vulkan: Add VK_EXT_calibrated_timestamps extension (radv and anv) [v5]

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.

v3:
	Set device->lost when returning DEVICE_LOST.
	Use MAX2 and DIV_ROUND_UP instead of open coding these.
	Delete spurious TIMESTAMP in radv version.

	Suggested-by: Jason Ekstrand <jason@jlekstrand.net>
	Suggested-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

v4:
	Add anv_gem_reg_read to anv_gem_stubs.c

	Suggested-by: Jason Ekstrand <jason@jlekstrand.net>

v5:
	Adjust maxDeviation computation to max(sampled_clock_period) +
	sample_interval.

	Suggested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
	Suggested-by: Jason Ekstrand <jason@jlekstrand.net>

Signed-off-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This commit is contained in:
Keith Packard 2018-10-11 16:05:18 -07:00
parent a11cafbd7a
commit 67a2c1493c
7 changed files with 270 additions and 0 deletions

View File

@ -4957,3 +4957,122 @@ 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, &current);
if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
ret = clock_gettime(CLOCK_MONOTONIC, &current);
if (ret < 0)
return 0;
return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
}
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;
uint64_t max_clock_period = 0;
begin = radv_clock_gettime(CLOCK_MONOTONIC_RAW);
for (d = 0; d < timestampCount; d++) {
switch (pTimestampInfos[d].timeDomain) {
case VK_TIME_DOMAIN_DEVICE_EXT:
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_EXT:
pTimestamps[d] = radv_clock_gettime(CLOCK_MONOTONIC);
max_clock_period = MAX2(max_clock_period, 1);
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);
/*
* The maximum deviation is the sum of the interval over which we
* perform the sampling and the maximum period of any sampled
* clock. That's because the maximum skew between any two sampled
* clock edges is when the sampled clock with the largest period is
* sampled at the end of that period but right at the beginning of the
* sampling interval and some other clock is sampled right at the
* begining of its sampling period and right at the end of the
* sampling interval. Let's assume the GPU has the longest clock
* period and that the application is sampling GPU and monotonic:
*
* s e
* w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
* Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
*
* g
* 0 1 2 3
* GPU -----_____-----_____-----_____-----_____
*
* m
* x y z 0 1 2 3 4 5 6 7 8 9 a b c
* Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
*
* Interval <----------------->
* Deviation <-------------------------->
*
* s = read(raw) 2
* g = read(GPU) 1
* m = read(monotonic) 2
* e = read(raw) b
*
* We round the sample interval up by one tick to cover sampling error
* in the interval clock
*/
uint64_t sample_interval = end - begin + 1;
*pMaxDeviation = sample_interval + max_clock_period;
return VK_SUCCESS;
}

View File

@ -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'),

View File

@ -3035,6 +3035,133 @@ 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, &current);
if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
ret = clock_gettime(CLOCK_MONOTONIC, &current);
if (ret < 0)
return 0;
return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
}
#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;
uint64_t max_clock_period = 0;
begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
for (d = 0; d < timestampCount; d++) {
switch (pTimestampInfos[d].timeDomain) {
case VK_TIME_DOMAIN_DEVICE_EXT:
ret = anv_gem_reg_read(device, TIMESTAMP | 1,
&pTimestamps[d]);
if (ret != 0) {
device->lost = TRUE;
return VK_ERROR_DEVICE_LOST;
}
uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
max_clock_period = MAX2(max_clock_period, device_period);
break;
case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
max_clock_period = MAX2(max_clock_period, 1);
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);
/*
* The maximum deviation is the sum of the interval over which we
* perform the sampling and the maximum period of any sampled
* clock. That's because the maximum skew between any two sampled
* clock edges is when the sampled clock with the largest period is
* sampled at the end of that period but right at the beginning of the
* sampling interval and some other clock is sampled right at the
* begining of its sampling period and right at the end of the
* sampling interval. Let's assume the GPU has the longest clock
* period and that the application is sampling GPU and monotonic:
*
* s e
* w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
* Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
*
* g
* 0 1 2 3
* GPU -----_____-----_____-----_____-----_____
*
* m
* x y z 0 1 2 3 4 5 6 7 8 9 a b c
* Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
*
* Interval <----------------->
* Deviation <-------------------------->
*
* s = read(raw) 2
* g = read(GPU) 1
* m = read(monotonic) 2
* e = read(raw) b
*
* We round the sample interval up by one tick to cover sampling error
* in the interval clock
*/
uint64_t sample_interval = end - begin + 1;
*pMaxDeviation = sample_interval + max_clock_period;
return VK_SUCCESS;
}
/* vk_icd.h does not declare this function, so we declare it here to
* suppress Wmissing-prototypes.
*/

View File

@ -127,6 +127,7 @@ 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:

View File

@ -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

View File

@ -251,3 +251,10 @@ anv_gem_syncobj_wait(struct anv_device *device,
{
unreachable("Unused");
}
int
anv_gem_reg_read(struct anv_device *device,
uint32_t offset, uint64_t *result)
{
unreachable("Unused");
}

View File

@ -1109,6 +1109,8 @@ 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,