Offers three clocks, device, clock monotonic and clock monotonic
raw. Could use some kernel support to reduce the deviation between
clock values.
v2:
Ensure deviation is at least as big as the GPU time interval.
Signed-off-by: Keith Packard <keithp@keithp.com>
---
src/amd/vulkan/radv_device.c | 84 ++++++++++++++++++++++++++++
src/amd/vulkan/radv_extensions.py | 1 +
src/intel/vulkan/anv_device.c | 88 ++++++++++++++++++++++++++++++
src/intel/vulkan/anv_extensions.py | 1 +
src/intel/vulkan/anv_gem.c | 13 +++++
src/intel/vulkan/anv_private.h | 2 +
6 files changed, 189 insertions(+)
diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
index 174922780fc..29f0afbc69b 100644
--- a/src/amd/vulkan/radv_device.c
+++ b/src/amd/vulkan/radv_device.c
@@ -4955,3 +4955,87 @@ radv_GetDeviceGroupPeerMemoryFeatures(
VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
}
+
+static const VkTimeDomainEXT radv_time_domains[] = {
+ VK_TIME_DOMAIN_DEVICE_EXT,
+ VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
+ VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
+};
+
+VkResult radv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
+ VkPhysicalDevice physicalDevice,
+ uint32_t *pTimeDomainCount,
+ VkTimeDomainEXT *pTimeDomains)
+{
+ int d;
+ VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
+
+ for (d = 0; d < ARRAY_SIZE(radv_time_domains); d++) {
+ vk_outarray_append(&out, i) {
+ *i = radv_time_domains[d];
+ }
+ }
+
+ return vk_outarray_status(&out);
+}
+
+static uint64_t
+radv_clock_gettime(clockid_t clock_id)
+{
+ struct timespec current;
+ int ret;
+
+ ret = clock_gettime(clock_id, ¤t);
+ if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
+ ret = clock_gettime(CLOCK_MONOTONIC, ¤t);
+ if (ret < 0)
+ return 0;
+
+ return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
+}
+
+#define TIMESTAMP 0x2358
+
+VkResult radv_GetCalibratedTimestampsEXT(
+ VkDevice _device,
+ uint32_t timestampCount,
+ const VkCalibratedTimestampInfoEXT *pTimestampInfos,
+ uint64_t *pTimestamps,
+ uint64_t *pMaxDeviation)
+{
+ RADV_FROM_HANDLE(radv_device, device, _device);
+ uint32_t clock_crystal_freq = device->physical_device->rad_info.clock_crystal_freq;
+ int d;
+ uint64_t begin, end;
+
+ begin = radv_clock_gettime(CLOCK_MONOTONIC_RAW);
+
+ for (d = 0; d < timestampCount; d++) {
+ switch (pTimestampInfos[d].timeDomain) {
+ case VK_TIME_DOMAIN_DEVICE_EXT:
+ /* XXX older kernels don't support this interface. */
+ pTimestamps[d] = device->ws->query_value(device->ws,
+ RADEON_TIMESTAMP);
+ break;
+ case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
+ pTimestamps[d] = radv_clock_gettime(CLOCK_MONOTONIC);
+ break;
+
+ case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
+ pTimestamps[d] = begin;
+ break;
+ default:
+ pTimestamps[d] = 0;
+ break;
+ }
+ }
+
+ end = radv_clock_gettime(CLOCK_MONOTONIC_RAW);
+
+ uint64_t clock_period = end - begin;
+ uint64_t device_period = (1000000 + clock_crystal_freq - 1) / clock_crystal_freq;
+
+ *pMaxDeviation = clock_period > device_period ? clock_period : device_period;
+
+ return VK_SUCCESS;
+}
diff --git a/src/amd/vulkan/radv_extensions.py b/src/amd/vulkan/radv_extensions.py
index 5dcedae1c63..4c81d3f0068 100644
--- a/src/amd/vulkan/radv_extensions.py
+++ b/src/amd/vulkan/radv_extensions.py
@@ -92,6 +92,7 @@ EXTENSIONS = [
Extension('VK_KHR_display', 23, 'VK_USE_PLATFORM_DISPLAY_KHR'),
Extension('VK_EXT_direct_mode_display', 1, 'VK_USE_PLATFORM_DISPLAY_KHR'),
Extension('VK_EXT_acquire_xlib_display', 1, 'VK_USE_PLATFORM_XLIB_XRANDR_EXT'),
+ Extension('VK_EXT_calibrated_timestamps', 1, True),
Extension('VK_EXT_conditional_rendering', 1, True),
Extension('VK_EXT_conservative_rasterization', 1, 'device->rad_info.chip_class >= GFX9'),
Extension('VK_EXT_display_surface_counter', 1, 'VK_USE_PLATFORM_DISPLAY_KHR'),
diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c
index a2551452eb1..6a6539c9685 100644
--- a/src/intel/vulkan/anv_device.c
+++ b/src/intel/vulkan/anv_device.c
@@ -3021,6 +3021,94 @@ void anv_DestroyFramebuffer(
vk_free2(&device->alloc, pAllocator, fb);
}
+static const VkTimeDomainEXT anv_time_domains[] = {
+ VK_TIME_DOMAIN_DEVICE_EXT,
+ VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
+ VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
+};
+
+VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
+ VkPhysicalDevice physicalDevice,
+ uint32_t *pTimeDomainCount,
+ VkTimeDomainEXT *pTimeDomains)
+{
+ int d;
+ VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
+
+ for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
+ vk_outarray_append(&out, i) {
+ *i = anv_time_domains[d];
+ }
+ }
+
+ return vk_outarray_status(&out);
+}
+
+static uint64_t
+anv_clock_gettime(clockid_t clock_id)
+{
+ struct timespec current;
+ int ret;
+
+ ret = clock_gettime(clock_id, ¤t);
+ if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
+ ret = clock_gettime(CLOCK_MONOTONIC, ¤t);
+ if (ret < 0)
+ return 0;
+
+ return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
+}
One lf these days, we should unify these. I know this is at least the second copy in anv and we've got it hand-rolled a couple other places. I don't care too much today though so meh.
+
+#define TIMESTAMP 0x2358
+
+VkResult anv_GetCalibratedTimestampsEXT(
+ VkDevice _device,
+ uint32_t timestampCount,
+ const VkCalibratedTimestampInfoEXT *pTimestampInfos,
+ uint64_t *pTimestamps,
+ uint64_t *pMaxDeviation)
+{
+ ANV_FROM_HANDLE(anv_device, device, _device);
+ uint64_t timestamp_frequency = device->info.timestamp_frequency;
+ int ret;
+ int d;
+ uint64_t begin, end;
+
+ begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
+
+ for (d = 0; d < timestampCount; d++) {
+ switch (pTimestampInfos[d].timeDomain) {
+ case VK_TIME_DOMAIN_DEVICE_EXT:
+ /* XXX older kernels don't support this interface. */
I don't think we support kernels that old, so you can drop the comment.
+ ret = anv_gem_reg_read(device, TIMESTAMP | 1,
+ &pTimestamps[d]);
+
+ if (ret != 0)
+ return VK_ERROR_DEVICE_LOST;
If you're going to return device_lost, you should set device->lost = true;
+ break;
+ case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
+ pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
+ break;
+
+ case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
+ pTimestamps[d] = begin;
+ break;
+ default:
+ pTimestamps[d] = 0;
+ break;
+ }
+ }
+
+ end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
+
+ uint64_t clock_period = end - begin;
+ uint64_t device_period = (1000000000 + timestamp_frequency - 1) / timestamp_frequency;
I think you just want DIV_ROUND_UP(a, b)
+
+ *pMaxDeviation = clock_period > device_period ? clock_period : device_period;
MAX2(a, b) is your friend.
+
+ return VK_SUCCESS;
+}
+
/* vk_icd.h does not declare this function, so we declare it here to
* suppress Wmissing-prototypes.
*/
diff --git a/src/intel/vulkan/anv_extensions.py b/src/intel/vulkan/anv_extensions.py
index d4915c95013..a8535964da7 100644
--- a/src/intel/vulkan/anv_extensions.py
+++ b/src/intel/vulkan/anv_extensions.py
@@ -126,6 +126,7 @@ EXTENSIONS = [
Extension('VK_EXT_vertex_attribute_divisor', 3, True),
Extension('VK_EXT_post_depth_coverage', 1, 'device->info.gen >= 9'),
Extension('VK_EXT_sampler_filter_minmax', 1, 'device->info.gen >= 9'),
+ Extension('VK_EXT_calibrated_timestamps', 1, True),
]
class VkVersion:
diff --git a/src/intel/vulkan/anv_gem.c b/src/intel/vulkan/anv_gem.c
index c43b5ef9e06..1bdf040c1a3 100644
--- a/src/intel/vulkan/anv_gem.c
+++ b/src/intel/vulkan/anv_gem.c
@@ -423,6 +423,19 @@ anv_gem_fd_to_handle(struct anv_device *device, int fd)
return args.handle;
}
+int
+anv_gem_reg_read(struct anv_device *device, uint32_t offset, uint64_t *result)
+{
+ struct drm_i915_reg_read args = {
+ .offset = offset
+ };
+
+ int ret = anv_ioctl(device->fd, DRM_IOCTL_I915_REG_READ, &args);
+
+ *result = args.val;
+ return ret;
+}
+
#ifndef SYNC_IOC_MAGIC
/* duplicated from linux/sync_file.h to avoid build-time dependency
* on new (v4.7) kernel headers. Once distro's are mostly using
diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h
index 599b903f25c..08376b00c8e 100644
--- a/src/intel/vulkan/anv_private.h
+++ b/src/intel/vulkan/anv_private.h
@@ -1103,6 +1103,8 @@ int anv_gem_get_aperture(int fd, uint64_t *size);
int anv_gem_gpu_get_reset_stats(struct anv_device *device,
uint32_t *active, uint32_t *pending);
int anv_gem_handle_to_fd(struct anv_device *device, uint32_t gem_handle);
+int anv_gem_reg_read(struct anv_device *device,
+ uint32_t offset, uint64_t *result);
uint32_t anv_gem_fd_to_handle(struct anv_device *device, int fd);
int anv_gem_set_caching(struct anv_device *device, uint32_t gem_handle, uint32_t caching);
int anv_gem_set_domain(struct anv_device *device, uint32_t gem_handle,
--
2.19.1
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel