xref: /aosp_15_r20/external/mesa3d/src/amd/vulkan/radv_device.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)

/*
 * Copyright © 2016 Red Hat.
 * Copyright © 2016 Bas Nieuwenhuizen
 *
 * based in part on anv driver which is:
 * Copyright © 2015 Intel Corporation
 *
 * SPDX-License-Identifier: MIT
 */

#include <fcntl.h>
#include <stdbool.h>
#include <string.h>

#ifdef __FreeBSD__
#include <sys/types.h>
#endif
#ifdef MAJOR_IN_MKDEV
#include <sys/mkdev.h>
#endif
#ifdef MAJOR_IN_SYSMACROS
#include <sys/sysmacros.h>
#endif

#ifdef __linux__
#include <sys/inotify.h>
#endif

#include "meta/radv_meta.h"
#include "util/disk_cache.h"
#include "util/u_debug.h"
#include "radv_cs.h"
#include "radv_debug.h"
#include "radv_entrypoints.h"
#include "radv_formats.h"
#include "radv_physical_device.h"
#include "radv_printf.h"
#include "radv_rmv.h"
#include "radv_shader.h"
#include "radv_spm.h"
#include "radv_sqtt.h"
#include "vk_common_entrypoints.h"
#include "vk_pipeline_cache.h"
#include "vk_semaphore.h"
#include "vk_util.h"
#ifdef _WIN32
typedef void *drmDevicePtr;
#include <io.h>
#else
#include <amdgpu.h>
#include <xf86drm.h>
#include "drm-uapi/amdgpu_drm.h"
#include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
#endif
#include "util/build_id.h"
#include "util/driconf.h"
#include "util/mesa-sha1.h"
#include "util/os_time.h"
#include "util/timespec.h"
#include "util/u_atomic.h"
#include "util/u_process.h"
#include "vulkan/vk_icd.h"
#include "winsys/null/radv_null_winsys_public.h"
#include "git_sha1.h"
#include "sid.h"
#include "vk_common_entrypoints.h"
#include "vk_format.h"
#include "vk_sync.h"
#include "vk_sync_dummy.h"

#if AMD_LLVM_AVAILABLE
#include "ac_llvm_util.h"
#endif

#include "ac_descriptors.h"
#include "ac_formats.h"

static bool
radv_spm_trace_enabled(const struct radv_instance *instance)
{
   return (instance->vk.trace_mode & RADV_TRACE_MODE_RGP) &&
          debug_get_bool_option("RADV_THREAD_TRACE_CACHE_COUNTERS", true);
}

static bool
radv_trap_handler_enabled()
{
   return !!getenv("RADV_TRAP_HANDLER");
}

VKAPI_ATTR VkResult VKAPI_CALL
radv_GetMemoryHostPointerPropertiesEXT(VkDevice _device, VkExternalMemoryHandleTypeFlagBits handleType,
                                       const void *pHostPointer,
                                       VkMemoryHostPointerPropertiesEXT *pMemoryHostPointerProperties)
{
   VK_FROM_HANDLE(radv_device, device, _device);
   const struct radv_physical_device *pdev = radv_device_physical(device);

   switch (handleType) {
   case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
      uint32_t memoryTypeBits = 0;
      for (int i = 0; i < pdev->memory_properties.memoryTypeCount; i++) {
         if (pdev->memory_domains[i] == RADEON_DOMAIN_GTT && !(pdev->memory_flags[i] & RADEON_FLAG_GTT_WC)) {
            memoryTypeBits = (1 << i);
            break;
         }
      }
      pMemoryHostPointerProperties->memoryTypeBits = memoryTypeBits;
      return VK_SUCCESS;
   }
   default:
      return VK_ERROR_INVALID_EXTERNAL_HANDLE;
   }
}

static VkResult
radv_device_init_border_color(struct radv_device *device)
{
   VkResult result;

   result = radv_bo_create(device, NULL, RADV_BORDER_COLOR_BUFFER_SIZE, 4096, RADEON_DOMAIN_VRAM,
                           RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_READ_ONLY | RADEON_FLAG_NO_INTERPROCESS_SHARING,
                           RADV_BO_PRIORITY_SHADER, 0, true, &device->border_color_data.bo);

   if (result != VK_SUCCESS)
      return vk_error(device, result);

   radv_rmv_log_border_color_palette_create(device, device->border_color_data.bo);

   result = device->ws->buffer_make_resident(device->ws, device->border_color_data.bo, true);
   if (result != VK_SUCCESS)
      return vk_error(device, result);

   device->border_color_data.colors_gpu_ptr = radv_buffer_map(device->ws, device->border_color_data.bo);
   if (!device->border_color_data.colors_gpu_ptr)
      return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
   mtx_init(&device->border_color_data.mutex, mtx_plain);

   return VK_SUCCESS;
}

static void
radv_device_finish_border_color(struct radv_device *device)
{
   if (device->border_color_data.bo) {
      radv_rmv_log_border_color_palette_destroy(device, device->border_color_data.bo);
      device->ws->buffer_make_resident(device->ws, device->border_color_data.bo, false);
      radv_bo_destroy(device, NULL, device->border_color_data.bo);

      mtx_destroy(&device->border_color_data.mutex);
   }
}

static struct radv_shader_part *
_radv_create_vs_prolog(struct radv_device *device, const void *_key)
{
   struct radv_vs_prolog_key *key = (struct radv_vs_prolog_key *)_key;
   return radv_create_vs_prolog(device, key);
}

static uint32_t
radv_hash_vs_prolog(const void *key_)
{
   const struct radv_vs_prolog_key *key = key_;
   return _mesa_hash_data(key, sizeof(*key));
}

static bool
radv_cmp_vs_prolog(const void *a_, const void *b_)
{
   const struct radv_vs_prolog_key *a = a_;
   const struct radv_vs_prolog_key *b = b_;

   return memcmp(a, b, sizeof(*a)) == 0;
}

static struct radv_shader_part_cache_ops vs_prolog_ops = {
   .create = _radv_create_vs_prolog,
   .hash = radv_hash_vs_prolog,
   .equals = radv_cmp_vs_prolog,
};

static VkResult
radv_device_init_vs_prologs(struct radv_device *device)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   const struct radv_instance *instance = radv_physical_device_instance(pdev);

   if (!radv_shader_part_cache_init(&device->vs_prologs, &vs_prolog_ops))
      return vk_error(instance, VK_ERROR_OUT_OF_HOST_MEMORY);

   /* don't pre-compile prologs if we want to print them */
   if (instance->debug_flags & RADV_DEBUG_DUMP_PROLOGS)
      return VK_SUCCESS;

   struct radv_vs_prolog_key key;
   memset(&key, 0, sizeof(key));
   key.as_ls = false;
   key.is_ngg = pdev->use_ngg;
   key.next_stage = MESA_SHADER_VERTEX;
   key.wave32 = pdev->ge_wave_size == 32;

   for (unsigned i = 1; i <= MAX_VERTEX_ATTRIBS; i++) {
      key.instance_rate_inputs = 0;
      key.num_attributes = i;

      device->simple_vs_prologs[i - 1] = radv_create_vs_prolog(device, &key);
      if (!device->simple_vs_prologs[i - 1])
         return vk_error(instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
   }

   unsigned idx = 0;
   for (unsigned num_attributes = 1; num_attributes <= 16; num_attributes++) {
      for (unsigned count = 1; count <= num_attributes; count++) {
         for (unsigned start = 0; start <= (num_attributes - count); start++) {
            key.instance_rate_inputs = u_bit_consecutive(start, count);
            key.num_attributes = num_attributes;

            struct radv_shader_part *prolog = radv_create_vs_prolog(device, &key);
            if (!prolog)
               return vk_error(instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);

            assert(idx == radv_instance_rate_prolog_index(num_attributes, key.instance_rate_inputs));
            device->instance_rate_vs_prologs[idx++] = prolog;
         }
      }
   }
   assert(idx == ARRAY_SIZE(device->instance_rate_vs_prologs));

   return VK_SUCCESS;
}

static void
radv_device_finish_vs_prologs(struct radv_device *device)
{
   if (device->vs_prologs.ops)
      radv_shader_part_cache_finish(device, &device->vs_prologs);

   for (unsigned i = 0; i < ARRAY_SIZE(device->simple_vs_prologs); i++) {
      if (!device->simple_vs_prologs[i])
         continue;

      radv_shader_part_unref(device, device->simple_vs_prologs[i]);
   }

   for (unsigned i = 0; i < ARRAY_SIZE(device->instance_rate_vs_prologs); i++) {
      if (!device->instance_rate_vs_prologs[i])
         continue;

      radv_shader_part_unref(device, device->instance_rate_vs_prologs[i]);
   }
}

static struct radv_shader_part *
_radv_create_ps_epilog(struct radv_device *device, const void *_key)
{
   struct radv_ps_epilog_key *key = (struct radv_ps_epilog_key *)_key;
   return radv_create_ps_epilog(device, key, NULL);
}

static uint32_t
radv_hash_ps_epilog(const void *key_)
{
   const struct radv_ps_epilog_key *key = key_;
   return _mesa_hash_data(key, sizeof(*key));
}

static bool
radv_cmp_ps_epilog(const void *a_, const void *b_)
{
   const struct radv_ps_epilog_key *a = a_;
   const struct radv_ps_epilog_key *b = b_;

   return memcmp(a, b, sizeof(*a)) == 0;
}

static struct radv_shader_part_cache_ops ps_epilog_ops = {
   .create = _radv_create_ps_epilog,
   .hash = radv_hash_ps_epilog,
   .equals = radv_cmp_ps_epilog,
};

VkResult
radv_device_init_vrs_state(struct radv_device *device)
{
   VkDeviceMemory mem;
   VkBuffer buffer;
   VkResult result;
   VkImage image;

   VkImageCreateInfo image_create_info = {
      .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
      .imageType = VK_IMAGE_TYPE_2D,
      .format = VK_FORMAT_D16_UNORM,
      .extent = {MAX_FRAMEBUFFER_WIDTH, MAX_FRAMEBUFFER_HEIGHT, 1},
      .mipLevels = 1,
      .arrayLayers = 1,
      .samples = VK_SAMPLE_COUNT_1_BIT,
      .tiling = VK_IMAGE_TILING_OPTIMAL,
      .usage = VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT,
      .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
      .queueFamilyIndexCount = 0,
      .pQueueFamilyIndices = NULL,
      .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
   };

   result =
      radv_image_create(radv_device_to_handle(device), &(struct radv_image_create_info){.vk_info = &image_create_info},
                        &device->meta_state.alloc, &image, true);
   if (result != VK_SUCCESS)
      return result;

   VkBufferCreateInfo buffer_create_info = {
      .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
      .pNext =
         &(VkBufferUsageFlags2CreateInfoKHR){
            .sType = VK_STRUCTURE_TYPE_BUFFER_USAGE_FLAGS_2_CREATE_INFO_KHR,
            .usage = VK_BUFFER_USAGE_2_STORAGE_BUFFER_BIT_KHR,
         },
      .size = radv_image_from_handle(image)->planes[0].surface.meta_size,
      .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
   };

   result = radv_create_buffer(device, &buffer_create_info, &device->meta_state.alloc, &buffer, true);
   if (result != VK_SUCCESS)
      goto fail_create;

   VkBufferMemoryRequirementsInfo2 info = {
      .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2,
      .buffer = buffer,
   };
   VkMemoryRequirements2 mem_req = {
      .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
   };
   vk_common_GetBufferMemoryRequirements2(radv_device_to_handle(device), &info, &mem_req);

   VkMemoryAllocateInfo alloc_info = {
      .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
      .allocationSize = mem_req.memoryRequirements.size,
   };

   result = radv_alloc_memory(device, &alloc_info, &device->meta_state.alloc, &mem, true);
   if (result != VK_SUCCESS)
      goto fail_alloc;

   VkBindBufferMemoryInfo bind_info = {.sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
                                       .buffer = buffer,
                                       .memory = mem,
                                       .memoryOffset = 0};

   result = radv_BindBufferMemory2(radv_device_to_handle(device), 1, &bind_info);
   if (result != VK_SUCCESS)
      goto fail_bind;

   device->vrs.image = radv_image_from_handle(image);
   device->vrs.buffer = radv_buffer_from_handle(buffer);
   device->vrs.mem = radv_device_memory_from_handle(mem);

   return VK_SUCCESS;

fail_bind:
   radv_FreeMemory(radv_device_to_handle(device), mem, &device->meta_state.alloc);
fail_alloc:
   radv_DestroyBuffer(radv_device_to_handle(device), buffer, &device->meta_state.alloc);
fail_create:
   radv_DestroyImage(radv_device_to_handle(device), image, &device->meta_state.alloc);

   return result;
}

static void
radv_device_finish_vrs_image(struct radv_device *device)
{
   if (!device->vrs.image)
      return;

   radv_FreeMemory(radv_device_to_handle(device), radv_device_memory_to_handle(device->vrs.mem),
                   &device->meta_state.alloc);
   radv_DestroyBuffer(radv_device_to_handle(device), radv_buffer_to_handle(device->vrs.buffer),
                      &device->meta_state.alloc);
   radv_DestroyImage(radv_device_to_handle(device), radv_image_to_handle(device->vrs.image), &device->meta_state.alloc);
}

static enum radv_force_vrs
radv_parse_vrs_rates(const char *str)
{
   if (!strcmp(str, "2x2")) {
      return RADV_FORCE_VRS_2x2;
   } else if (!strcmp(str, "2x1")) {
      return RADV_FORCE_VRS_2x1;
   } else if (!strcmp(str, "1x2")) {
      return RADV_FORCE_VRS_1x2;
   } else if (!strcmp(str, "1x1")) {
      return RADV_FORCE_VRS_1x1;
   }

   fprintf(stderr, "radv: Invalid VRS rates specified (valid values are 2x2, 2x1, 1x2 and 1x1)\n");
   return RADV_FORCE_VRS_1x1;
}

static const char *
radv_get_force_vrs_config_file(void)
{
   return getenv("RADV_FORCE_VRS_CONFIG_FILE");
}

static enum radv_force_vrs
radv_parse_force_vrs_config_file(const char *config_file)
{
   enum radv_force_vrs force_vrs = RADV_FORCE_VRS_1x1;
   char buf[4];
   FILE *f;

   f = fopen(config_file, "r");
   if (!f) {
      fprintf(stderr, "radv: Can't open file: '%s'.\n", config_file);
      return force_vrs;
   }

   if (fread(buf, sizeof(buf), 1, f) == 1) {
      buf[3] = '\0';
      force_vrs = radv_parse_vrs_rates(buf);
   }

   fclose(f);
   return force_vrs;
}

#ifdef __linux__

#define BUF_LEN ((10 * (sizeof(struct inotify_event) + NAME_MAX + 1)))

static int
radv_notifier_thread_run(void *data)
{
   struct radv_device *device = data;
   struct radv_notifier *notifier = &device->notifier;
   char buf[BUF_LEN];

   while (!notifier->quit) {
      const char *file = radv_get_force_vrs_config_file();
      struct timespec tm = {.tv_nsec = 100000000}; /* 1OOms */
      int length, i = 0;

      length = read(notifier->fd, buf, BUF_LEN);
      while (i < length) {
         struct inotify_event *event = (struct inotify_event *)&buf[i];

         i += sizeof(struct inotify_event) + event->len;
         if (event->mask & IN_MODIFY || event->mask & IN_DELETE_SELF) {
            /* Sleep 100ms for editors that use a temporary file and delete the original. */
            thrd_sleep(&tm, NULL);
            device->force_vrs = radv_parse_force_vrs_config_file(file);

            fprintf(stderr, "radv: Updated the per-vertex VRS rate to '%d'.\n", device->force_vrs);

            if (event->mask & IN_DELETE_SELF) {
               inotify_rm_watch(notifier->fd, notifier->watch);
               notifier->watch = inotify_add_watch(notifier->fd, file, IN_MODIFY | IN_DELETE_SELF);
            }
         }
      }

      thrd_sleep(&tm, NULL);
   }

   return 0;
}

#endif

static int
radv_device_init_notifier(struct radv_device *device)
{
#ifndef __linux__
   return true;
#else
   struct radv_notifier *notifier = &device->notifier;
   const char *file = radv_get_force_vrs_config_file();
   int ret;

   notifier->fd = inotify_init1(IN_NONBLOCK);
   if (notifier->fd < 0)
      return false;

   notifier->watch = inotify_add_watch(notifier->fd, file, IN_MODIFY | IN_DELETE_SELF);
   if (notifier->watch < 0)
      goto fail_watch;

   ret = thrd_create(&notifier->thread, radv_notifier_thread_run, device);
   if (ret)
      goto fail_thread;

   return true;

fail_thread:
   inotify_rm_watch(notifier->fd, notifier->watch);
fail_watch:
   close(notifier->fd);

   return false;
#endif
}

static void
radv_device_finish_notifier(struct radv_device *device)
{
#ifdef __linux__
   struct radv_notifier *notifier = &device->notifier;

   if (!notifier->thread)
      return;

   notifier->quit = true;
   thrd_join(notifier->thread, NULL);
   inotify_rm_watch(notifier->fd, notifier->watch);
   close(notifier->fd);
#endif
}

static VkResult
radv_device_init_perf_counter(struct radv_device *device)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   const size_t bo_size = PERF_CTR_BO_PASS_OFFSET + sizeof(uint64_t) * PERF_CTR_MAX_PASSES;
   VkResult result;

   result = radv_bo_create(device, NULL, bo_size, 4096, RADEON_DOMAIN_GTT,
                           RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING, RADV_BO_PRIORITY_UPLOAD_BUFFER,
                           0, true, &device->perf_counter_bo);
   if (result != VK_SUCCESS)
      return result;

   device->perf_counter_lock_cs = calloc(sizeof(struct radeon_winsys_cs *), 2 * PERF_CTR_MAX_PASSES);
   if (!device->perf_counter_lock_cs)
      return VK_ERROR_OUT_OF_HOST_MEMORY;

   if (!pdev->ac_perfcounters.blocks)
      return VK_ERROR_INITIALIZATION_FAILED;

   return VK_SUCCESS;
}

static void
radv_device_finish_perf_counter(struct radv_device *device)
{
   if (device->perf_counter_bo)
      radv_bo_destroy(device, NULL, device->perf_counter_bo);

   if (!device->perf_counter_lock_cs)
      return;

   for (unsigned i = 0; i < 2 * PERF_CTR_MAX_PASSES; ++i) {
      if (device->perf_counter_lock_cs[i])
         device->ws->cs_destroy(device->perf_counter_lock_cs[i]);
   }

   free(device->perf_counter_lock_cs);
}

static VkResult
radv_device_init_memory_cache(struct radv_device *device)
{
   struct vk_pipeline_cache_create_info info = {.weak_ref = true};

   device->mem_cache = vk_pipeline_cache_create(&device->vk, &info, NULL);
   if (!device->mem_cache)
      return VK_ERROR_OUT_OF_HOST_MEMORY;

   return VK_SUCCESS;
}

static void
radv_device_finish_memory_cache(struct radv_device *device)
{
   if (device->mem_cache)
      vk_pipeline_cache_destroy(device->mem_cache, NULL);
}

static VkResult
radv_device_init_rgp(struct radv_device *device)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   const struct radv_instance *instance = radv_physical_device_instance(pdev);

   if (!(instance->vk.trace_mode & RADV_TRACE_MODE_RGP))
      return VK_SUCCESS;

   if (pdev->info.gfx_level < GFX8 || pdev->info.gfx_level > GFX11_5) {
      fprintf(stderr, "GPU hardware not supported: refer to "
                      "the RGP documentation for the list of "
                      "supported GPUs!\n");
      abort();
   }

   if (!radv_sqtt_init(device))
      return VK_ERROR_INITIALIZATION_FAILED;

   fprintf(stderr,
           "radv: Thread trace support is enabled (initial buffer size: %u MiB, "
           "instruction timing: %s, cache counters: %s, queue events: %s).\n",
           device->sqtt.buffer_size / (1024 * 1024), radv_is_instruction_timing_enabled() ? "enabled" : "disabled",
           radv_spm_trace_enabled(instance) ? "enabled" : "disabled",
           radv_sqtt_queue_events_enabled() ? "enabled" : "disabled");

   if (radv_spm_trace_enabled(instance)) {
      if (pdev->info.gfx_level >= GFX10 && pdev->info.gfx_level < GFX11_5) {
         if (!radv_spm_init(device))
            return VK_ERROR_INITIALIZATION_FAILED;
      } else {
         fprintf(stderr, "radv: SPM isn't supported for this GPU (%s)!\n", pdev->name);
      }
   }

   return VK_SUCCESS;
}

static void
radv_device_finish_rgp(struct radv_device *device)
{
   radv_sqtt_finish(device);
   radv_spm_finish(device);
}

static void
radv_device_init_rmv(struct radv_device *device)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   const struct radv_instance *instance = radv_physical_device_instance(pdev);

   if (!(instance->vk.trace_mode & VK_TRACE_MODE_RMV))
      return;

   struct vk_rmv_device_info info;
   memset(&info, 0, sizeof(struct vk_rmv_device_info));
   radv_rmv_fill_device_info(pdev, &info);
   vk_memory_trace_init(&device->vk, &info);
   radv_memory_trace_init(device);
}

static VkResult
radv_device_init_trap_handler(struct radv_device *device)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);

   if (!radv_trap_handler_enabled())
      return VK_SUCCESS;

   /* TODO: Add support for more hardware. */
   assert(pdev->info.gfx_level == GFX8);

   fprintf(stderr, "**********************************************************************\n");
   fprintf(stderr, "* WARNING: RADV_TRAP_HANDLER is experimental and only for debugging! *\n");
   fprintf(stderr, "**********************************************************************\n");

   if (!radv_trap_handler_init(device))
      return VK_ERROR_INITIALIZATION_FAILED;

   return VK_SUCCESS;
}

static VkResult
radv_device_init_device_fault_detection(struct radv_device *device)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   struct radv_instance *instance = radv_physical_device_instance(pdev);

   if (!radv_device_fault_detection_enabled(device))
      return VK_SUCCESS;

   if (!radv_init_trace(device))
      return VK_ERROR_INITIALIZATION_FAILED;

   fprintf(stderr, "*****************************************************************************\n");
   fprintf(stderr, "* WARNING: RADV_DEBUG=hang is costly and should only be used for debugging! *\n");
   fprintf(stderr, "*****************************************************************************\n");

   /* Wait for idle after every draw/dispatch to identify the
    * first bad call.
    */
   instance->debug_flags |= RADV_DEBUG_SYNC_SHADERS;

   radv_dump_enabled_options(device, stderr);

   return VK_SUCCESS;
}

static void
radv_device_finish_device_fault_detection(struct radv_device *device)
{
   radv_finish_trace(device);
   ralloc_free(device->gpu_hang_report);
}

static VkResult
radv_device_init_tools(struct radv_device *device)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   struct radv_instance *instance = radv_physical_device_instance(pdev);
   VkResult result;

   result = radv_device_init_device_fault_detection(device);
   if (result != VK_SUCCESS)
      return result;

   result = radv_device_init_rgp(device);
   if (result != VK_SUCCESS)
      return result;

   radv_device_init_rmv(device);

   result = radv_device_init_trap_handler(device);
   if (result != VK_SUCCESS)
      return result;

   if ((instance->vk.trace_mode & RADV_TRACE_MODE_RRA) && radv_enable_rt(pdev, false)) {
      result = radv_rra_trace_init(device);
      if (result != VK_SUCCESS)
         return result;
   }

   result = radv_printf_data_init(device);
   if (result != VK_SUCCESS)
      return result;

   return VK_SUCCESS;
}

static void
radv_device_finish_tools(struct radv_device *device)
{
   radv_printf_data_finish(device);
   radv_rra_trace_finish(radv_device_to_handle(device), &device->rra_trace);
   radv_trap_handler_finish(device);
   radv_memory_trace_finish(device);
   radv_device_finish_rgp(device);
   radv_device_finish_device_fault_detection(device);
}

struct dispatch_table_builder {
   struct vk_device_dispatch_table *tables[RADV_DISPATCH_TABLE_COUNT];
   bool used[RADV_DISPATCH_TABLE_COUNT];
   bool initialized[RADV_DISPATCH_TABLE_COUNT];
};

static void
add_entrypoints(struct dispatch_table_builder *b, const struct vk_device_entrypoint_table *entrypoints,
                enum radv_dispatch_table table)
{
   for (int32_t i = table - 1; i >= RADV_DEVICE_DISPATCH_TABLE; i--) {
      if (i == RADV_DEVICE_DISPATCH_TABLE || b->used[i]) {
         vk_device_dispatch_table_from_entrypoints(b->tables[i], entrypoints, !b->initialized[i]);
         b->initialized[i] = true;
      }
   }

   if (table < RADV_DISPATCH_TABLE_COUNT)
      b->used[table] = true;
}

static void
init_dispatch_tables(struct radv_device *device, struct radv_physical_device *pdev)
{
   const struct radv_instance *instance = radv_physical_device_instance(pdev);
   struct dispatch_table_builder b = {0};
   b.tables[RADV_DEVICE_DISPATCH_TABLE] = &device->vk.dispatch_table;
   b.tables[RADV_ANNOTATE_DISPATCH_TABLE] = &device->layer_dispatch.annotate;
   b.tables[RADV_APP_DISPATCH_TABLE] = &device->layer_dispatch.app;
   b.tables[RADV_RGP_DISPATCH_TABLE] = &device->layer_dispatch.rgp;
   b.tables[RADV_RRA_DISPATCH_TABLE] = &device->layer_dispatch.rra;
   b.tables[RADV_RMV_DISPATCH_TABLE] = &device->layer_dispatch.rmv;
   b.tables[RADV_CTX_ROLL_DISPATCH_TABLE] = &device->layer_dispatch.ctx_roll;

   bool gather_ctx_rolls = instance->vk.trace_mode & RADV_TRACE_MODE_CTX_ROLLS;
   if (radv_device_fault_detection_enabled(device) || gather_ctx_rolls)
      add_entrypoints(&b, &annotate_device_entrypoints, RADV_ANNOTATE_DISPATCH_TABLE);

   if (!strcmp(instance->drirc.app_layer, "metroexodus")) {
      add_entrypoints(&b, &metro_exodus_device_entrypoints, RADV_APP_DISPATCH_TABLE);
   } else if (!strcmp(instance->drirc.app_layer, "rage2")) {
      add_entrypoints(&b, &rage2_device_entrypoints, RADV_APP_DISPATCH_TABLE);
   } else if (!strcmp(instance->drirc.app_layer, "quanticdream")) {
      add_entrypoints(&b, &quantic_dream_device_entrypoints, RADV_APP_DISPATCH_TABLE);
   }

   if (instance->vk.trace_mode & RADV_TRACE_MODE_RGP)
      add_entrypoints(&b, &sqtt_device_entrypoints, RADV_RGP_DISPATCH_TABLE);

   if ((instance->vk.trace_mode & RADV_TRACE_MODE_RRA) && radv_enable_rt(pdev, false))
      add_entrypoints(&b, &rra_device_entrypoints, RADV_RRA_DISPATCH_TABLE);

#ifndef _WIN32
   if (instance->vk.trace_mode & VK_TRACE_MODE_RMV)
      add_entrypoints(&b, &rmv_device_entrypoints, RADV_RMV_DISPATCH_TABLE);
#endif

   if (gather_ctx_rolls)
      add_entrypoints(&b, &ctx_roll_device_entrypoints, RADV_CTX_ROLL_DISPATCH_TABLE);

   add_entrypoints(&b, &radv_device_entrypoints, RADV_DISPATCH_TABLE_COUNT);
   add_entrypoints(&b, &wsi_device_entrypoints, RADV_DISPATCH_TABLE_COUNT);
   add_entrypoints(&b, &vk_common_device_entrypoints, RADV_DISPATCH_TABLE_COUNT);
}

static VkResult
capture_trace(VkQueue _queue)
{
   VK_FROM_HANDLE(radv_queue, queue, _queue);
   struct radv_device *device = radv_queue_device(queue);
   const struct radv_physical_device *pdev = radv_device_physical(device);
   const struct radv_instance *instance = radv_physical_device_instance(pdev);

   VkResult result = VK_SUCCESS;

   if (instance->vk.trace_mode & RADV_TRACE_MODE_RRA)
      device->rra_trace.triggered = true;

   if (device->vk.memory_trace_data.is_enabled) {
      simple_mtx_lock(&device->vk.memory_trace_data.token_mtx);
      radv_rmv_collect_trace_events(device);
      vk_dump_rmv_capture(&device->vk.memory_trace_data);
      simple_mtx_unlock(&device->vk.memory_trace_data.token_mtx);
   }

   if (instance->vk.trace_mode & RADV_TRACE_MODE_RGP)
      device->sqtt_triggered = true;

   if (instance->vk.trace_mode & RADV_TRACE_MODE_CTX_ROLLS) {
      char filename[2048];
      time_t t = time(NULL);
      struct tm now = *localtime(&t);
      snprintf(filename, sizeof(filename), "/tmp/%s_%04d.%02d.%02d_%02d.%02d.%02d.ctxroll", util_get_process_name(),
               1900 + now.tm_year, now.tm_mon + 1, now.tm_mday, now.tm_hour, now.tm_min, now.tm_sec);

      simple_mtx_lock(&device->ctx_roll_mtx);

      device->ctx_roll_file = fopen(filename, "w");
      if (device->ctx_roll_file)
         fprintf(stderr, "radv: Writing context rolls to '%s'...\n", filename);

      simple_mtx_unlock(&device->ctx_roll_mtx);
   }

   return result;
}

static void
radv_device_init_cache_key(struct radv_device *device)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   struct radv_device_cache_key *key = &device->cache_key;

   key->disable_trunc_coord = device->disable_trunc_coord;
   key->image_2d_view_of_3d = device->vk.enabled_features.image2DViewOf3D && pdev->info.gfx_level == GFX9;
   key->mesh_shader_queries = device->vk.enabled_features.meshShaderQueries;
   key->primitives_generated_query = radv_uses_primitives_generated_query(device);

   /* The Vulkan spec says:
    *  "Binary shaders retrieved from a physical device with a certain shaderBinaryUUID are
    *   guaranteed to be compatible with all other physical devices reporting the same
    *   shaderBinaryUUID and the same or higher shaderBinaryVersion."
    *
    * That means the driver should compile shaders for the "worst" case of all features being
    * enabled, regardless of what features are actually enabled on the logical device.
    */
   if (device->vk.enabled_features.shaderObject) {
      key->image_2d_view_of_3d = pdev->info.gfx_level == GFX9;
      key->primitives_generated_query = true;
   }

   _mesa_blake3_compute(key, sizeof(*key), device->cache_hash);
}

static void
radv_create_gfx_preamble(struct radv_device *device)
{
   struct radeon_cmdbuf *cs = device->ws->cs_create(device->ws, AMD_IP_GFX, false);
   if (!cs)
      return;

   radeon_check_space(device->ws, cs, 512);

   radv_emit_graphics(device, cs);

   device->ws->cs_pad(cs, 0);

   VkResult result = radv_bo_create(
      device, NULL, cs->cdw * 4, 4096, device->ws->cs_domain(device->ws),
      RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_READ_ONLY | RADEON_FLAG_GTT_WC,
      RADV_BO_PRIORITY_CS, 0, true, &device->gfx_init);
   if (result != VK_SUCCESS)
      goto fail;

   void *map = radv_buffer_map(device->ws, device->gfx_init);
   if (!map) {
      radv_bo_destroy(device, NULL, device->gfx_init);
      device->gfx_init = NULL;
      goto fail;
   }
   memcpy(map, cs->buf, cs->cdw * 4);

   device->ws->buffer_unmap(device->ws, device->gfx_init, false);
   device->gfx_init_size_dw = cs->cdw;
fail:
   device->ws->cs_destroy(cs);
}

/* For MSAA sample positions. */
#define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y)                                                              \
   ((((unsigned)(s0x)&0xf) << 0) | (((unsigned)(s0y)&0xf) << 4) | (((unsigned)(s1x)&0xf) << 8) |                       \
    (((unsigned)(s1y)&0xf) << 12) | (((unsigned)(s2x)&0xf) << 16) | (((unsigned)(s2y)&0xf) << 20) |                    \
    (((unsigned)(s3x)&0xf) << 24) | (((unsigned)(s3y)&0xf) << 28))

/* For obtaining location coordinates from registers */
#define SEXT4(x)               ((int)((x) | ((x)&0x8 ? 0xfffffff0 : 0)))
#define GET_SFIELD(reg, index) SEXT4(((reg) >> ((index)*4)) & 0xf)
#define GET_SX(reg, index)     GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2)
#define GET_SY(reg, index)     GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2 + 1)

/* 1x MSAA */
static const uint32_t sample_locs_1x = FILL_SREG(0, 0, 0, 0, 0, 0, 0, 0);
static const unsigned max_dist_1x = 0;
static const uint64_t centroid_priority_1x = 0x0000000000000000ull;

/* 2xMSAA */
static const uint32_t sample_locs_2x = FILL_SREG(4, 4, -4, -4, 0, 0, 0, 0);
static const unsigned max_dist_2x = 4;
static const uint64_t centroid_priority_2x = 0x1010101010101010ull;

/* 4xMSAA */
static const uint32_t sample_locs_4x = FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6);
static const unsigned max_dist_4x = 6;
static const uint64_t centroid_priority_4x = 0x3210321032103210ull;

/* 8xMSAA */
static const uint32_t sample_locs_8x[] = {
   FILL_SREG(1, -3, -1, 3, 5, 1, -3, -5),
   FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
   /* The following are unused by hardware, but we emit them to IBs
    * instead of multiple SET_CONTEXT_REG packets. */
   0,
   0,
};
static const unsigned max_dist_8x = 7;
static const uint64_t centroid_priority_8x = 0x7654321076543210ull;

unsigned
radv_get_default_max_sample_dist(int log_samples)
{
   unsigned max_dist[] = {
      max_dist_1x,
      max_dist_2x,
      max_dist_4x,
      max_dist_8x,
   };
   return max_dist[log_samples];
}

void
radv_emit_default_sample_locations(const struct radv_physical_device *pdev, struct radeon_cmdbuf *cs, int nr_samples)
{
   uint64_t centroid_priority;

   switch (nr_samples) {
   default:
   case 1:
      centroid_priority = centroid_priority_1x;

      radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, sample_locs_1x);
      radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, sample_locs_1x);
      radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, sample_locs_1x);
      radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, sample_locs_1x);
      break;
   case 2:
      centroid_priority = centroid_priority_2x;

      radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, sample_locs_2x);
      radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, sample_locs_2x);
      radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, sample_locs_2x);
      radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, sample_locs_2x);
      break;
   case 4:
      centroid_priority = centroid_priority_4x;

      radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, sample_locs_4x);
      radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, sample_locs_4x);
      radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, sample_locs_4x);
      radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, sample_locs_4x);
      break;
   case 8:
      centroid_priority = centroid_priority_8x;

      radeon_set_context_reg_seq(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 14);
      radeon_emit_array(cs, sample_locs_8x, 4);
      radeon_emit_array(cs, sample_locs_8x, 4);
      radeon_emit_array(cs, sample_locs_8x, 4);
      radeon_emit_array(cs, sample_locs_8x, 2);
      break;
   }

   if (pdev->info.gfx_level >= GFX12) {
      radeon_set_context_reg_seq(cs, R_028BF0_PA_SC_CENTROID_PRIORITY_0, 2);
   } else {
      radeon_set_context_reg_seq(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
   }
   radeon_emit(cs, centroid_priority);
   radeon_emit(cs, centroid_priority >> 32);
}

static void
radv_get_sample_position(struct radv_device *device, unsigned sample_count, unsigned sample_index, float *out_value)
{
   const uint32_t *sample_locs;

   switch (sample_count) {
   case 1:
   default:
      sample_locs = &sample_locs_1x;
      break;
   case 2:
      sample_locs = &sample_locs_2x;
      break;
   case 4:
      sample_locs = &sample_locs_4x;
      break;
   case 8:
      sample_locs = sample_locs_8x;
      break;
   }

   out_value[0] = (GET_SX(sample_locs, sample_index) + 8) / 16.0f;
   out_value[1] = (GET_SY(sample_locs, sample_index) + 8) / 16.0f;
}

static void
radv_device_init_msaa(struct radv_device *device)
{
   int i;

   radv_get_sample_position(device, 1, 0, device->sample_locations_1x[0]);

   for (i = 0; i < 2; i++)
      radv_get_sample_position(device, 2, i, device->sample_locations_2x[i]);
   for (i = 0; i < 4; i++)
      radv_get_sample_position(device, 4, i, device->sample_locations_4x[i]);
   for (i = 0; i < 8; i++)
      radv_get_sample_position(device, 8, i, device->sample_locations_8x[i]);
}

VKAPI_ATTR VkResult VKAPI_CALL
radv_CreateDevice(VkPhysicalDevice physicalDevice, const VkDeviceCreateInfo *pCreateInfo,
                  const VkAllocationCallbacks *pAllocator, VkDevice *pDevice)
{
   VK_FROM_HANDLE(radv_physical_device, pdev, physicalDevice);
   struct radv_instance *instance = radv_physical_device_instance(pdev);
   VkResult result;
   struct radv_device *device;

   bool overallocation_disallowed = false;

   vk_foreach_struct_const (ext, pCreateInfo->pNext) {
      switch (ext->sType) {
      case VK_STRUCTURE_TYPE_DEVICE_MEMORY_OVERALLOCATION_CREATE_INFO_AMD: {
         const VkDeviceMemoryOverallocationCreateInfoAMD *overallocation = (const void *)ext;
         if (overallocation->overallocationBehavior == VK_MEMORY_OVERALLOCATION_BEHAVIOR_DISALLOWED_AMD)
            overallocation_disallowed = true;
         break;
      }
      default:
         break;
      }
   }

   device = vk_zalloc2(&instance->vk.alloc, pAllocator, sizeof(*device), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
   if (!device)
      return vk_error(instance, VK_ERROR_OUT_OF_HOST_MEMORY);

   result = vk_device_init(&device->vk, &pdev->vk, NULL, pCreateInfo, pAllocator);
   if (result != VK_SUCCESS) {
      vk_free(&device->vk.alloc, device);
      return result;
   }

   device->vk.capture_trace = capture_trace;

   device->vk.command_buffer_ops = &radv_cmd_buffer_ops;

   init_dispatch_tables(device, pdev);

   simple_mtx_init(&device->ctx_roll_mtx, mtx_plain);
   simple_mtx_init(&device->trace_mtx, mtx_plain);
   simple_mtx_init(&device->pstate_mtx, mtx_plain);
   simple_mtx_init(&device->rt_handles_mtx, mtx_plain);
   simple_mtx_init(&device->compute_scratch_mtx, mtx_plain);
   simple_mtx_init(&device->pso_cache_stats_mtx, mtx_plain);

   device->rt_handles = _mesa_hash_table_create(NULL, _mesa_hash_u32, _mesa_key_u32_equal);

   device->ws = pdev->ws;
   vk_device_set_drm_fd(&device->vk, device->ws->get_fd(device->ws));

   /* With update after bind we can't attach bo's to the command buffer
    * from the descriptor set anymore, so we have to use a global BO list.
    */
   device->use_global_bo_list =
      (instance->perftest_flags & RADV_PERFTEST_BO_LIST) || device->vk.enabled_features.bufferDeviceAddress ||
      device->vk.enabled_features.descriptorIndexing || device->vk.enabled_extensions.EXT_descriptor_indexing ||
      device->vk.enabled_extensions.EXT_buffer_device_address ||
      device->vk.enabled_extensions.KHR_buffer_device_address ||
      device->vk.enabled_extensions.KHR_ray_tracing_pipeline ||
      device->vk.enabled_extensions.KHR_acceleration_structure ||
      device->vk.enabled_extensions.VALVE_descriptor_set_host_mapping;

   radv_init_shader_arenas(device);

   device->overallocation_disallowed = overallocation_disallowed;
   mtx_init(&device->overallocation_mutex, mtx_plain);

   if (pdev->info.register_shadowing_required || instance->debug_flags & RADV_DEBUG_SHADOW_REGS)
      device->uses_shadow_regs = true;

   /* Create one context per queue priority. */
   for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
      const VkDeviceQueueCreateInfo *queue_create = &pCreateInfo->pQueueCreateInfos[i];
      const VkDeviceQueueGlobalPriorityCreateInfoKHR *global_priority =
         vk_find_struct_const(queue_create->pNext, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_KHR);
      enum radeon_ctx_priority priority = radv_get_queue_global_priority(global_priority);

      if (device->hw_ctx[priority])
         continue;

      result = device->ws->ctx_create(device->ws, priority, &device->hw_ctx[priority]);
      if (result != VK_SUCCESS)
         goto fail_queue;
   }

   for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
      const VkDeviceQueueCreateInfo *queue_create = &pCreateInfo->pQueueCreateInfos[i];
      uint32_t qfi = queue_create->queueFamilyIndex;
      const VkDeviceQueueGlobalPriorityCreateInfoKHR *global_priority =
         vk_find_struct_const(queue_create->pNext, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_KHR);

      device->queues[qfi] = vk_zalloc(&device->vk.alloc, queue_create->queueCount * sizeof(struct radv_queue), 8,
                                      VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
      if (!device->queues[qfi]) {
         result = VK_ERROR_OUT_OF_HOST_MEMORY;
         goto fail_queue;
      }

      device->queue_count[qfi] = queue_create->queueCount;

      for (unsigned q = 0; q < queue_create->queueCount; q++) {
         result = radv_queue_init(device, &device->queues[qfi][q], q, queue_create, global_priority);
         if (result != VK_SUCCESS)
            goto fail_queue;
      }
   }
   device->private_sdma_queue = VK_NULL_HANDLE;

   device->shader_use_invisible_vram = (instance->perftest_flags & RADV_PERFTEST_DMA_SHADERS) &&
                                       /* SDMA buffer copy is only implemented for GFX7+. */
                                       pdev->info.gfx_level >= GFX7;
   result = radv_init_shader_upload_queue(device);
   if (result != VK_SUCCESS)
      goto fail;

   device->pbb_allowed = pdev->info.gfx_level >= GFX9 && !(instance->debug_flags & RADV_DEBUG_NOBINNING);

   device->disable_trunc_coord = instance->drirc.disable_trunc_coord;

   if (instance->vk.app_info.engine_name && !strcmp(instance->vk.app_info.engine_name, "DXVK")) {
      /* For DXVK 2.3.0 and older, use dualSrcBlend to determine if this is D3D9. */
      bool is_d3d9 = !device->vk.enabled_features.dualSrcBlend;
      if (instance->vk.app_info.engine_version > VK_MAKE_VERSION(2, 3, 0))
         is_d3d9 = instance->vk.app_info.app_version & 0x1;

      device->disable_trunc_coord &= !is_d3d9;
   }

   /* The maximum number of scratch waves. Scratch space isn't divided
    * evenly between CUs. The number is only a function of the number of CUs.
    * We can decrease the constant to decrease the scratch buffer size.
    *
    * sctx->scratch_waves must be >= the maximum possible size of
    * 1 threadgroup, so that the hw doesn't hang from being unable
    * to start any.
    *
    * The recommended value is 4 per CU at most. Higher numbers don't
    * bring much benefit, but they still occupy chip resources (think
    * async compute). I've seen ~2% performance difference between 4 and 32.
    */
   uint32_t max_threads_per_block = 2048;
   device->scratch_waves = MAX2(32 * pdev->info.num_cu, max_threads_per_block / 64);

   device->dispatch_initiator = S_00B800_COMPUTE_SHADER_EN(1);

   if (pdev->info.gfx_level >= GFX7) {
      /* If the KMD allows it (there is a KMD hw register for it),
       * allow launching waves out-of-order.
       */
      device->dispatch_initiator |= S_00B800_ORDER_MODE(1);
   }
   if (pdev->info.gfx_level >= GFX10) {
      /* Enable asynchronous compute tunneling. The KMD restricts this feature
       * to high-priority compute queues, so setting the bit on any other queue
       * is a no-op. PAL always sets this bit as well.
       */
      device->dispatch_initiator |= S_00B800_TUNNEL_ENABLE(1);
   }

   /* Disable partial preemption for task shaders.
    * The kernel may not support preemption, but PAL always sets this bit,
    * so let's also set it here for consistency.
    */
   device->dispatch_initiator_task = device->dispatch_initiator | S_00B800_DISABLE_DISP_PREMPT_EN(1);

   if (pdev->info.gfx_level == GFX10_3) {
      if (getenv("RADV_FORCE_VRS_CONFIG_FILE")) {
         const char *file = radv_get_force_vrs_config_file();

         device->force_vrs = radv_parse_force_vrs_config_file(file);

         if (radv_device_init_notifier(device)) {
            device->force_vrs_enabled = true;
         } else {
            fprintf(stderr, "radv: Failed to initialize the notifier for RADV_FORCE_VRS_CONFIG_FILE!\n");
         }
      } else if (getenv("RADV_FORCE_VRS")) {
         const char *vrs_rates = getenv("RADV_FORCE_VRS");

         device->force_vrs = radv_parse_vrs_rates(vrs_rates);
         device->force_vrs_enabled = device->force_vrs != RADV_FORCE_VRS_1x1;
      }
   }

   /* PKT3_LOAD_SH_REG_INDEX is supported on GFX8+, but it hangs with compute queues until GFX10.3. */
   device->load_grid_size_from_user_sgpr = pdev->info.gfx_level >= GFX10_3;

   /* Keep shader info for GPU hangs debugging. */
   device->keep_shader_info = radv_device_fault_detection_enabled(device) || radv_trap_handler_enabled();

   /* Initialize the per-device cache key before compiling meta shaders. */
   radv_device_init_cache_key(device);

   result = radv_device_init_tools(device);
   if (result != VK_SUCCESS)
      goto fail;

   result = radv_device_init_meta(device);
   if (result != VK_SUCCESS)
      goto fail;

   radv_device_init_msaa(device);

   /* If the border color extension is enabled, let's create the buffer we need. */
   if (device->vk.enabled_features.customBorderColors) {
      result = radv_device_init_border_color(device);
      if (result != VK_SUCCESS)
         goto fail;
   }

   if (device->vk.enabled_features.vertexInputDynamicState || device->vk.enabled_features.graphicsPipelineLibrary ||
       device->vk.enabled_features.shaderObject) {
      result = radv_device_init_vs_prologs(device);
      if (result != VK_SUCCESS)
         goto fail;
   }

   if (device->vk.enabled_features.graphicsPipelineLibrary || device->vk.enabled_features.shaderObject ||
       device->vk.enabled_features.extendedDynamicState3ColorBlendEnable ||
       device->vk.enabled_features.extendedDynamicState3ColorWriteMask ||
       device->vk.enabled_features.extendedDynamicState3AlphaToCoverageEnable ||
       device->vk.enabled_features.extendedDynamicState3ColorBlendEquation) {
      if (!radv_shader_part_cache_init(&device->ps_epilogs, &ps_epilog_ops)) {
         result = VK_ERROR_OUT_OF_HOST_MEMORY;
         goto fail;
      }
   }

   if (!(instance->debug_flags & RADV_DEBUG_NO_IBS))
      radv_create_gfx_preamble(device);

   if (!device->vk.disable_internal_cache) {
      result = radv_device_init_memory_cache(device);
      if (result != VK_SUCCESS)
         goto fail_meta;
   }

   device->force_aniso = MIN2(16, (int)debug_get_num_option("RADV_TEX_ANISO", -1));
   if (device->force_aniso >= 0) {
      fprintf(stderr, "radv: Forcing anisotropy filter to %ix\n", 1 << util_logbase2(device->force_aniso));
   }

   if (device->vk.enabled_features.performanceCounterQueryPools) {
      result = radv_device_init_perf_counter(device);
      if (result != VK_SUCCESS)
         goto fail_cache;
   }

   if (device->vk.enabled_features.rayTracingPipelineShaderGroupHandleCaptureReplay) {
      device->capture_replay_arena_vas = _mesa_hash_table_u64_create(NULL);
   }

   if (pdev->info.gfx_level == GFX11 && pdev->info.has_dedicated_vram && instance->drirc.force_pstate_peak_gfx11_dgpu) {
      if (!radv_device_acquire_performance_counters(device))
         fprintf(stderr, "radv: failed to set pstate to profile_peak.\n");
   }

   *pDevice = radv_device_to_handle(device);
   return VK_SUCCESS;

fail_cache:
   radv_device_finish_memory_cache(device);
fail_meta:
   radv_device_finish_meta(device);
fail:
   radv_device_finish_perf_counter(device);

   radv_device_finish_tools(device);

   if (device->gfx_init)
      radv_bo_destroy(device, NULL, device->gfx_init);

   radv_device_finish_notifier(device);
   radv_device_finish_vs_prologs(device);
   if (device->ps_epilogs.ops)
      radv_shader_part_cache_finish(device, &device->ps_epilogs);
   radv_device_finish_border_color(device);

   radv_destroy_shader_upload_queue(device);

fail_queue:
   for (unsigned i = 0; i < RADV_MAX_QUEUE_FAMILIES; i++) {
      for (unsigned q = 0; q < device->queue_count[i]; q++)
         radv_queue_finish(&device->queues[i][q]);
      if (device->queue_count[i])
         vk_free(&device->vk.alloc, device->queues[i]);
   }

   for (unsigned i = 0; i < RADV_NUM_HW_CTX; i++) {
      if (device->hw_ctx[i])
         device->ws->ctx_destroy(device->hw_ctx[i]);
   }

   radv_destroy_shader_arenas(device);

   _mesa_hash_table_destroy(device->rt_handles, NULL);

   simple_mtx_destroy(&device->ctx_roll_mtx);
   simple_mtx_destroy(&device->pstate_mtx);
   simple_mtx_destroy(&device->trace_mtx);
   simple_mtx_destroy(&device->rt_handles_mtx);
   simple_mtx_destroy(&device->compute_scratch_mtx);
   simple_mtx_destroy(&device->pso_cache_stats_mtx);
   mtx_destroy(&device->overallocation_mutex);

   vk_device_finish(&device->vk);
   vk_free(&device->vk.alloc, device);
   return result;
}

VKAPI_ATTR void VKAPI_CALL
radv_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
{
   VK_FROM_HANDLE(radv_device, device, _device);

   if (!device)
      return;

   radv_device_finish_perf_counter(device);

   if (device->gfx_init)
      radv_bo_destroy(device, NULL, device->gfx_init);

   radv_device_finish_notifier(device);
   radv_device_finish_vs_prologs(device);
   if (device->ps_epilogs.ops)
      radv_shader_part_cache_finish(device, &device->ps_epilogs);
   radv_device_finish_border_color(device);
   radv_device_finish_vrs_image(device);

   for (unsigned i = 0; i < RADV_MAX_QUEUE_FAMILIES; i++) {
      for (unsigned q = 0; q < device->queue_count[i]; q++)
         radv_queue_finish(&device->queues[i][q]);
      if (device->queue_count[i])
         vk_free(&device->vk.alloc, device->queues[i]);
   }
   if (device->private_sdma_queue != VK_NULL_HANDLE) {
      radv_queue_finish(device->private_sdma_queue);
      vk_free(&device->vk.alloc, device->private_sdma_queue);
   }

   _mesa_hash_table_destroy(device->rt_handles, NULL);

   radv_device_finish_meta(device);

   radv_device_finish_memory_cache(device);

   radv_destroy_shader_upload_queue(device);

   for (unsigned i = 0; i < RADV_NUM_HW_CTX; i++) {
      if (device->hw_ctx[i])
         device->ws->ctx_destroy(device->hw_ctx[i]);
   }

   mtx_destroy(&device->overallocation_mutex);
   simple_mtx_destroy(&device->ctx_roll_mtx);
   simple_mtx_destroy(&device->pstate_mtx);
   simple_mtx_destroy(&device->trace_mtx);
   simple_mtx_destroy(&device->rt_handles_mtx);
   simple_mtx_destroy(&device->compute_scratch_mtx);
   simple_mtx_destroy(&device->pso_cache_stats_mtx);

   radv_destroy_shader_arenas(device);
   if (device->capture_replay_arena_vas)
      _mesa_hash_table_u64_destroy(device->capture_replay_arena_vas);

   vk_device_finish(&device->vk);
   vk_free(&device->vk.alloc, device);
}

bool
radv_get_memory_fd(struct radv_device *device, struct radv_device_memory *memory, int *pFD)
{
   /* Set BO metadata for dedicated image allocations.  We don't need it for import when the image
    * tiling is VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT, but we set it anyway for foreign consumers.
    */
   if (memory->image) {
      struct radeon_bo_metadata metadata;

      assert(memory->image->bindings[0].offset == 0);
      radv_init_metadata(device, memory->image, &metadata);
      device->ws->buffer_set_metadata(device->ws, memory->bo, &metadata);
   }

   return device->ws->buffer_get_fd(device->ws, memory->bo, pFD);
}

VKAPI_ATTR void VKAPI_CALL
radv_GetImageMemoryRequirements2(VkDevice _device, const VkImageMemoryRequirementsInfo2 *pInfo,
                                 VkMemoryRequirements2 *pMemoryRequirements)
{
   VK_FROM_HANDLE(radv_device, device, _device);
   VK_FROM_HANDLE(radv_image, image, pInfo->image);
   const struct radv_physical_device *pdev = radv_device_physical(device);
   uint32_t alignment;
   uint64_t size;

   const VkImagePlaneMemoryRequirementsInfo *plane_info =
      vk_find_struct_const(pInfo->pNext, IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO);

   if (plane_info) {
      const uint32_t plane = radv_plane_from_aspect(plane_info->planeAspect);

      size = image->planes[plane].surface.total_size;
      alignment = 1 << image->planes[plane].surface.alignment_log2;
   } else {
      size = image->size;
      alignment = image->alignment;
   }

   pMemoryRequirements->memoryRequirements.memoryTypeBits =
      ((1u << pdev->memory_properties.memoryTypeCount) - 1u) & ~pdev->memory_types_32bit;

   pMemoryRequirements->memoryRequirements.size = size;
   pMemoryRequirements->memoryRequirements.alignment = alignment;

   vk_foreach_struct (ext, pMemoryRequirements->pNext) {
      switch (ext->sType) {
      case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
         VkMemoryDedicatedRequirements *req = (VkMemoryDedicatedRequirements *)ext;
         req->requiresDedicatedAllocation = image->shareable && image->vk.tiling != VK_IMAGE_TILING_LINEAR;
         req->prefersDedicatedAllocation = req->requiresDedicatedAllocation;
         break;
      }
      default:
         break;
      }
   }
}

VKAPI_ATTR void VKAPI_CALL
radv_GetDeviceImageMemoryRequirements(VkDevice device, const VkDeviceImageMemoryRequirements *pInfo,
                                      VkMemoryRequirements2 *pMemoryRequirements)
{
   UNUSED VkResult result;
   VkImage image;

   /* Determining the image size/alignment require to create a surface, which is complicated without
    * creating an image.
    * TODO: Avoid creating an image.
    */
   result =
      radv_image_create(device, &(struct radv_image_create_info){.vk_info = pInfo->pCreateInfo}, NULL, &image, true);
   assert(result == VK_SUCCESS);

   VkImageMemoryRequirementsInfo2 info2 = {
      .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
      .image = image,
   };

   radv_GetImageMemoryRequirements2(device, &info2, pMemoryRequirements);

   radv_DestroyImage(device, image, NULL);
}

static uint32_t
radv_surface_max_layer_count(struct radv_image_view *iview)
{
   return iview->vk.view_type == VK_IMAGE_VIEW_TYPE_3D ? iview->extent.depth
                                                       : (iview->vk.base_array_layer + iview->vk.layer_count);
}

unsigned
radv_get_dcc_max_uncompressed_block_size(const struct radv_device *device, const struct radv_image *image)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);

   if (pdev->info.gfx_level < GFX10 && image->vk.samples > 1) {
      if (image->planes[0].surface.bpe == 1)
         return V_028C78_MAX_BLOCK_SIZE_64B;
      else if (image->planes[0].surface.bpe == 2)
         return V_028C78_MAX_BLOCK_SIZE_128B;
   }

   return V_028C78_MAX_BLOCK_SIZE_256B;
}

void
radv_initialise_color_surface(struct radv_device *device, struct radv_color_buffer_info *cb,
                              struct radv_image_view *iview)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   const struct radv_instance *instance = radv_physical_device_instance(pdev);
   uint64_t va;
   const struct radv_image_plane *plane = &iview->image->planes[iview->plane_id];
   const struct radeon_surf *surf = &plane->surface;

   memset(cb, 0, sizeof(*cb));

   const unsigned num_layers =
      iview->image->vk.image_type == VK_IMAGE_TYPE_3D ? (iview->extent.depth - 1) : (iview->image->vk.array_layers - 1);

   const struct ac_cb_state cb_state = {
      .surf = surf,
      .format = vk_format_to_pipe_format(iview->vk.format),
      .width = vk_format_get_plane_width(iview->image->vk.format, iview->plane_id, iview->extent.width),
      .height = vk_format_get_plane_height(iview->image->vk.format, iview->plane_id, iview->extent.height),
      .first_layer = iview->vk.base_array_layer,
      .last_layer = radv_surface_max_layer_count(iview) - 1,
      .num_layers = num_layers,
      .num_samples = iview->image->vk.samples,
      .num_storage_samples = iview->image->vk.samples,
      .base_level = iview->vk.base_mip_level,
      .num_levels = iview->image->vk.mip_levels,
      .gfx10 =
         {
            .nbc_view = iview->nbc_view.valid ? &iview->nbc_view : NULL,
         },
   };

   ac_init_cb_surface(&pdev->info, &cb_state, &cb->ac);

   uint32_t plane_id = iview->image->disjoint ? iview->plane_id : 0;
   va = radv_image_get_va(iview->image, plane_id);

   const struct ac_mutable_cb_state mutable_cb_state = {
      .surf = surf,
      .cb = &cb->ac,
      .va = va,
      .base_level = iview->vk.base_mip_level,
      .num_samples = iview->image->vk.samples,
      .fmask_enabled = radv_image_has_fmask(iview->image),
      .cmask_enabled = radv_image_has_cmask(iview->image),
      .fast_clear_enabled = !(instance->debug_flags & RADV_DEBUG_NO_FAST_CLEARS),
      .tc_compat_cmask_enabled = radv_image_is_tc_compat_cmask(iview->image),
      .dcc_enabled = radv_dcc_enabled(iview->image, iview->vk.base_mip_level) &&
                     (pdev->info.gfx_level >= GFX11 || !iview->disable_dcc_mrt),
      .gfx10 =
         {
            .nbc_view = iview->nbc_view.valid ? &iview->nbc_view : NULL,
         },
   };

   ac_set_mutable_cb_surface_fields(&pdev->info, &mutable_cb_state, &cb->ac);
}

void
radv_initialise_vrs_surface(struct radv_image *image, struct radv_buffer *htile_buffer, struct radv_ds_buffer_info *ds)
{
   const struct radeon_surf *surf = &image->planes[0].surface;

   assert(image->vk.format == VK_FORMAT_D16_UNORM);
   memset(ds, 0, sizeof(*ds));

   ds->ac.db_z_info = S_028038_FORMAT(V_028040_Z_16) | S_028038_SW_MODE(surf->u.gfx9.swizzle_mode) |
                      S_028038_ZRANGE_PRECISION(1) | S_028038_TILE_SURFACE_ENABLE(1);
   ds->ac.db_stencil_info = S_02803C_FORMAT(V_028044_STENCIL_INVALID);

   ds->ac.db_depth_size = S_02801C_X_MAX(image->vk.extent.width - 1) | S_02801C_Y_MAX(image->vk.extent.height - 1);

   ds->ac.u.gfx6.db_htile_data_base = radv_buffer_get_va(htile_buffer->bo) >> 8;
   ds->ac.u.gfx6.db_htile_surface =
      S_028ABC_FULL_CACHE(1) | S_028ABC_PIPE_ALIGNED(1) | S_028ABC_VRS_HTILE_ENCODING(V_028ABC_VRS_HTILE_4BIT_ENCODING);
}

void
radv_initialise_ds_surface(const struct radv_device *device, struct radv_ds_buffer_info *ds,
                           struct radv_image_view *iview, VkImageAspectFlags ds_aspects)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   unsigned level = iview->vk.base_mip_level;
   bool stencil_only = iview->image->vk.format == VK_FORMAT_S8_UINT;

   assert(vk_format_get_plane_count(iview->image->vk.format) == 1);

   memset(ds, 0, sizeof(*ds));

   uint32_t max_slice = radv_surface_max_layer_count(iview) - 1;

   /* Recommended value for better performance with 4x and 8x. */
   ds->db_render_override2 = S_028010_DECOMPRESS_Z_ON_FLUSH(iview->image->vk.samples >= 4) |
                             S_028010_CENTROID_COMPUTATION_MODE(pdev->info.gfx_level >= GFX10_3);

   const struct ac_ds_state ds_state = {
      .surf = &iview->image->planes[0].surface,
      .va = radv_image_get_va(iview->image, 0),
      .format = vk_format_to_pipe_format(iview->image->vk.format),
      .width = iview->image->vk.extent.width,
      .height = iview->image->vk.extent.height,
      .level = level,
      .num_levels = iview->image->vk.mip_levels,
      .num_samples = iview->image->vk.samples,
      .first_layer = iview->vk.base_array_layer,
      .last_layer = max_slice,
      .stencil_only = stencil_only,
      .z_read_only = !(ds_aspects & VK_IMAGE_ASPECT_DEPTH_BIT),
      .stencil_read_only = !(ds_aspects & VK_IMAGE_ASPECT_STENCIL_BIT),
      .htile_enabled = radv_htile_enabled(iview->image, level),
      .htile_stencil_disabled = radv_image_tile_stencil_disabled(device, iview->image),
      .vrs_enabled = radv_image_has_vrs_htile(device, iview->image),
   };

   ac_init_ds_surface(&pdev->info, &ds_state, &ds->ac);

   const struct ac_mutable_ds_state mutable_ds_state = {
      .ds = &ds->ac,
      .format = vk_format_to_pipe_format(iview->image->vk.format),
      .tc_compat_htile_enabled = radv_htile_enabled(iview->image, level) && radv_image_is_tc_compat_htile(iview->image),
      .zrange_precision = true,
      .no_d16_compression = true,
   };

   ac_set_mutable_ds_surface_fields(&pdev->info, &mutable_ds_state, &ds->ac);

   if (pdev->info.gfx_level >= GFX11) {
      radv_gfx11_set_db_render_control(device, iview->image->vk.samples, &ds->db_render_control);
   }
}

void
radv_gfx11_set_db_render_control(const struct radv_device *device, unsigned num_samples, unsigned *db_render_control)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   unsigned max_allowed_tiles_in_wave = 0;

   if (pdev->info.has_dedicated_vram) {
      if (num_samples == 8)
         max_allowed_tiles_in_wave = 6;
      else if (num_samples == 4)
         max_allowed_tiles_in_wave = 13;
      else
         max_allowed_tiles_in_wave = 0;
   } else {
      if (num_samples == 8)
         max_allowed_tiles_in_wave = 7;
      else if (num_samples == 4)
         max_allowed_tiles_in_wave = 15;
      else
         max_allowed_tiles_in_wave = 0;
   }

   *db_render_control |= S_028000_MAX_ALLOWED_TILES_IN_WAVE(max_allowed_tiles_in_wave);
}

VKAPI_ATTR VkResult VKAPI_CALL
radv_GetMemoryFdKHR(VkDevice _device, const VkMemoryGetFdInfoKHR *pGetFdInfo, int *pFD)
{
   VK_FROM_HANDLE(radv_device, device, _device);
   VK_FROM_HANDLE(radv_device_memory, memory, pGetFdInfo->memory);

   assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);

   /* At the moment, we support only the below handle types. */
   assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
          pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);

   bool ret = radv_get_memory_fd(device, memory, pFD);
   if (ret == false)
      return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
   return VK_SUCCESS;
}

static uint32_t
radv_compute_valid_memory_types_attempt(struct radv_physical_device *pdev, enum radeon_bo_domain domains,
                                        enum radeon_bo_flag flags, enum radeon_bo_flag ignore_flags)
{
   /* Don't count GTT/CPU as relevant:
    *
    * - We're not fully consistent between the two.
    * - Sometimes VRAM gets VRAM|GTT.
    */
   const enum radeon_bo_domain relevant_domains = RADEON_DOMAIN_VRAM | RADEON_DOMAIN_GDS | RADEON_DOMAIN_OA;
   uint32_t bits = 0;
   for (unsigned i = 0; i < pdev->memory_properties.memoryTypeCount; ++i) {
      if ((domains & relevant_domains) != (pdev->memory_domains[i] & relevant_domains))
         continue;

      if ((flags & ~ignore_flags) != (pdev->memory_flags[i] & ~ignore_flags))
         continue;

      bits |= 1u << i;
   }

   return bits;
}

static uint32_t
radv_compute_valid_memory_types(struct radv_physical_device *pdev, enum radeon_bo_domain domains,
                                enum radeon_bo_flag flags)
{
   enum radeon_bo_flag ignore_flags = ~(RADEON_FLAG_NO_CPU_ACCESS | RADEON_FLAG_GTT_WC);
   uint32_t bits = radv_compute_valid_memory_types_attempt(pdev, domains, flags, ignore_flags);

   if (!bits) {
      ignore_flags |= RADEON_FLAG_GTT_WC;
      bits = radv_compute_valid_memory_types_attempt(pdev, domains, flags, ignore_flags);
   }

   if (!bits) {
      ignore_flags |= RADEON_FLAG_NO_CPU_ACCESS;
      bits = radv_compute_valid_memory_types_attempt(pdev, domains, flags, ignore_flags);
   }

   /* Avoid 32-bit memory types for shared memory. */
   bits &= ~pdev->memory_types_32bit;

   return bits;
}
VKAPI_ATTR VkResult VKAPI_CALL
radv_GetMemoryFdPropertiesKHR(VkDevice _device, VkExternalMemoryHandleTypeFlagBits handleType, int fd,
                              VkMemoryFdPropertiesKHR *pMemoryFdProperties)
{
   VK_FROM_HANDLE(radv_device, device, _device);
   struct radv_physical_device *pdev = radv_device_physical(device);

   switch (handleType) {
   case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT: {
      enum radeon_bo_domain domains;
      enum radeon_bo_flag flags;
      if (!device->ws->buffer_get_flags_from_fd(device->ws, fd, &domains, &flags))
         return vk_error(device, VK_ERROR_INVALID_EXTERNAL_HANDLE);

      pMemoryFdProperties->memoryTypeBits = radv_compute_valid_memory_types(pdev, domains, flags);
      return VK_SUCCESS;
   }
   default:
      /* The valid usage section for this function says:
       *
       *    "handleType must not be one of the handle types defined as
       *    opaque."
       *
       * So opaque handle types fall into the default "unsupported" case.
       */
      return vk_error(device, VK_ERROR_INVALID_EXTERNAL_HANDLE);
   }
}

VKAPI_ATTR VkResult VKAPI_CALL
radv_GetCalibratedTimestampsKHR(VkDevice _device, uint32_t timestampCount,
                                const VkCalibratedTimestampInfoKHR *pTimestampInfos, uint64_t *pTimestamps,
                                uint64_t *pMaxDeviation)
{
#ifndef _WIN32
   VK_FROM_HANDLE(radv_device, device, _device);
   const struct radv_physical_device *pdev = radv_device_physical(device);
   uint32_t clock_crystal_freq = pdev->info.clock_crystal_freq;
   int d;
   uint64_t begin, end;
   uint64_t max_clock_period = 0;

#ifdef CLOCK_MONOTONIC_RAW
   begin = vk_clock_gettime(CLOCK_MONOTONIC_RAW);
#else
   begin = vk_clock_gettime(CLOCK_MONOTONIC);
#endif

   for (d = 0; d < timestampCount; d++) {
      switch (pTimestampInfos[d].timeDomain) {
      case VK_TIME_DOMAIN_DEVICE_KHR:
         pTimestamps[d] = device->ws->query_value(device->ws, RADEON_TIMESTAMP);
         uint64_t device_period = DIV_ROUND_UP(1000000, clock_crystal_freq);
         max_clock_period = MAX2(max_clock_period, device_period);
         break;
      case VK_TIME_DOMAIN_CLOCK_MONOTONIC_KHR:
         pTimestamps[d] = vk_clock_gettime(CLOCK_MONOTONIC);
         max_clock_period = MAX2(max_clock_period, 1);
         break;

#ifdef CLOCK_MONOTONIC_RAW
      case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_KHR:
         pTimestamps[d] = begin;
         break;
#endif
      default:
         pTimestamps[d] = 0;
         break;
      }
   }

#ifdef CLOCK_MONOTONIC_RAW
   end = vk_clock_gettime(CLOCK_MONOTONIC_RAW);
#else
   end = vk_clock_gettime(CLOCK_MONOTONIC);
#endif

   *pMaxDeviation = vk_time_max_deviation(begin, end, max_clock_period);

   return VK_SUCCESS;
#else
   return VK_ERROR_FEATURE_NOT_PRESENT;
#endif
}

bool
radv_device_set_pstate(struct radv_device *device, bool enable)
{
   const struct radv_physical_device *pdev = radv_device_physical(device);
   const struct radv_instance *instance = radv_physical_device_instance(pdev);
   struct radeon_winsys *ws = device->ws;
   enum radeon_ctx_pstate pstate = enable ? instance->profile_pstate : RADEON_CTX_PSTATE_NONE;

   if (pdev->info.has_stable_pstate) {
      /* pstate is per-device; setting it for one ctx is sufficient.
       * We pick the first initialized one below. */
      for (unsigned i = 0; i < RADV_NUM_HW_CTX; i++)
         if (device->hw_ctx[i])
            return ws->ctx_set_pstate(device->hw_ctx[i], pstate) >= 0;
   }

   return true;
}

bool
radv_device_acquire_performance_counters(struct radv_device *device)
{
   bool result = true;
   simple_mtx_lock(&device->pstate_mtx);

   if (device->pstate_cnt == 0) {
      result = radv_device_set_pstate(device, true);
      if (result)
         ++device->pstate_cnt;
   }

   simple_mtx_unlock(&device->pstate_mtx);
   return result;
}

void
radv_device_release_performance_counters(struct radv_device *device)
{
   simple_mtx_lock(&device->pstate_mtx);

   if (--device->pstate_cnt == 0)
      radv_device_set_pstate(device, false);

   simple_mtx_unlock(&device->pstate_mtx);
}

VKAPI_ATTR VkResult VKAPI_CALL
radv_AcquireProfilingLockKHR(VkDevice _device, const VkAcquireProfilingLockInfoKHR *pInfo)
{
   VK_FROM_HANDLE(radv_device, device, _device);
   bool result = radv_device_acquire_performance_counters(device);
   return result ? VK_SUCCESS : VK_ERROR_UNKNOWN;
}

VKAPI_ATTR void VKAPI_CALL
radv_ReleaseProfilingLockKHR(VkDevice _device)
{
   VK_FROM_HANDLE(radv_device, device, _device);
   radv_device_release_performance_counters(device);
}

VKAPI_ATTR void VKAPI_CALL
radv_GetDeviceImageSubresourceLayoutKHR(VkDevice device, const VkDeviceImageSubresourceInfoKHR *pInfo,
                                        VkSubresourceLayout2KHR *pLayout)
{
   UNUSED VkResult result;
   VkImage image;

   result =
      radv_image_create(device, &(struct radv_image_create_info){.vk_info = pInfo->pCreateInfo}, NULL, &image, true);
   assert(result == VK_SUCCESS);

   radv_GetImageSubresourceLayout2KHR(device, image, pInfo->pSubresource, pLayout);

   radv_DestroyImage(device, image, NULL);
}