[PATCH v4 0/1] MR9859: Draft: winevulkan: Use correct time domain for calibrated timestamps

[Etaash Mathamsetty (＠etaash.mathamsetty)]

I am not sure how to address this fully, but the following patch should at least make it accurate before the machine suspends.

--
  v4: winevulkan: Use correct time domain for calibrated timestamps
 
https://gitlab.winehq.org/wine/wine/-/merge_requests/9859

[PATCH v4 1/1] winevulkan: Use correct time domain for calibrated timestamps

[Etaash Mathamsetty]

From: Etaash Mathamsetty <[e-mail]>

---
 dlls/winevulkan/vulkan.c | 14 +++-----------
 1 file changed, 3 insertions(+), 11 deletions(-)

diff --git a/dlls/winevulkan/vulkan.c b/dlls/winevulkan/vulkan.c
index 0f49148ee15..498a6a42c7b 100644
--- a/dlls/winevulkan/vulkan.c
+++ b/dlls/winevulkan/vulkan.c
@@ -585,11 +585,7 @@ void wine_vkGetPhysicalDeviceExternalFencePropertiesKHR(VkPhysicalDevice client_
 static inline VkTimeDomainEXT get_performance_counter_time_domain(void)
 {
 #if !defined(__APPLE__) && defined(HAVE_CLOCK_GETTIME)
-# ifdef CLOCK_MONOTONIC_RAW
-    return VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT;
-# else
     return VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT;
-# endif
 #else
     FIXME("No mapping for VK_TIME_DOMAIN_QUERY_PERFORMANCE_COUNTER_EXT on this platform.\n");
     return VK_TIME_DOMAIN_QUERY_PERFORMANCE_COUNTER_EXT;
@@ -616,7 +612,7 @@ static inline uint64_t convert_timestamp(VkTimeDomainEXT host_domain, VkTimeDoma
         return value;
 
     /* Convert between MONOTONIC time in ns -> QueryPerformanceCounter */
-    if ((host_domain == VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT || host_domain == VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT)
+    if (host_domain == VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
             && target_domain == VK_TIME_DOMAIN_QUERY_PERFORMANCE_COUNTER_EXT)
         return convert_monotonic_timestamp(value);
 
@@ -663,7 +659,7 @@ static VkResult wine_vk_get_time_domains(struct vulkan_physical_device *physical
                                          VkTimeDomainEXT *time_domains,
                                          VkResult (*get_domains)(VkPhysicalDevice, uint32_t *, VkTimeDomainEXT *))
 {
-    BOOL supports_device = FALSE, supports_monotonic = FALSE, supports_monotonic_raw = FALSE;
+    BOOL supports_device = FALSE, supports_monotonic = FALSE;
     const VkTimeDomainEXT performance_counter_domain = get_performance_counter_time_domain();
     VkTimeDomainEXT *host_time_domains;
     uint32_t host_time_domain_count;
@@ -693,8 +689,6 @@ static VkResult wine_vk_get_time_domains(struct vulkan_physical_device *physical
             supports_device = TRUE;
         else if (host_time_domains[i] == VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT)
             supports_monotonic = TRUE;
-        else if (host_time_domains[i] == VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT)
-            supports_monotonic_raw = TRUE;
         else
             FIXME("Unknown time domain %d\n", host_time_domains[i]);
     }
@@ -704,9 +698,7 @@ static VkResult wine_vk_get_time_domains(struct vulkan_physical_device *physical
     out_time_domain_count = 0;
 
     /* Map our monotonic times -> QPC */
-    if (supports_monotonic_raw && performance_counter_domain == VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT)
-        out_time_domains[out_time_domain_count++] = VK_TIME_DOMAIN_QUERY_PERFORMANCE_COUNTER_EXT;
-    else if (supports_monotonic && performance_counter_domain == VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT)
+    if (supports_monotonic && performance_counter_domain == VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT)
         out_time_domains[out_time_domain_count++] = VK_TIME_DOMAIN_QUERY_PERFORMANCE_COUNTER_EXT;
     else
         FIXME("VK_TIME_DOMAIN_QUERY_PERFORMANCE_COUNTER_EXT not supported on this platform.\n");
-- 
GitLab

https://gitlab.winehq.org/wine/wine/-/merge_requests/9859

[Etaash Mathamsetty (＠etaash.mathamsetty)]

On Wed Jan  7 22:53:47 2026 +0000, Rémi Bernon wrote:
> Is this because we don't use CLOCK_MONOTONIC_RAW in ntdll for QPC when
> we need to make sure both QPC / vulkan clock source match?
> Yes, similar fix needs to be applied to vkd3d as well.

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/9859#note_126521

[Etaash Mathamsetty (＠etaash.mathamsetty)]

On Wed Jan  7 22:55:25 2026 +0000, Rémi Bernon wrote:
> If that's the case, isn't this actually a regression from
> https://gitlab.winehq.org/wine/wine/-/merge_requests/8916? If we cannot
> have a BOOTTIME clock source here, IMO we should revert that MR instead.
> I have an approach that should alleviate most of the issues. Essentially it is QueryPerformanceCounter and clock_gettime(monotonic) and then computing the difference. This would be inaccurate on a stale timestamp before suspend, but Paul Gofman said that this isn't too much of an issue (unless proven otherwise in some application).

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/9859#note_126523

[Paul Gofman (＠gofman)]

On Wed Jan  7 22:55:39 2026 +0000, Etaash Mathamsetty wrote:
> I have an approach that should alleviate most of the issues. Essentially
> it is QueryPerformanceCounter and clock_gettime(monotonic) and then
> computing the difference. This would be inaccurate on a stale timestamp
> before suspend, but Paul Gofman said that this isn't too much of an
> issue (unless proven otherwise in some application).
> Meanwhile, if we can reconsider using BOOTTIME for QPC, I'd also think (as Rémi suggested) that would be a cleaner solution.

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/9859#note_126524