radv_init_metadata(device, memory->image, &metadata);
5092 device->ws->buffer_set_metadata(device->ws, memory->bo, &metadata);
5095 return device->ws->buffer_get_fd(device->ws, memory->bo, pFD);
5099 radv_device_memory_init(struct radv_device_memory *mem, struct radv_device *device,
5100 struct radeon_winsys_bo *bo)
5102 memset(mem, 0, sizeof(*mem));
5103 vk_object_base_init(&device->vk, &mem->base, VK_OBJECT_TYPE_DEVICE_MEMORY);
5109 radv_device_memory_finish(struct radv_device_memory *mem)
5111 vk_object_base_finish(&mem->base);
5115 radv_free_memory(struct radv_device *device, const VkAllocationCallbacks *pAllocator,
5116 struct radv_device_memory *mem)
5121 #if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
5122 if (mem->android_hardware_buffer)
5123 AHardwareBuffer_release(mem->android_hardware_buffer);
5127 if (device->overallocation_disallowed) {
5128 mtx_lock(&device->overallocation_mutex);
5129 device->allocated_memory_size[mem->heap_index] -= mem->alloc_size;
5130 mtx_unlock(&device->overallocation_mutex);
5133 if (device->use_global_bo_list)
5134 device->ws->buffer_make_resident(device->ws, mem->bo, false);
5135 device->ws->buffer_destroy(device->ws, mem->bo);
5139 radv_device_memory_finish(mem);
5140 vk_free2(&device->vk.alloc, pAllocator, mem);
5144 radv_alloc_memory(struct radv_device *device, const VkMemoryAllocateInfo *pAllocateInfo,
5145 const VkAllocationCallbacks *pAllocator, VkDeviceMemory *pMem)
5147 struct radv_device_memory *mem;
5149 enum radeon_bo_domain domain;
5152 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
5154 const VkImportMemoryFdInfoKHR *import_info =
5155 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
5156 const VkMemoryDedicatedAllocateInfo *dedicate_info =
5157 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO);
5158 const VkExportMemoryAllocateInfo *export_info =
5159 vk_find_struct_const(pAllocateInfo->pNext, EXPORT_MEMORY_ALLOCATE_INFO);
5160 const struct VkImportAndroidHardwareBufferInfoANDROID *ahb_import_info =
5161 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID);
5162 const VkImportMemoryHostPointerInfoEXT *host_ptr_info =
5163 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_HOST_POINTER_INFO_EXT);
5165 const struct wsi_memory_allocate_info *wsi_info =
5166 vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
5168 if (pAllocateInfo->allocationSize == 0 && !ahb_import_info &&
5169 !(export_info && (export_info->handleTypes &
5170 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID))) {
5171 /* Apparently, this is allowed */
5172 *pMem = VK_NULL_HANDLE;
5177 vk_alloc2(&device->vk.alloc, pAllocator, sizeof(*mem), 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
5179 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
5181 radv_device_memory_init(mem, device, NULL);
5184 if(wsi_info->implicit_sync)
5185 flags |= RADEON_FLAG_IMPLICIT_SYNC;
5187 /* In case of prime, linear buffer is allocated in default heap which is VRAM.
5188 * Due to this when display is connected to iGPU and render on dGPU, ddx
5189 * function amdgpu_present_check_flip() fails due to which there is blit
5190 * instead of flip. Setting the flag RADEON_FLAG_GTT_WC allows kernel to
5191 * allocate GTT memory in supported hardware where GTT can be directly scanout.
5192 * Using wsi_info variable check to set the flag RADEON_FLAG_GTT_WC so that
5193 * only for memory allocated by driver this flag is set.
5195 flags |= RADEON_FLAG_GTT_WC;
5198 if (dedicate_info) {
5199 mem->image = radv_image_from_handle(dedicate_info->image);
5200 mem->buffer = radv_buffer_from_handle(dedicate_info->buffer);
5206 if (wsi_info && wsi_info->implicit_sync && mem->buffer) {
5207 /* Mark the linear prime buffer (aka the destination of the prime blit
5210 flags |= RADEON_FLAG_VA_UNCACHED;
5213 float priority_float = 0.5;
5214 const struct VkMemoryPriorityAllocateInfoEXT *priority_ext =
5215 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_PRIORITY_ALLOCATE_INFO_EXT);
5217 priority_float = priority_ext->priority;
5219 uint64_t replay_address = 0;
5220 const VkMemoryOpaqueCaptureAddressAllocateInfo *replay_info =
5221 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO);
5222 if (replay_info && replay_info->opaqueCaptureAddress)
5223 replay_address = replay_info->opaqueCaptureAddress;
5225 unsigned priority = MIN2(RADV_BO_PRIORITY_APPLICATION_MAX - 1,
5226 (int)(priority_float * RADV_BO_PRIORITY_APPLICATION_MAX));
5228 mem->user_ptr = NULL;
5230 #if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
5231 mem->android_hardware_buffer = NULL;
5234 if (ahb_import_info) {
5235 result = radv_import_ahb_memory(device, mem, priority, ahb_import_info);
5236 if (result != VK_SUCCESS)
5238 } else if (export_info && (export_info->handleTypes &
5239 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)) {
5240 result = radv_create_ahb_memory(device, mem, priority, pAllocateInfo);
5241 if (result != VK_SUCCESS)
5243 } else if (import_info) {
5244 assert(import_info->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
5245 import_info->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
5246 result = device->ws->buffer_from_fd(device->ws, import_info->fd, priority, &mem->bo, NULL);
5247 if (result != VK_SUCCESS) {
5250 close(import_info->fd);
5253 if (mem->image && mem->image->plane_count == 1 &&
5254 !vk_format_is_depth_or_stencil(mem->image->vk.format) && mem->image->info.samples == 1 &&
5255 mem->image->vk.tiling != VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT) {
5256 struct radeon_bo_metadata metadata;
5257 device->ws->buffer_get_metadata(device->ws, mem->bo, &metadata);
5259 struct radv_image_create_info create_info = {.no_metadata_planes = true,
5260 .bo_metadata = &metadata};
5262 /* This gives a basic ability to import radeonsi images
5263 * that don't have DCC. This is not guaranteed by any
5264 * spec and can be removed after we support modifiers. */
5265 result = radv_image_create_layout(device, create_info, NULL, mem->image);
5266 if (result != VK_SUCCESS) {
5267 device->ws->buffer_destroy(device->ws, mem->bo);
5271 } else if (host_ptr_info) {
5272 assert(host_ptr_info->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
5273 result = device->ws->buffer_from_ptr(device->ws, host_ptr_info->pHostPointer,
5274 pAllocateInfo->allocationSize, priority, &mem->bo);
5275 if (result != VK_SUCCESS) {
5278 mem->user_ptr = host_ptr_info->pHostPointer;
5281 uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
5282 uint32_t heap_index;
5285 device->physical_device->memory_properties.memoryTypes[pAllocateInfo->memoryTypeIndex]
5287 domain = device->physical_device->memory_domains[pAllocateInfo->memoryTypeIndex];
5288 flags |= device->physical_device->memory_flags[pAllocateInfo->memoryTypeIndex];
5290 if (!import_info && (!export_info || !export_info->handleTypes)) {
5291 flags |= RADEON_FLAG_NO_INTERPROCESS_SHARING;
5292 if (device->use_global_bo_list) {
5293 flags |= RADEON_FLAG_PREFER_LOCAL_BO;
5297 const VkMemoryAllocateFlagsInfo *flags_info = vk_find_struct_const(pAllocateInfo->pNext, MEMORY_ALLOCATE_FLAGS_INFO);
5298 if (flags_info && flags_info->flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT)
5299 flags |= RADEON_FLAG_REPLAYABLE;
5301 if (device->instance->zero_vram)
5302 flags |= RADEON_FLAG_ZERO_VRAM;
5304 if (device->overallocation_disallowed) {
5305 uint64_t total_size =
5306 device->physical_device->memory_properties.memoryHeaps[heap_index].size;
5308 mtx_lock(&device->overallocation_mutex);
5309 if (device->allocated_memory_size[heap_index] + alloc_size > total_size) {
5310 mtx_unlock(&device->overallocation_mutex);
5311 result = VK_ERROR_OUT_OF_DEVICE_MEMORY;
5314 device->allocated_memory_size[heap_index] += alloc_size;
5315 mtx_unlock(&device->overallocation_mutex);
5318 result = device->ws->buffer_create(device->ws, alloc_size,
5319 device->physical_device->rad_info.max_alignment, domain,
5320 flags, priority, replay_address, &mem->bo);
5322 if (result != VK_SUCCESS) {
5323 if (device->overallocation_disallowed) {
5324 mtx_lock(&device->overallocation_mutex);
5325 device->allocated_memory_size[heap_index] -= alloc_size;
5326 mtx_unlock(&device->overallocation_mutex);
5331 mem->heap_index = heap_index;
5332 mem->alloc_size = alloc_size;
5336 if (device->use_global_bo_list) {
5337 result = device->ws->buffer_make_resident(device->ws, mem->bo, true);
5338 if (result != VK_SUCCESS)
5343 *pMem = radv_device_memory_to_handle(mem);
5348 radv_free_memory(device, pAllocator, mem);
5353 VKAPI_ATTR VkResult VKAPI_CALL
5354 radv_AllocateMemory(VkDevice _device, const VkMemoryAllocateInfo *pAllocateInfo,
5355 const VkAllocationCallbacks *pAllocator, VkDeviceMemory *pMem)
5357 RADV_FROM_HANDLE(radv_device, device, _device);
5358 return radv_alloc_memory(device, pAllocateInfo, pAllocator, pMem);
5361 VKAPI_ATTR void VKAPI_CALL
5362 radv_FreeMemory(VkDevice _device, VkDeviceMemory _mem, const VkAllocationCallbacks *pAllocator)
5364 RADV_FROM_HANDLE(radv_device, device, _device);
5365 RADV_FROM_HANDLE(radv_device_memory, mem, _mem);
5367 radv_free_memory(device, pAllocator, mem);
5370 VKAPI_ATTR VkResult VKAPI_CALL
5371 radv_MapMemory(VkDevice _device, VkDeviceMemory _memory, VkDeviceSize offset, VkDeviceSize size,
5372 VkMemoryMapFlags flags, void **ppData)
5374 RADV_FROM_HANDLE(radv_device, device, _device);
5375 RADV_FROM_HANDLE(radv_device_memory, mem, _memory);
5378 *ppData = mem->user_ptr;
5380 *ppData = device->ws->buffer_map(mem->bo);
5383 *ppData = (uint8_t *)*ppData + offset;
5387 return vk_error(device, VK_ERROR_MEMORY_MAP_FAILED);
5390 VKAPI_ATTR void VKAPI_CALL
5391 radv_UnmapMemory(VkDevice _device, VkDeviceMemory _memory)
5393 RADV_FROM_HANDLE(radv_device, device, _device);
5394 RADV_FROM_HANDLE(radv_device_memory, mem, _memory);
5396 if (mem->user_ptr == NULL)
5397 device->ws->buffer_unmap(mem->bo);
5400 VKAPI_ATTR VkResult VKAPI_CALL
5401 radv_FlushMappedMemoryRanges(VkDevice _device, uint32_t memoryRangeCount,
5402 const VkMappedMemoryRange *pMemoryRanges)
5407 VKAPI_ATTR VkResult VKAPI_CALL
5408 radv_InvalidateMappedMemoryRanges(VkDevice _device, uint32_t memoryRangeCount,
5409 const VkMappedMemoryRange *pMemoryRanges)
5415 radv_get_buffer_memory_requirements(struct radv_device *device, VkDeviceSize size,
5416 VkBufferCreateFlags flags, VkBufferCreateFlags usage,
5417 VkMemoryRequirements2 *pMemoryRequirements)
5419 pMemoryRequirements->memoryRequirements.memoryTypeBits =
5420 (1u << device->physical_device->memory_properties.memoryTypeCount) - 1;
5422 if (flags & VK_BUFFER_CREATE_SPARSE_BINDING_BIT)
5423 pMemoryRequirements->memoryRequirements.alignment = 4096;
5425 pMemoryRequirements->memoryRequirements.alignment = 16;
5427 /* Top level acceleration structures need the bottom 6 bits to store
5428 * the root ids of instances. The hardware also needs bvh nodes to
5429 * be 64 byte aligned.
5431 if (usage & VK_BUFFER_USAGE_ACCELERATION_STRUCTURE_STORAGE_BIT_KHR)
5432 pMemoryRequirements->memoryRequirements.alignment =
5433 MAX2(pMemoryRequirements->memoryRequirements.alignment, 64);
5435 pMemoryRequirements->memoryRequirements.size =
5436 align64(size, pMemoryRequirements->memoryRequirements.alignment);
5438 vk_foreach_struct(ext, pMemoryRequirements->pNext)
5440 switch (ext->sType) {
5441 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
5442 VkMemoryDedicatedRequirements *req = (VkMemoryDedicatedRequirements *)ext;
5443 req->requiresDedicatedAllocation = false;
5444 req->prefersDedicatedAllocation = req->requiresDedicatedAllocation;
5453 VKAPI_ATTR void VKAPI_CALL
5454 radv_GetBufferMemoryRequirements2(VkDevice _device, const VkBufferMemoryRequirementsInfo2 *pInfo,
5455 VkMemoryRequirements2 *pMemoryRequirements)
5457 RADV_FROM_HANDLE(radv_device, device, _device);
5458 RADV_FROM_HANDLE(radv_buffer, buffer, pInfo->buffer);
5460 radv_get_buffer_memory_requirements(device, buffer->vk.size, buffer->vk.create_flags,
5461 buffer->vk.usage, pMemoryRequirements);
5464 VKAPI_ATTR void VKAPI_CALL
5465 radv_GetDeviceBufferMemoryRequirements(VkDevice _device,
5466 const VkDeviceBufferMemoryRequirements *pInfo,
5467 VkMemoryRequirements2 *pMemoryRequirements)
5469 RADV_FROM_HANDLE(radv_device, device, _device);
5471 radv_get_buffer_memory_requirements(device, pInfo->pCreateInfo->size, pInfo->pCreateInfo->flags,
5472 pInfo->pCreateInfo->usage, pMemoryRequirements);
5475 VKAPI_ATTR void VKAPI_CALL
5476 radv_GetImageMemoryRequirements2(VkDevice _device, const VkImageMemoryRequirementsInfo2 *pInfo,
5477 VkMemoryRequirements2 *pMemoryRequirements)
5479 RADV_FROM_HANDLE(radv_device, device, _device);
5480 RADV_FROM_HANDLE(radv_image, image, pInfo->image);
5482 pMemoryRequirements->memoryRequirements.memoryTypeBits =
5483 (1u << device->physical_device->memory_properties.memoryTypeCount) - 1;
5485 pMemoryRequirements->memoryRequirements.size = image->size;
5486 pMemoryRequirements->memoryRequirements.alignment = image->alignment;
5488 vk_foreach_struct(ext, pMemoryRequirements->pNext)
5490 switch (ext->sType) {
5491 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
5492 VkMemoryDedicatedRequirements *req = (VkMemoryDedicatedRequirements *)ext;
5493 req->requiresDedicatedAllocation =
5494 image->shareable && image->vk.tiling != VK_IMAGE_TILING_LINEAR;
5495 req->prefersDedicatedAllocation = req->requiresDedicatedAllocation;
5504 VKAPI_ATTR void VKAPI_CALL
5505 radv_GetDeviceImageMemoryRequirements(VkDevice device,
5506 const VkDeviceImageMemoryRequirements *pInfo,
5507 VkMemoryRequirements2 *pMemoryRequirements)
5509 UNUSED VkResult result;
5512 /* Determining the image size/alignment require to create a surface, which is complicated without
5513 * creating an image.
5514 * TODO: Avoid creating an image.
5516 result = radv_CreateImage(device, pInfo->pCreateInfo, NULL, &image);
5517 assert(result == VK_SUCCESS);
5519 VkImageMemoryRequirementsInfo2 info2 = {
5520 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
5524 radv_GetImageMemoryRequirements2(device, &info2, pMemoryRequirements);
5526 radv_DestroyImage(device, image, NULL);
5529 VKAPI_ATTR void VKAPI_CALL
5530 radv_GetDeviceMemoryCommitment(VkDevice device, VkDeviceMemory memory,
5531 VkDeviceSize *pCommittedMemoryInBytes)
5533 *pCommittedMemoryInBytes = 0;
5536 VKAPI_ATTR VkResult VKAPI_CALL
5537 radv_BindBufferMemory2(VkDevice _device, uint32_t bindInfoCount,
5538 const VkBindBufferMemoryInfo *pBindInfos)
5540 RADV_FROM_HANDLE(radv_device, device, _device);
5542 for (uint32_t i = 0; i < bindInfoCount; ++i) {
5543 RADV_FROM_HANDLE(radv_device_memory, mem, pBindInfos[i].memory);
5544 RADV_FROM_HANDLE(radv_buffer, buffer, pBindInfos[i].buffer);
5546 if (mem->alloc_size) {
5547 VkBufferMemoryRequirementsInfo2 info = {
5548 .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2,
5549 .buffer = pBindInfos[i].buffer,
5551 VkMemoryRequirements2 reqs = {
5552 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
5555 radv_GetBufferMemoryRequirements2(_device, &info, &reqs);
5557 if (pBindInfos[i].memoryOffset + reqs.memoryRequirements.size > mem->alloc_size) {
5558 return vk_errorf(device, VK_ERROR_UNKNOWN,
5559 "Device memory object too small for the buffer.\n");
5563 buffer->bo = mem->bo;
5564 buffer->offset = pBindInfos[i].memoryOffset;
5569 VKAPI_ATTR VkResult VKAPI_CALL
5570 radv_BindImageMemory2(VkDevice _device, uint32_t bindInfoCount,
5571 const VkBindImageMemoryInfo *pBindInfos)
5573 RADV_FROM_HANDLE(radv_device, device, _device);
5575 for (uint32_t i = 0; i < bindInfoCount; ++i) {
5576 RADV_FROM_HANDLE(radv_device_memory, mem, pBindInfos[i].memory);
5577 RADV_FROM_HANDLE(radv_image, image, pBindInfos[i].image);
5579 if (mem->alloc_size) {
5580 VkImageMemoryRequirementsInfo2 info = {
5581 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
5582 .image = pBindInfos[i].image,
5584 VkMemoryRequirements2 reqs = {
5585 .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
5588 radv_GetImageMemoryRequirements2(_device, &info, &reqs);
5590 if (pBindInfos[i].memoryOffset + reqs.memoryRequirements.size > mem->alloc_size) {
5591 return vk_errorf(device, VK_ERROR_UNKNOWN,
5592 "Device memory object too small for the image.\n");
5596 image->bo = mem->bo;
5597 image->offset = pBindInfos[i].memoryOffset;
5603 radv_destroy_event(struct radv_device *device, const VkAllocationCallbacks *pAllocator,
5604 struct radv_event *event)
5607 device->ws->buffer_destroy(device->ws, event->bo);
5609 vk_object_base_finish(&event->base);
5610 vk_free2(&device->vk.alloc, pAllocator, event);
5613 VKAPI_ATTR VkResult VKAPI_CALL
5614 radv_CreateEvent(VkDevice _device, const VkEventCreateInfo *pCreateInfo,
5615 const VkAllocationCallbacks *pAllocator, VkEvent *pEvent)
5617 RADV_FROM_HANDLE(radv_device, device, _device);
5618 enum radeon_bo_domain bo_domain;
5619 enum radeon_bo_flag bo_flags;
5620 struct radv_event *event;
5623 event = vk_alloc2(&device->vk.alloc, pAllocator, sizeof(*event), 8,
5624 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
5626 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
5628 vk_object_base_init(&device->vk, &event->base, VK_OBJECT_TYPE_EVENT);
5630 if (pCreateInfo->flags & VK_EVENT_CREATE_DEVICE_ONLY_BIT) {
5631 bo_domain = RADEON_DOMAIN_VRAM;
5632 bo_flags = RADEON_FLAG_NO_CPU_ACCESS;
5634 bo_domain = RADEON_DOMAIN_GTT;
5635 bo_flags = RADEON_FLAG_CPU_ACCESS;
5638 result = device->ws->buffer_create(
5639 device->ws, 8, 8, bo_domain,
5640 RADEON_FLAG_VA_UNCACHED | RADEON_FLAG_NO_INTERPROCESS_SHARING | bo_flags,
5641 RADV_BO_PRIORITY_FENCE, 0, &event->bo);
5642 if (result != VK_SUCCESS) {
5643 radv_destroy_event(device, pAllocator, event);
5644 return vk_error(device, result);
5647 if (!(pCreateInfo->flags & VK_EVENT_CREATE_DEVICE_ONLY_BIT)) {
5648 event->map = (uint64_t *)device->ws->buffer_map(event->bo);
5650 radv_destroy_event(device, pAllocator, event);
5651 return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
5655 *pEvent = radv_event_to_handle(event);
5660 VKAPI_ATTR void VKAPI_CALL
5661 radv_DestroyEvent(VkDevice _device, VkEvent _event, const VkAllocationCallbacks *pAllocator)
5663 RADV_FROM_HANDLE(radv_device, device, _device);
5664 RADV_FROM_HANDLE(radv_event, event, _event);
5669 radv_destroy_event(device, pAllocator, event);
5672 VKAPI_ATTR VkResult VKAPI_CALL
5673 radv_GetEventStatus(VkDevice _device, VkEvent _event)
5675 RADV_FROM_HANDLE(radv_device, device, _device);
5676 RADV_FROM_HANDLE(radv_event, event, _event);
5678 if (vk_device_is_lost(&device->vk))
5679 return VK_ERROR_DEVICE_LOST;
5681 if (*event->map == 1)
5682 return VK_EVENT_SET;
5683 return VK_EVENT_RESET;
5686 VKAPI_ATTR VkResult VKAPI_CALL
5687 radv_SetEvent(VkDevice _device, VkEvent _event)
5689 RADV_FROM_HANDLE(radv_event, event, _event);
5695 VKAPI_ATTR VkResult VKAPI_CALL
5696 radv_ResetEvent(VkDevice _device, VkEvent _event)
5698 RADV_FROM_HANDLE(radv_event, event, _event);
5705 radv_buffer_init(struct radv_buffer *buffer, struct radv_device *device,
5706 struct radeon_winsys_bo *bo, uint64_t size,
5709 VkBufferCreateInfo createInfo = {
5710 .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
5714 vk_buffer_init(&device->vk, &buffer->vk, &createInfo);
5717 buffer->offset = offset;
5721 radv_buffer_finish(struct radv_buffer *buffer)
5723 vk_buffer_finish(&buffer->vk);
5727 radv_destroy_buffer(struct radv_device *device, const VkAllocationCallbacks *pAllocator,
5728 struct radv_buffer *buffer)
5730 if ((buffer->vk.create_flags & VK_BUFFER_CREATE_SPARSE_BINDING_BIT) && buffer->bo)
5731 device->ws->buffer_destroy(device->ws, buffer->bo);
5733 radv_buffer_finish(buffer);
5734 vk_free2(&device->vk.alloc, pAllocator, buffer);
5737 VKAPI_ATTR VkResult VKAPI_CALL
5738 radv_CreateBuffer(VkDevice _device, const VkBufferCreateInfo *pCreateInfo,
5739 const VkAllocationCallbacks *pAllocator, VkBuffer *pBuffer)
5741 RADV_FROM_HANDLE(radv_device, device, _device);
5742 struct radv_buffer *buffer;
5744 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
5746 buffer = vk_alloc2(&device->vk.alloc, pAllocator, sizeof(*buffer), 8,
5747 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
5749 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
5751 vk_buffer_init(&device->vk, &buffer->vk, pCreateInfo);
5755 if (pCreateInfo->flags & VK_BUFFER_CREATE_SPARSE_BINDING_BIT) {
5756 enum radeon_bo_flag flags = RADEON_FLAG_VIRTUAL;
5757 if (pCreateInfo->flags & VK_BUFFER_CREATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT)
5758 flags |= RADEON_FLAG_REPLAYABLE;
5760 uint64_t replay_address = 0;
5761 const VkBufferOpaqueCaptureAddressCreateInfo *replay_info =
5762 vk_find_struct_const(pCreateInfo->pNext, BUFFER_OPAQUE_CAPTURE_ADDRESS_CREATE_INFO);
5763 if (replay_info && replay_info->opaqueCaptureAddress)
5764 replay_address = replay_info->opaqueCaptureAddress;
5767 device->ws->buffer_create(device->ws, align64(buffer->vk.size, 4096), 4096, 0, flags,
5768 RADV_BO_PRIORITY_VIRTUAL, replay_address, &buffer->bo);
5769 if (result != VK_SUCCESS) {
5770 radv_destroy_buffer(device, pAllocator, buffer);
5771 return vk_error(device, result);
5775 *pBuffer = radv_buffer_to_handle(buffer);
5780 VKAPI_ATTR void VKAPI_CALL
5781 radv_DestroyBuffer(VkDevice _device, VkBuffer _buffer, const VkAllocationCallbacks *pAllocator)
5783 RADV_FROM_HANDLE(radv_device, device, _device);
5784 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
5789 radv_destroy_buffer(device, pAllocator, buffer);
5792 VKAPI_ATTR VkDeviceAddress VKAPI_CALL
5793 radv_GetBufferDeviceAddress(VkDevice device, const VkBufferDeviceAddressInfo *pInfo)
5795 RADV_FROM_HANDLE(radv_buffer, buffer, pInfo->buffer);
5796 return radv_buffer_get_va(buffer->bo) + buffer->offset;
5799 VKAPI_ATTR uint64_t VKAPI_CALL
5800 radv_GetBufferOpaqueCaptureAddress(VkDevice device, const VkBufferDeviceAddressInfo *pInfo)
5802 RADV_FROM_HANDLE(radv_buffer, buffer, pInfo->buffer);
5803 return buffer->bo ? radv_buffer_get_va(buffer->bo) + buffer->offset : 0;
5806 VKAPI_ATTR uint64_t VKAPI_CALL
5807 radv_GetDeviceMemoryOpaqueCaptureAddress(VkDevice device,
5808 const VkDeviceMemoryOpaqueCaptureAddressInfo *pInfo)
5810 RADV_FROM_HANDLE(radv_device_memory, mem, pInfo->memory);
5811 return radv_buffer_get_va(mem->bo);
5814 static inline unsigned
5815 si_tile_mode_index(const struct radv_image_plane *plane, unsigned level, bool stencil)
5818 return plane->surface.u.legacy.zs.stencil_tiling_index[level];
5820 return plane->surface.u.legacy.tiling_index[level];
5824 radv_surface_max_layer_count(struct radv_image_view *iview)
5826 return iview->vk.view_type == VK_IMAGE_VIEW_TYPE_3D ? iview->extent.depth
5827 : (iview->vk.base_array_layer + iview->vk.layer_count);
5831 get_dcc_max_uncompressed_block_size(const struct radv_device *device,
5832 const struct radv_image_view *iview)
5834 if (device->physical_device->rad_info.gfx_level < GFX10 && iview->image->info.samples > 1) {
5835 if (iview->image->planes[0].surface.bpe == 1)
5836 return V_028C78_MAX_BLOCK_SIZE_64B;
5837 else if (iview->image->planes[0].surface.bpe == 2)
5838 return V_028C78_MAX_BLOCK_SIZE_128B;
5841 return V_028C78_MAX_BLOCK_SIZE_256B;
5845 get_dcc_min_compressed_block_size(const struct radv_device *device)
5847 if (!device->physical_device->rad_info.has_dedicated_vram) {
5848 /* amdvlk: [min-compressed-block-size] should be set to 32 for
5849 * dGPU and 64 for APU because all of our APUs to date use
5850 * DIMMs which have a request granularity size of 64B while all
5851 * other chips have a 32B request size.
5853 return V_028C78_MIN_BLOCK_SIZE_64B;
5856 return V_028C78_MIN_BLOCK_SIZE_32B;
5860 radv_init_dcc_control_reg(struct radv_device *device, struct radv_image_view *iview)
5862 unsigned max_uncompressed_block_size = get_dcc_max_uncompressed_block_size(device, iview);
5863 unsigned min_compressed_block_size = get_dcc_min_compressed_block_size(device);
5864 unsigned max_compressed_block_size;
5865 unsigned independent_128b_blocks;
5866 unsigned independent_64b_blocks;
5868 if (!radv_dcc_enabled(iview->image, iview->vk.base_mip_level))
5871 /* For GFX9+ ac_surface computes values for us (except min_compressed
5872 * and max_uncompressed) */
5873 if (device->physical_device->rad_info.gfx_level >= GFX9) {
5874 max_compressed_block_size =
5875 iview->image->planes[0].surface.u.gfx9.color.dcc.max_compressed_block_size;
5876 independent_128b_blocks = iview->image->planes[0].surface.u.gfx9.color.dcc.independent_128B_blocks;
5877 independent_64b_blocks = iview->image->planes[0].surface.u.gfx9.color.dcc.independent_64B_blocks;
5879 independent_128b_blocks = 0;
5881 if (iview->image->vk.usage & (VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT |
5882 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT)) {
5883 /* If this DCC image is potentially going to be used in texture
5884 * fetches, we need some special settings.
5886 independent_64b_blocks = 1;
5887 max_compressed_block_size = V_028C78_MAX_BLOCK_SIZE_64B;
5889 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
5890 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
5891 * big as possible for better compression state.
5893 independent_64b_blocks = 0;
5894 max_compressed_block_size = max_uncompressed_block_size;
5898 uint32_t result = S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size) |
5899 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size) |
5900 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size) |
5901 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks);
5903 if (device->physical_device->rad_info.gfx_level >= GFX11) {
5904 result |= S_028C78_INDEPENDENT_128B_BLOCKS_GFX11(independent_128b_blocks) |
5905 S_028C78_DISABLE_CONSTANT_ENCODE_REG(1) |
5906 S_028C78_FDCC_ENABLE(radv_dcc_enabled(iview->image, iview->vk.base_mip_level));
5908 result |= S_028C78_INDEPENDENT_128B_BLOCKS_GFX10(independent_128b_blocks);
5915 radv_initialise_color_surface(struct radv_device *device, struct radv_color_buffer_info *cb,
5916 struct radv_image_view *iview)
5918 const struct util_format_description *desc;
5919 unsigned ntype, format, swap, endian;
5920 unsigned blend_clamp = 0, blend_bypass = 0;
5922 const struct radv_image_plane *plane = &iview->image->planes[iview->plane_id];
5923 const struct radeon_surf *surf = &plane->surface;
5925 desc = vk_format_description(iview->vk.format);
5927 memset(cb, 0, sizeof(*cb));
5929 /* Intensity is implemented as Red, so treat it that way. */
5930 if (device->physical_device->rad_info.gfx_level >= GFX11)
5931 cb->cb_color_attrib = S_028C74_FORCE_DST_ALPHA_1_GFX11(desc->swizzle[3] == PIPE_SWIZZLE_1);
5933 cb->cb_color_attrib = S_028C74_FORCE_DST_ALPHA_1_GFX6(desc->swizzle[3] == PIPE_SWIZZLE_1);
5935 va = radv_buffer_get_va(iview->image->bo) + iview->image->offset;
5937 cb->cb_color_base = va >> 8;
5939 if (device->physical_device->rad_info.gfx_level >= GFX9) {
5940 if (device->physical_device->rad_info.gfx_level >= GFX11) {
5941 cb->cb_color_attrib3 |= S_028EE0_COLOR_SW_MODE(surf->u.gfx9.swizzle_mode) |
5942 S_028EE0_DCC_PIPE_ALIGNED(surf->u.gfx9.color.dcc.pipe_aligned);
5943 } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
5944 cb->cb_color_attrib3 |= S_028EE0_COLOR_SW_MODE(surf->u.gfx9.swizzle_mode) |
5945 S_028EE0_FMASK_SW_MODE(surf->u.gfx9.color.fmask_swizzle_mode) |
5946 S_028EE0_CMASK_PIPE_ALIGNED(1) |
5947 S_028EE0_DCC_PIPE_ALIGNED(surf->u.gfx9.color.dcc.pipe_aligned);
5949 struct gfx9_surf_meta_flags meta = {
5954 if (surf->meta_offset)
5955 meta = surf->u.gfx9.color.dcc;
5957 cb->cb_color_attrib |= S_028C74_COLOR_SW_MODE(surf->u.gfx9.swizzle_mode) |
5958 S_028C74_FMASK_SW_MODE(surf->u.gfx9.color.fmask_swizzle_mode) |
5959 S_028C74_RB_ALIGNED(meta.rb_aligned) |
5960 S_028C74_PIPE_ALIGNED(meta.pipe_aligned);
5961 cb->cb_mrt_epitch = S_0287A0_EPITCH(surf->u.gfx9.epitch);
5964 cb->cb_color_base += surf->u.gfx9.surf_offset >> 8;
5965 cb->cb_color_base |= surf->tile_swizzle;
5967 const struct legacy_surf_level *level_info = &surf->u.legacy.level[iview->vk.base_mip_level];
5968 unsigned pitch_tile_max, slice_tile_max, tile_mode_index;
5970 cb->cb_color_base += level_info->offset_256B;
5971 if (level_info->mode == RADEON_SURF_MODE_2D)
5972 cb->cb_color_base |= surf->tile_swizzle;
5974 pitch_tile_max = level_info->nblk_x / 8 - 1;
5975 slice_tile_max = (level_info->nblk_x * level_info->nblk_y) / 64 - 1;
5976 tile_mode_index = si_tile_mode_index(plane, iview->vk.base_mip_level, false);
5978 cb->cb_color_pitch = S_028C64_TILE_MAX(pitch_tile_max);
5979 cb->cb_color_slice = S_028C68_TILE_MAX(slice_tile_max);
5980 cb->cb_color_cmask_slice = surf->u.legacy.color.cmask_slice_tile_max;
5982 cb->cb_color_attrib |= S_028C74_TILE_MODE_INDEX(tile_mode_index);
5984 if (radv_image_has_fmask(iview->image)) {
5985 if (device->physical_device->rad_info.gfx_level >= GFX7)
5986 cb->cb_color_pitch |=
5987 S_028C64_FMASK_TILE_MAX(surf->u.legacy.color.fmask.pitch_in_pixels / 8 - 1);
5988 cb->cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(surf->u.legacy.color.fmask.tiling_index);
5989 cb->cb_color_fmask_slice = S_028C88_TILE_MAX(surf->u.legacy.color.fmask.slice_tile_max);
5991 /* This must be set for fast clear to work without FMASK. */
5992 if (device->physical_device->rad_info.gfx_level >= GFX7)
5993 cb->cb_color_pitch |= S_028C64_FMASK_TILE_MAX(pitch_tile_max);
5994 cb->cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index);
5995 cb->cb_color_fmask_slice = S_028C88_TILE_MAX(slice_tile_max);
5999 /* CMASK variables */
6000 va = radv_buffer_get_va(iview->image->bo) + iview->image->offset;
6001 va += surf->cmask_offset;
6002 cb->cb_color_cmask = va >> 8;
6004 va = radv_buffer_get_va(iview->image->bo) + iview->image->offset;
6005 va += surf->meta_offset;
6007 if (radv_dcc_enabled(iview->image, iview->vk.base_mip_level) &&
6008 device->physical_device->rad_info.gfx_level <= GFX8)
6009 va += plane->surface.u.legacy.color.dcc_level[iview->vk.base_mip_level].dcc_offset;
6011 unsigned dcc_tile_swizzle = surf->tile_swizzle;
6012 dcc_tile_swizzle &= ((1 << surf->meta_alignment_log2) - 1) >> 8;
6014 cb->cb_dcc_base = va >> 8;
6015 cb->cb_dcc_base |= dcc_tile_swizzle;
6017 /* GFX10 field has the same base shift as the GFX6 field. */
6018 uint32_t max_slice = radv_surface_max_layer_count(iview) - 1;
6020 S_028C6C_SLICE_START(iview->vk.base_array_layer) | S_028C6C_SLICE_MAX_GFX10(max_slice);
6022 if (iview->image->info.samples > 1) {
6023 unsigned log_samples = util_logbase2(iview->image->info.samples);
6025 if (device->physical_device->rad_info.gfx_level >= GFX11)
6026 cb->cb_color_attrib |= S_028C74_NUM_FRAGMENTS_GFX11(log_samples);
6028 cb->cb_color_attrib |=
6029 S_028C74_NUM_SAMPLES(log_samples) | S_028C74_NUM_FRAGMENTS_GFX6(log_samples);
6032 if (radv_image_has_fmask(iview->image)) {
6033 va = radv_buffer_get_va(iview->image->bo) + iview->image->offset + surf->fmask_offset;
6034 cb->cb_color_fmask = va >> 8;
6035 cb->cb_color_fmask |= surf->fmask_tile_swizzle;
6037 cb->cb_color_fmask = cb->cb_color_base;
6040 ntype = radv_translate_color_numformat(iview->vk.format, desc,
6041 vk_format_get_first_non_void_channel(iview->vk.format));
6042 format = radv_translate_colorformat(iview->vk.format);
6043 assert(format != V_028C70_COLOR_INVALID);
6045 swap = radv_translate_colorswap(iview->vk.format, false);
6046 endian = radv_colorformat_endian_swap(format);
6048 /* blend clamp should be set for all NORM/SRGB types */
6049 if (ntype == V_028C70_NUMBER_UNORM || ntype == V_028C70_NUMBER_SNORM ||
6050 ntype == V_028C70_NUMBER_SRGB)
6053 /* set blend bypass according to docs if SINT/UINT or
6054 8/24 COLOR variants */
6055 if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT ||
6056 format == V_028C70_COLOR_8_24 || format == V_028C70_COLOR_24_8 ||
6057 format == V_028C70_COLOR_X24_8_32_FLOAT) {
6062 if ((ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT) &&
6063 (format == V_028C70_COLOR_8 ||
6064 format == V_028C70_COLOR_8_8 ||
6065 format == V_028C70_COLOR_8_8_8_8))
6066 ->color_is_int8 = true;
6069 S_028C70_COMP_SWAP(swap) | S_028C70_BLEND_CLAMP(blend_clamp) |
6070 S_028C70_BLEND_BYPASS(blend_bypass) | S_028C70_SIMPLE_FLOAT(1) |
6071 S_028C70_ROUND_MODE(ntype != V_028C70_NUMBER_UNORM && ntype != V_028C70_NUMBER_SNORM &&
6072 ntype != V_028C70_NUMBER_SRGB && format != V_028C70_COLOR_8_24 &&
6073 format != V_028C70_COLOR_24_8) |
6074 S_028C70_NUMBER_TYPE(ntype);
6076 if (device->physical_device->rad_info.gfx_level >= GFX11)
6077 cb->cb_color_info |= S_028C70_FORMAT_GFX11(format);
6079 cb->cb_color_info |= S_028C70_FORMAT_GFX6(format) | S_028C70_ENDIAN(endian);
6081 if (radv_image_has_fmask(iview->image)) {
6082 cb->cb_color_info |= S_028C70_COMPRESSION(1);
6083 if (device->physical_device->rad_info.gfx_level == GFX6) {
6084 unsigned fmask_bankh = util_logbase2(surf->u.legacy.color.fmask.bankh);
6085 cb->cb_color_attrib |= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh);
6088 if (radv_image_is_tc_compat_cmask(iview->image)) {
6089 /* Allow the texture block to read FMASK directly
6090 * without decompressing it. This bit must be cleared
6091 * when performing FMASK_DECOMPRESS or DCC_COMPRESS,
6092 * otherwise the operation doesn't happen.
6094 cb->cb_color_info |= S_028C70_FMASK_COMPRESS_1FRAG_ONLY(1);
6096 if (device->physical_device->rad_info.gfx_level == GFX8) {
6097 /* Set CMASK into a tiling format that allows
6098 * the texture block to read it.
6100 cb->cb_color_info |= S_028C70_CMASK_ADDR_TYPE(2);
6105 if (radv_image_has_cmask(iview->image) &&
6106 !(device->instance->debug_flags & RADV_DEBUG_NO_FAST_CLEARS))
6107 cb->cb_color_info |= S_028C70_FAST_CLEAR(1);
6109 if (radv_dcc_enabled(iview->image, iview->vk.base_mip_level) && !iview->disable_dcc_mrt &&
6110 device->physical_device->rad_info.gfx_level < GFX11)
6111 cb->cb_color_info |= S_028C70_DCC_ENABLE(1);
6113 cb->cb_dcc_control = radv_init_dcc_control_reg(device, iview);
6115 /* This must be set for fast clear to work without FMASK. */
6116 if (!radv_image_has_fmask(iview->image) && device->physical_device->rad_info.gfx_level == GFX6) {
6117 unsigned bankh = util_logbase2(surf->u.legacy.bankh);
6118 cb->cb_color_attrib |= S_028C74_FMASK_BANK_HEIGHT(bankh);
6121 if (device->physical_device->rad_info.gfx_level >= GFX9) {
6122 unsigned mip0_depth = iview->image->vk.image_type == VK_IMAGE_TYPE_3D
6123 ? (iview->extent.depth - 1)
6124 : (iview->image->info.array_size - 1);
6126 vk_format_get_plane_width(iview->image->vk.format, iview->plane_id, iview->extent.width);
6128 vk_format_get_plane_height(iview->image->vk.format, iview->plane_id, iview->extent.height);
6130 if (device->physical_device->rad_info.gfx_level >= GFX10) {
6131 cb->cb_color_view |= S_028C6C_MIP_LEVEL_GFX10(iview->vk.base_mip_level);
6133 cb->cb_color_attrib3 |=
6134 S_028EE0_MIP0_DEPTH(mip0_depth) | S_028EE0_RESOURCE_TYPE(surf->u.gfx9.resource_type) |
6135 S_028EE0_RESOURCE_LEVEL(device->physical_device->rad_info.gfx_level >= GFX11 ? 0 : 1);
6137 cb->cb_color_view |= S_028C6C_MIP_LEVEL_GFX9(iview->vk.base_mip_level);
6138 cb->cb_color_attrib |=
6139 S_028C74_MIP0_DEPTH(mip0_depth) | S_028C74_RESOURCE_TYPE(surf->u.gfx9.resource_type);
6142 cb->cb_color_attrib2 = S_028C68_MIP0_WIDTH(width - 1) | S_028C68_MIP0_HEIGHT(height - 1) |
6143 S_028C68_MAX_MIP(iview->image->info.levels - 1);
6148 radv_calc_decompress_on_z_planes(struct radv_device *device, struct radv_image_view *iview)
6150 unsigned max_zplanes = 0;
6152 assert(radv_image_is_tc_compat_htile(iview->image));
6154 if (device->physical_device->rad_info.gfx_level >= GFX9) {
6155 /* Default value for 32-bit depth surfaces. */
6158 if (iview->vk.format == VK_FORMAT_D16_UNORM && iview->image->info.samples > 1)
6161 /* Workaround for a DB hang when ITERATE_256 is set to 1. Only affects 4X MSAA D/S images. */
6162 if (device->physical_device->rad_info.has_two_planes_iterate256_bug &&
6163 radv_image_get_iterate256(device, iview->image) &&
6164 !radv_image_tile_stencil_disabled(device, iview->image) &&
6165 iview->image->info.samples == 4) {
6169 max_zplanes = max_zplanes + 1;
6171 if (iview->vk.format == VK_FORMAT_D16_UNORM) {
6172 /* Do not enable Z plane compression for 16-bit depth
6173 * surfaces because isn't supported on GFX8. Only
6174 * 32-bit depth surfaces are supported by the hardware.
6175 * This allows to maintain shader compatibility and to
6176 * reduce the number of depth decompressions.
6180 if (iview->image->info.samples <= 1)
6182 else if (iview->image->info.samples <= 4)
6193 radv_initialise_vrs_surface(struct radv_image *image, struct radv_buffer *htile_buffer,
6194 struct radv_ds_buffer_info *ds)
6196 const struct radeon_surf *surf = &image->planes[0].surface;
6198 assert(image->vk.format == VK_FORMAT_D16_UNORM);
6199 memset(ds, 0, sizeof(*ds));
6201 ds->pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
6203 ds->db_z_info = S_028038_FORMAT(V_028040_Z_16) |
6204 S_028038_SW_MODE(surf->u.gfx9.swizzle_mode) |
6205 S_028038_ZRANGE_PRECISION(1) |
6206 S_028038_TILE_SURFACE_ENABLE(1);
6207 ds->db_stencil_info = S_02803C_FORMAT(V_028044_STENCIL_INVALID);
6209 ds->db_depth_size = S_02801C_X_MAX(image->info.width - 1) |
6210 S_02801C_Y_MAX(image->info.height - 1);
6212 ds->db_htile_data_base = radv_buffer_get_va(htile_buffer->bo) >> 8;
6213 ds->db_htile_surface = S_028ABC_FULL_CACHE(1) | S_028ABC_PIPE_ALIGNED(1) |
6214 S_028ABC_VRS_HTILE_ENCODING(V_028ABC_VRS_HTILE_4BIT_ENCODING);
6218 radv_initialise_ds_surface(struct radv_device *device, struct radv_ds_buffer_info *ds,
6219 struct radv_image_view *iview)
6221 unsigned level = iview->vk.base_mip_level;
6222 unsigned format, stencil_format;
6223 uint64_t va, s_offs, z_offs;
6224 bool stencil_only = iview->image->vk.format == VK_FORMAT_S8_UINT;
6225 const struct radv_image_plane *plane = &iview->image->planes[0];
6226 const struct radeon_surf *surf = &plane->surface;
6228 assert(vk_format_get_plane_count(iview->image->vk.format) == 1);
6230 memset(ds, 0, sizeof(*ds));
6231 if (!device->instance->absolute_depth_bias) {
6232 switch (iview->image->vk.format) {
6233 case VK_FORMAT_D24_UNORM_S8_UINT:
6234 case VK_FORMAT_X8_D24_UNORM_PACK32:
6235 ds->pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
6237 case VK_FORMAT_D16_UNORM:
6238 case VK_FORMAT_D16_UNORM_S8_UINT:
6239 ds->pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
6241 case VK_FORMAT_D32_SFLOAT:
6242 case VK_FORMAT_D32_SFLOAT_S8_UINT:
6243 ds->pa_su_poly_offset_db_fmt_cntl =
6244 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) | S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
6251 format = radv_translate_dbformat(iview->image->vk.format);
6252 stencil_format = surf->has_stencil ? V_028044_STENCIL_8 : V_028044_STENCIL_INVALID;
6254 uint32_t max_slice = radv_surface_max_layer_count(iview) - 1;
6255 ds->db_depth_view = S_028008_SLICE_START(iview->vk.base_array_layer) |
6256 S_028008_SLICE_MAX(max_slice);
6257 if (device->physical_device->rad_info.gfx_level >= GFX10) {
6258 ds->db_depth_view |= S_028008_SLICE_START_HI(iview->vk.base_array_layer >> 11) |
6259 S_028008_SLICE_MAX_HI(max_slice >> 11);
6262 ds->db_htile_data_base = 0;
6263 ds->db_htile_surface = 0;
6265 va = radv_buffer_get_va(iview->image->bo) + iview->image->offset;
6266 s_offs = z_offs = va;
6268 if (device->physical_device->rad_info.gfx_level >= GFX9) {
6269 assert(surf->u.gfx9.surf_offset == 0);
6270 s_offs += surf->u.gfx9.zs.stencil_offset;
6272 ds->db_z_info = S_028038_FORMAT(format) |
6273 S_028038_NUM_SAMPLES(util_logbase2(iview->image->info.samples)) |
6274 S_028038_SW_MODE(surf->u.gfx9.swizzle_mode) |
6275 S_028038_MAXMIP(iview->image->info.levels - 1) |
6276 S_028038_ZRANGE_PRECISION(1) |
6277 S_028040_ITERATE_256(device->physical_device->rad_info.gfx_level >= GFX11);
6278 ds->db_stencil_info = S_02803C_FORMAT(stencil_format) |
6279 S_02803C_SW_MODE(surf->u.gfx9.zs.stencil_swizzle_mode) |
6280 S_028044_ITERATE_256(device->physical_device->rad_info.gfx_level >= GFX11);
6282 if (device->physical_device->rad_info.gfx_level == GFX9) {
6283 ds->db_z_info2 = S_028068_EPITCH(surf->u.gfx9.epitch);
6284 ds->db_stencil_info2 = S_02806C_EPITCH(surf->u.gfx9.zs.stencil_epitch);
6287 ds->db_depth_view |= S_028008_MIPID(level);
6288 ds->db_depth_size = S_02801C_X_MAX(iview->image->info.width - 1) |
6289 S_02801C_Y_MAX(iview->image->info.height - 1);
6291 if (radv_htile_enabled(iview->image, level)) {
6292 ds->db_z_info |= S_028038_TILE_SURFACE_ENABLE(1);
6294 if (radv_image_is_tc_compat_htile(iview->image)) {
6295 unsigned max_zplanes = radv_calc_decompress_on_z_planes(device, iview);
6297 ds->db_z_info |= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes);
6299 if (device->physical_device->rad_info.gfx_level >= GFX10) {
6300 bool iterate256 = radv_image_get_iterate256(device, iview->image);
6302 ds->db_z_info |= S_028040_ITERATE_FLUSH(1);
6303 ds->db_stencil_info |= S_028044_ITERATE_FLUSH(1);
6304 ds->db_z_info |= S_028040_ITERATE_256(iterate256);
6305 ds->db_stencil_info |= S_028044_ITERATE_256(iterate256);
6307 ds->db_z_info |= S_028038_ITERATE_FLUSH(1);
6308 ds->db_stencil_info |= S_02803C_ITERATE_FLUSH(1);
6312 if (radv_image_tile_stencil_disabled(device, iview->image)) {
6313 ds->db_stencil_info |= S_02803C_TILE_STENCIL_DISABLE(1);
6316 va = radv_buffer_get_va(iview->image->bo) + iview->image->offset + surf->meta_offset;
6317 ds->db_htile_data_base = va >> 8;
6318 ds->db_htile_surface = S_028ABC_FULL_CACHE(1) | S_028ABC_PIPE_ALIGNED(1);
6320 if (device->physical_device->rad_info.gfx_level == GFX9) {
6321 ds->db_htile_surface |= S_028ABC_RB_ALIGNED(1);
6324 if (radv_image_has_vrs_htile(device, iview->image)) {
6325 ds->db_htile_surface |= S_028ABC_VRS_HTILE_ENCODING(V_028ABC_VRS_HTILE_4BIT_ENCODING);
6329 const struct legacy_surf_level *level_info = &surf->u.legacy.level[level];
6332 level_info = &surf->u.legacy.zs.stencil_level[level];
6334 z_offs += (uint64_t)surf->u.legacy.level[level].offset_256B * 256;
6335 s_offs += (uint64_t)surf->u.legacy.zs.stencil_level[level].offset_256B * 256;
6337 ds->db_depth_info = S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview->image));
6338 ds->db_z_info = S_028040_FORMAT(format) | S_028040_ZRANGE_PRECISION(1);
6339 ds->db_stencil_info = S_028044_FORMAT(stencil_format);
6341 if (iview->image->info.samples > 1)
6342 ds->db_z_info |= S_028040_NUM_SAMPLES(util_logbase2(iview->image->info.samples));
6344 if (device->physical_device->rad_info.gfx_level >= GFX7) {
6345 struct radeon_info *info = &device->physical_device->rad_info;
6346 unsigned tiling_index = surf->u.legacy.tiling_index[level];
6347 unsigned stencil_index = surf->u.legacy.zs.stencil_tiling_index[level];
6348 unsigned macro_index = surf->u.legacy.macro_tile_index;
6349 unsigned tile_mode = info->si_tile_mode_array[tiling_index];
6350 unsigned stencil_tile_mode = info->si_tile_mode_array[stencil_index];
6351 unsigned macro_mode = info->cik_macrotile_mode_array[macro_index];
6354 tile_mode = stencil_tile_mode;
6356 ds->db_depth_info |= S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode)) |
6357 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode)) |
6358 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode)) |
6359 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode)) |
6360 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode)) |
6361 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode));
6362 ds->db_z_info |= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode));
6363 ds->db_stencil_info |= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode));
6365 unsigned tile_mode_index = si_tile_mode_index(&iview->image->planes[0], level, false);
6366 ds->db_z_info |= S_028040_TILE_MODE_INDEX(tile_mode_index);
6367 tile_mode_index = si_tile_mode_index(&iview->image->planes[0], level, true);
6368 ds->db_stencil_info |= S_028044_TILE_MODE_INDEX(tile_mode_index);
6370 ds->db_z_info |= S_028040_TILE_MODE_INDEX(tile_mode_index);
6373 ds->db_depth_size = S_028058_PITCH_TILE_MAX((level_info->nblk_x / 8) - 1) |
6374 S_028058_HEIGHT_TILE_MAX((level_info->nblk_y / 8) - 1);
6375 ds->db_depth_slice =
6376 S_02805C_SLICE_TILE_MAX((level_info->nblk_x * level_info->nblk_y) / 64 - 1);
6378 if (radv_htile_enabled(iview->image, level)) {
6379 ds->db_z_info |= S_028040_TILE_SURFACE_ENABLE(1);
6381 if (radv_image_tile_stencil_disabled(device, iview->image)) {
6382 ds->db_stencil_info |= S_028044_TILE_STENCIL_DISABLE(1);
6385 va = radv_buffer_get_va(iview->image->bo) + iview->image->offset + surf->meta_offset;
6386 ds->db_htile_data_base = va >> 8;
6387 ds->db_htile_surface = S_028ABC_FULL_CACHE(1);
6389 if (radv_image_is_tc_compat_htile(iview->image)) {
6390 unsigned max_zplanes = radv_calc_decompress_on_z_planes(device, iview);
6392 ds->db_htile_surface |= S_028ABC_TC_COMPATIBLE(1);
6393 ds->db_z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes);
6398 ds->db_z_read_base = ds->db_z_write_base = z_offs >> 8;
6399 ds->db_stencil_read_base = ds->db_stencil_write_base = s_offs >> 8;
6403 radv_tex_wrap(VkSamplerAddressMode address_mode)
6405 switch (address_mode) {
6406 case VK_SAMPLER_ADDRESS_MODE_REPEAT:
6407 return V_008F30_SQ_TEX_WRAP;
6408 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT:
6409 return V_008F30_SQ_TEX_MIRROR;
6410 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE:
6411 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL;
6412 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER:
6413 return V_008F30_SQ_TEX_CLAMP_BORDER;
6414 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE:
6415 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL;
6417 unreachable("illegal tex wrap mode");
6423 radv_tex_compare(VkCompareOp op)
6426 case VK_COMPARE_OP_NEVER:
6427 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
6428 case VK_COMPARE_OP_LESS:
6429 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS;
6430 case VK_COMPARE_OP_EQUAL:
6431 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL;
6432 case VK_COMPARE_OP_LESS_OR_EQUAL:
6433 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL;
6434 case VK_COMPARE_OP_GREATER:
6435 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER;
6436 case VK_COMPARE_OP_NOT_EQUAL:
6437 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL;
6438 case VK_COMPARE_OP_GREATER_OR_EQUAL:
6439 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL;
6440 case VK_COMPARE_OP_ALWAYS:
6441 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS;
6443 unreachable("illegal compare mode");
6449 radv_tex_filter(VkFilter filter, unsigned max_ansio)
6452 case VK_FILTER_NEAREST:
6453 return (max_ansio > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
6454 : V_008F38_SQ_TEX_XY_FILTER_POINT);
6455 case VK_FILTER_LINEAR:
6456 return (max_ansio > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
6457 : V_008F38_SQ_TEX_XY_FILTER_BILINEAR);
6458 case VK_FILTER_CUBIC_IMG:
6460 fprintf(stderr, "illegal texture filter");
6466 radv_tex_mipfilter(VkSamplerMipmapMode mode)
6469 case VK_SAMPLER_MIPMAP_MODE_NEAREST:
6470 return V_008F38_SQ_TEX_Z_FILTER_POINT;
6471 case VK_SAMPLER_MIPMAP_MODE_LINEAR:
6472 return V_008F38_SQ_TEX_Z_FILTER_LINEAR;
6474 return V_008F38_SQ_TEX_Z_FILTER_NONE;
6479 radv_tex_bordercolor(VkBorderColor bcolor)
6482 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK:
6483 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK:
6484 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK;
6485 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK:
6486 case VK_BORDER_COLOR_INT_OPAQUE_BLACK:
6487 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK;
6488 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE:
6489 case VK_BORDER_COLOR_INT_OPAQUE_WHITE:
6490 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE;
6491 case VK_BORDER_COLOR_FLOAT_CUSTOM_EXT:
6492 case VK_BORDER_COLOR_INT_CUSTOM_EXT:
6493 return V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER;
6501 radv_tex_aniso_filter(unsigned filter)
6503 return MIN2(util_logbase2(filter), 4);
6507 radv_tex_filter_mode(VkSamplerReductionMode mode)
6510 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT:
6511 return V_008F30_SQ_IMG_FILTER_MODE_BLEND;
6512 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT:
6513 return V_008F30_SQ_IMG_FILTER_MODE_MIN;
6514 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT:
6515 return V_008F30_SQ_IMG_FILTER_MODE_MAX;
6523 radv_get_max_anisotropy(struct radv_device *device, const VkSamplerCreateInfo *pCreateInfo)
6525 if (device->force_aniso >= 0)
6526 return device->force_aniso;
6528 if (pCreateInfo->anisotropyEnable && pCreateInfo->maxAnisotropy > 1.0f)
6529 return (uint32_t)pCreateInfo->maxAnisotropy;
6535 radv_register_border_color(struct radv_device *device, VkClearColorValue value)
6539 mtx_lock(&device->border_color_data.mutex);
6541 for (slot = 0; slot < RADV_BORDER_COLOR_COUNT; slot++) {
6542 if (!device->border_color_data.used[slot]) {
6543 /* Copy to the GPU wrt endian-ness. */
6544 util_memcpy_cpu_to_le32(&device->border_color_data.colors_gpu_ptr[slot], &value,
6545 sizeof(VkClearColorValue));
6547 device->border_color_data.used[slot] = true;
6552 mtx_unlock(&device->border_color_data.mutex);
6558 radv_unregister_border_color(struct radv_device *device, uint32_t slot)
6560 mtx_lock(&device->border_color_data.mutex);
6562 device->border_color_data.used[slot] = false;
6564 mtx_unlock(&device->border_color_data.mutex);
6568 radv_init_sampler(struct radv_device *device, struct radv_sampler *sampler,
6569 const VkSamplerCreateInfo *pCreateInfo)
6571 uint32_t max_aniso = radv_get_max_anisotropy(device, pCreateInfo);
6572 uint32_t max_aniso_ratio = radv_tex_aniso_filter(max_aniso);
6573 bool compat_mode = device->physical_device->rad_info.gfx_level == GFX8 ||
6574 device->physical_device->rad_info.gfx_level == GFX9;
6575 unsigned filter_mode = V_008F30_SQ_IMG_FILTER_MODE_BLEND;
6576 unsigned depth_compare_func = V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
6578 pCreateInfo->minFilter == VK_FILTER_NEAREST && pCreateInfo->magFilter == VK_FILTER_NEAREST;
6579 bool uses_border_color = pCreateInfo->addressModeU == VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER ||
6580 pCreateInfo->addressModeV == VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER ||
6581 pCreateInfo->addressModeW == VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER;
6582 VkBorderColor border_color =
6583 uses_border_color ? pCreateInfo->borderColor : VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK;
6584 uint32_t border_color_ptr;
6585 bool disable_cube_wrap = pCreateInfo->flags & VK_SAMPLER_CREATE_NON_SEAMLESS_CUBE_MAP_BIT_EXT;
6587 const struct VkSamplerReductionModeCreateInfo *sampler_reduction =
6588 vk_find_struct_const(pCreateInfo->pNext, SAMPLER_REDUCTION_MODE_CREATE_INFO);
6589 if (sampler_reduction)
6590 filter_mode = radv_tex_filter_mode(sampler_reduction->reductionMode);
6592 if (pCreateInfo->compareEnable)
6593 depth_compare_func = radv_tex_compare(pCreateInfo->compareOp);
6595 sampler->border_color_slot = RADV_BORDER_COLOR_COUNT;
6597 if (border_color == VK_BORDER_COLOR_FLOAT_CUSTOM_EXT ||
6598 border_color == VK_BORDER_COLOR_INT_CUSTOM_EXT) {
6599 const VkSamplerCustomBorderColorCreateInfoEXT *custom_border_color =
6600 vk_find_struct_const(pCreateInfo->pNext, SAMPLER_CUSTOM_BORDER_COLOR_CREATE_INFO_EXT);
6602 assert(custom_border_color);
6604 sampler->border_color_slot =
6605 radv_register_border_color(device, custom_border_color->customBorderColor);
6607 /* Did we fail to find a slot? */
6608 if (sampler->border_color_slot == RADV_BORDER_COLOR_COUNT) {
6609 fprintf(stderr, "WARNING: no free border color slots, defaulting to TRANS_BLACK.\n");
6610 border_color = VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK;
6614 /* If we don't have a custom color, set the ptr to 0 */
6616 sampler->border_color_slot != RADV_BORDER_COLOR_COUNT ? sampler->border_color_slot : 0;
6619 (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo->addressModeU)) |
6620 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo->addressModeV)) |
6621 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo->addressModeW)) |
6622 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio) | S_008F30_DEPTH_COMPARE_FUNC(depth_compare_func) |
6623 S_008F30_FORCE_UNNORMALIZED(pCreateInfo->unnormalizedCoordinates ? 1 : 0) |
6624 S_008F30_ANISO_THRESHOLD(max_aniso_ratio >> 1) | S_008F30_ANISO_BIAS(max_aniso_ratio) |
6625 S_008F30_DISABLE_CUBE_WRAP(disable_cube_wrap) | S_008F30_COMPAT_MODE(compat_mode) |
6626 S_008F30_FILTER_MODE(filter_mode) | S_008F30_TRUNC_COORD(trunc_coord));
6627 sampler->state[1] = (S_008F34_MIN_LOD(radv_float_to_ufixed(CLAMP(pCreateInfo->minLod, 0, 15), 8)) |
6628 S_008F34_MAX_LOD(radv_float_to_ufixed(CLAMP(pCreateInfo->maxLod, 0, 15), 8)) |
6629 S_008F34_PERF_MIP(max_aniso_ratio ? max_aniso_ratio + 6 : 0));
6630 sampler->state[2] = (S_008F38_LOD_BIAS(radv_float_to_sfixed(CLAMP(pCreateInfo->mipLodBias, -16, 16), 8)) |
6631 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo->magFilter, max_aniso)) |
6632 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo->minFilter, max_aniso)) |
6633 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo->mipmapMode)));
6634 sampler->state[3] = S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(border_color));
6636 if (device->physical_device->rad_info.gfx_level >= GFX10) {
6637 sampler->state[2] |=
6638 S_008F38_ANISO_OVERRIDE_GFX10(device->instance->disable_aniso_single_level);
6640 sampler->state[2] |=
6641 S_008F38_DISABLE_LSB_CEIL(device->physical_device->rad_info.gfx_level <= GFX8) |
6642 S_008F38_FILTER_PREC_FIX(1) |
6643 S_008F38_ANISO_OVERRIDE_GFX8(device->instance->disable_aniso_single_level &&
6644 device->physical_device->rad_info.gfx_level >= GFX8);
6647 if (device->physical_device->rad_info.gfx_level >= GFX11) {
6648 sampler->state[3] |= S_008F3C_BORDER_COLOR_PTR_GFX11(border_color_ptr);
6650 sampler->state[3] |= S_008F3C_BORDER_COLOR_PTR_GFX6(border_color_ptr);
6654 VKAPI_ATTR VkResult VKAPI_CALL
6655 radv_CreateSampler(VkDevice _device, const VkSamplerCreateInfo *pCreateInfo,
6656 const VkAllocationCallbacks *pAllocator, VkSampler *pSampler)
6658 RADV_FROM_HANDLE(radv_device, device, _device);
6659 struct radv_sampler *sampler;
6661 const struct VkSamplerYcbcrConversionInfo *ycbcr_conversion =
6662 vk_find_struct_const(pCreateInfo->pNext, SAMPLER_YCBCR_CONVERSION_INFO);
6664 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
6666 sampler = vk_alloc2(&device->vk.alloc, pAllocator, sizeof(*sampler), 8,
6667 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
6669 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
6671 vk_object_base_init(&device->vk, &sampler->base, VK_OBJECT_TYPE_SAMPLER);
6673 radv_init_sampler(device, sampler, pCreateInfo);
6675 sampler->ycbcr_sampler =
6676 ycbcr_conversion ? radv_sampler_ycbcr_conversion_from_handle(ycbcr_conversion->conversion)
6678 *pSampler = radv_sampler_to_handle(sampler);
6683 VKAPI_ATTR void VKAPI_CALL
6684 radv_DestroySampler(VkDevice _device, VkSampler _sampler, const VkAllocationCallbacks *pAllocator)
6686 RADV_FROM_HANDLE(radv_device, device, _device);
6687 RADV_FROM_HANDLE(radv_sampler, sampler, _sampler);
6692 if (sampler->border_color_slot != RADV_BORDER_COLOR_COUNT)
6693 radv_unregister_border_color(device, sampler->border_color_slot);
6695 vk_object_base_finish(&sampler->base);
6696 vk_free2(&device->vk.alloc, pAllocator, sampler);
6699 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
6700 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion)
6702 /* For the full details on loader interface versioning, see
6703 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
6704 * What follows is a condensed summary, to help you navigate the large and
6705 * confusing official doc.
6707 * - Loader interface v0 is incompatible with later versions. We don't
6710 * - In loader interface v1:
6711 * - The first ICD entrypoint called by the loader is
6712 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
6714 * - The ICD must statically expose no other Vulkan symbol unless it is
6715 * linked with -Bsymbolic.
6716 * - Each dispatchable Vulkan handle created by the ICD must be
6717 * a pointer to a struct whose first member is VK_LOADER_DATA. The
6718 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
6719 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
6720 * vkDestroySurfaceKHR(). The ICD must be capable of working with
6721 * such loader-managed surfaces.
6723 * - Loader interface v2 differs from v1 in:
6724 * - The first ICD entrypoint called by the loader is
6725 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
6726 * statically expose this entrypoint.
6728 * - Loader interface v3 differs from v2 in:
6729 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
6730 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
6731 * because the loader no longer does so.
6733 * - Loader interface v4 differs from v3 in:
6734 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
6736 * - Loader interface v5 differs from v4 in:
6737 * - The ICD must support Vulkan API version 1.1 and must not return
6738 * VK_ERROR_INCOMPATIBLE_DRIVER from vkCreateInstance() unless a
6739 * Vulkan Loader with interface v4 or smaller is being used and the
6740 * application provides an API version that is greater than 1.0.
6742 *pSupportedVersion = MIN2(*pSupportedVersion, 5u);
6746 VKAPI_ATTR VkResult VKAPI_CALL
6747 radv_GetMemoryFdKHR(VkDevice _device, const VkMemoryGetFdInfoKHR *pGetFdInfo, int *pFD)
6749 RADV_FROM_HANDLE(radv_device, device, _device);
6750 RADV_FROM_HANDLE(radv_device_memory, memory, pGetFdInfo->memory);
6752 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
6754 /* At the moment, we support only the below handle types. */
6755 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
6756 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
6758 bool ret = radv_get_memory_fd(device, memory, pFD);
6760 return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
6765 radv_compute_valid_memory_types_attempt(struct radv_physical_device *dev,
6766 enum radeon_bo_domain domains, enum radeon_bo_flag flags,
6767 enum radeon_bo_flag ignore_flags)
6769 /* Don't count GTT/CPU as relevant:
6771 * - We're not fully consistent between the two.
6772 * - Sometimes VRAM gets VRAM|GTT.
6774 const enum radeon_bo_domain relevant_domains =
6775 RADEON_DOMAIN_VRAM | RADEON_DOMAIN_GDS | RADEON_DOMAIN_OA;
6777 for (unsigned i = 0; i < dev->memory_properties.memoryTypeCount; ++i) {
6778 if ((domains & relevant_domains) != (dev->memory_domains[i] & relevant_domains))
6781 if ((flags & ~ignore_flags) != (dev->memory_flags[i] & ~ignore_flags))
6791 radv_compute_valid_memory_types(struct radv_physical_device *dev, enum radeon_bo_domain domains,
6792 enum radeon_bo_flag flags)
6794 enum radeon_bo_flag ignore_flags = ~(RADEON_FLAG_NO_CPU_ACCESS | RADEON_FLAG_GTT_WC);
6795 uint32_t bits = radv_compute_valid_memory_types_attempt(dev, domains, flags, ignore_flags);
6798 ignore_flags |= RADEON_FLAG_GTT_WC;
6799 bits = radv_compute_valid_memory_types_attempt(dev, domains, flags, ignore_flags);
6803 ignore_flags |= RADEON_FLAG_NO_CPU_ACCESS;
6804 bits = radv_compute_valid_memory_types_attempt(dev, domains, flags, ignore_flags);
6809 VKAPI_ATTR VkResult VKAPI_CALL
6810 radv_GetMemoryFdPropertiesKHR(VkDevice _device, VkExternalMemoryHandleTypeFlagBits handleType,
6811 int fd, VkMemoryFdPropertiesKHR *pMemoryFdProperties)
6813 RADV_FROM_HANDLE(radv_device, device, _device);
6815 switch (handleType) {
6816 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT: {
6817 enum radeon_bo_domain domains;
6818 enum radeon_bo_flag flags;
6819 if (!device->ws->buffer_get_flags_from_fd(device->ws, fd, &domains, &flags))
6820 return vk_error(device, VK_ERROR_INVALID_EXTERNAL_HANDLE);
6822 pMemoryFdProperties->memoryTypeBits =
6823 radv_compute_valid_memory_types(device->physical_device, domains, flags);
6827 /* The valid usage section for this function says:
6829 * "handleType must not be one of the handle types defined as
6832 * So opaque handle types fall into the default "unsupported" case.
6834 return vk_error(device, VK_ERROR_INVALID_EXTERNAL_HANDLE);
6838 VKAPI_ATTR void VKAPI_CALL
6839 radv_GetDeviceGroupPeerMemoryFeatures(VkDevice device, uint32_t heapIndex,
6840 uint32_t localDeviceIndex, uint32_t remoteDeviceIndex,
6841 VkPeerMemoryFeatureFlags *pPeerMemoryFeatures)
6843 assert(localDeviceIndex == remoteDeviceIndex);
6845 *pPeerMemoryFeatures =
6846 VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT | VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
6847 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT | VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
6850 static const VkTimeDomainEXT radv_time_domains[] = {
6851 VK_TIME_DOMAIN_DEVICE_EXT,
6852 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
6853 #ifdef CLOCK_MONOTONIC_RAW
6854 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
6858 VKAPI_ATTR VkResult VKAPI_CALL
6859 radv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(VkPhysicalDevice physicalDevice,
6860 uint32_t *pTimeDomainCount,
6861 VkTimeDomainEXT *pTimeDomains)
6864 VK_OUTARRAY_MAKE_TYPED(VkTimeDomainEXT, out, pTimeDomains, pTimeDomainCount);
6866 for (d = 0; d < ARRAY_SIZE(radv_time_domains); d++) {
6867 vk_outarray_append_typed(VkTimeDomainEXT, &out, i)
6869 *i = radv_time_domains[d];
6873 return vk_outarray_status(&out);
6878 radv_clock_gettime(clockid_t clock_id)
6880 struct timespec current;
6883 ret = clock_gettime(clock_id, ¤t);
6884 #ifdef CLOCK_MONOTONIC_RAW
6885 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
6886 ret = clock_gettime(CLOCK_MONOTONIC, ¤t);
6891 return (uint64_t)current.tv_sec * 1000000000ULL + current.tv_nsec;
6894 VKAPI_ATTR VkResult VKAPI_CALL
6895 radv_GetCalibratedTimestampsEXT(VkDevice _device, uint32_t timestampCount,
6896 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
6897 uint64_t *pTimestamps, uint64_t *pMaxDeviation)
6899 RADV_FROM_HANDLE(radv_device, device, _device);
6900 uint32_t clock_crystal_freq = device->physical_device->rad_info.clock_crystal_freq;
6902 uint64_t begin, end;
6903 uint64_t max_clock_period = 0;
6905 #ifdef CLOCK_MONOTONIC_RAW
6906 begin = radv_clock_gettime(CLOCK_MONOTONIC_RAW);
6908 begin = radv_clock_gettime(CLOCK_MONOTONIC);
6911 for (d = 0; d < timestampCount; d++) {
6912 switch (pTimestampInfos[d].timeDomain) {
6913 case VK_TIME_DOMAIN_DEVICE_EXT:
6914 pTimestamps[d] = device->ws->query_value(device->ws, RADEON_TIMESTAMP);
6915 uint64_t device_period = DIV_ROUND_UP(1000000, clock_crystal_freq);
6916 max_clock_period = MAX2(max_clock_period, device_period);
6918 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
6919 pTimestamps[d] = radv_clock_gettime(CLOCK_MONOTONIC);
6920 max_clock_period = MAX2(max_clock_period, 1);
6923 #ifdef CLOCK_MONOTONIC_RAW
6924 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
6925 pTimestamps[d] = begin;
6934 #ifdef CLOCK_MONOTONIC_RAW
6935 end = radv_clock_gettime(CLOCK_MONOTONIC_RAW);
6937 end = radv_clock_gettime(CLOCK_MONOTONIC);
6941 * The maximum deviation is the sum of the interval over which we
6942 * perform the sampling and the maximum period of any sampled
6943 * clock. That's because the maximum skew between any two sampled
6944 * clock edges is when the sampled clock with the largest period is
6945 * sampled at the end of that period but right at the beginning of the
6946 * sampling interval and some other clock is sampled right at the
6947 * begining of its sampling period and right at the end of the
6948 * sampling interval. Let's assume the GPU has the longest clock
6949 * period and that the application is sampling GPU and monotonic:
6952 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
6953 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
6957 * GPU -----_____-----_____-----_____-----_____
6960 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
6961 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
6963 * Interval <----------------->
6964 * Deviation <-------------------------->
6968 * m = read(monotonic) 2
6971 * We round the sample interval up by one tick to cover sampling error
6972 * in the interval clock
6975 uint64_t sample_interval = end - begin + 1;
6977 *pMaxDeviation = sample_interval + max_clock_period;
6983 VKAPI_ATTR void VKAPI_CALL
6984 radv_GetPhysicalDeviceMultisamplePropertiesEXT(VkPhysicalDevice physicalDevice,
6985 VkSampleCountFlagBits samples,
6986 VkMultisamplePropertiesEXT *pMultisampleProperties)
6988 VkSampleCountFlagBits supported_samples = VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT |
6989 VK_SAMPLE_COUNT_8_BIT;
6991 if (samples & supported_samples) {
6992 pMultisampleProperties->maxSampleLocationGridSize = (VkExtent2D){2, 2};
6994 pMultisampleProperties->maxSampleLocationGridSize = (VkExtent2D){0, 0};
6998 VKAPI_ATTR VkResult VKAPI_CALL
6999 radv_GetPhysicalDeviceFragmentShadingRatesKHR(
7000 VkPhysicalDevice physicalDevice, uint32_t *pFragmentShadingRateCount,
7001 VkPhysicalDeviceFragmentShadingRateKHR *pFragmentShadingRates)
7003 VK_OUTARRAY_MAKE_TYPED(VkPhysicalDeviceFragmentShadingRateKHR, out, pFragmentShadingRates,
7004 pFragmentShadingRateCount);
7006 #define append_rate(w, h, s) \
7008 VkPhysicalDeviceFragmentShadingRateKHR rate = { \
7009 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADING_RATE_PROPERTIES_KHR, \
7010 .sampleCounts = s, \
7011 .fragmentSize = {.width = w, .height = h}, \
7013 vk_outarray_append_typed(VkPhysicalDeviceFragmentShadingRateKHR, &out, r) *r = rate; \
7016 for (uint32_t x = 2; x >= 1; x--) {
7017 for (uint32_t y = 2; y >= 1; y--) {
7018 VkSampleCountFlagBits samples;
7020 if (x == 1 && y == 1) {
7023 samples = VK_SAMPLE_COUNT_1_BIT | VK_SAMPLE_COUNT_2_BIT |
7024 VK_SAMPLE_COUNT_4_BIT | VK_SAMPLE_COUNT_8_BIT;
7027 append_rate(x, y, samples);
7032 return vk_outarray_status(&out);
7036 radv_thread_trace_set_pstate(struct radv_device *device, bool enable)
7038 struct radeon_winsys *ws = device->ws;
7039 enum radeon_ctx_pstate pstate = enable ? RADEON_CTX_PSTATE_PEAK : RADEON_CTX_PSTATE_NONE;
7041 if (device->physical_device->rad_info.has_stable_pstate) {
7042 for (unsigned i = 0; i < RADV_MAX_QUEUE_FAMILIES; i++) {
7043 for (unsigned q = 0; q < device->queue_count[i]; q++) {
7044 struct radv_queue *queue = &device->queues[i][q];
7046 if (ws->ctx_set_pstate(queue->hw_ctx, pstate) < 0)
7056 radv_device_acquire_performance_counters(struct radv_device *device)
7059 simple_mtx_lock(&device->pstate_mtx);
7061 if (device->pstate_cnt == 0) {
7062 result = radv_thread_trace_set_pstate(device, true);
7064 ++device->pstate_cnt;
7067 simple_mtx_unlock(&device->pstate_mtx);
7072 radv_device_release_performance_counters(struct radv_device *device)
7074 simple_mtx_lock(&device->pstate_mtx);
7076 if (--device->pstate_cnt == 0)
7077 radv_thread_trace_set_pstate(device, false);
7079 simple_mtx_unlock(&device->pstate_mtx);
7082 VKAPI_ATTR VkResult VKAPI_CALL
7083 radv_AcquireProfilingLockKHR(VkDevice _device, const VkAcquireProfilingLockInfoKHR *pInfo)
7085 RADV_FROM_HANDLE(radv_device, device, _device);
7086 bool result = radv_device_acquire_performance_counters(device);
7087 return result ? VK_SUCCESS : VK_ERROR_UNKNOWN;
7090 VKAPI_ATTR void VKAPI_CALL
7091 radv_ReleaseProfilingLockKHR(VkDevice _device)
7093 RADV_FROM_HANDLE(radv_device, device, _device);
7094 radv_device_release_performance_counters(device);