diff mbox series

[i-g-t,v6,2/2] tests/xe/pmu: Add pmu tests

Message ID 20250127223301.2819108-3-vinay.belgaumkar@intel.com (mailing list archive)
State New
Headers show
Series tests/intel/xe_pmu: Add PMU tests | expand

Commit Message

Belgaumkar, Vinay Jan. 27, 2025, 10:33 p.m. UTC 
Simple tests for validating the PMU implementation for GT C6
residencies.

v2: Rename rc6-residency-* to gt-c6-residency and remove freq tests.
v3: Keep just gt-c6 tests, add frequency tests later.
v4: Review comments (Riana)

Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Cc: Riana Tauro <riana.tauro@intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
---
 tests/intel/xe_pmu.c | 176 +++++++++++++++++++++++++++++++++++++++++++
 tests/meson.build    |   1 +
 2 files changed, 177 insertions(+)
 create mode 100644 tests/intel/xe_pmu.c

Comments

Lucas De Marchi Jan. 27, 2025, 11:56 p.m. UTC | #1 
On Mon, Jan 27, 2025 at 02:33:01PM -0800, Vinay Belgaumkar wrote:
>Simple tests for validating the PMU implementation for GT C6
>residencies.
>
>v2: Rename rc6-residency-* to gt-c6-residency and remove freq tests.
>v3: Keep just gt-c6 tests, add frequency tests later.
>v4: Review comments (Riana)
>
>Cc: Lucas De Marchi <lucas.demarchi@intel.com>
>Cc: Riana Tauro <riana.tauro@intel.com>
>Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
>Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
>---
> tests/intel/xe_pmu.c | 176 +++++++++++++++++++++++++++++++++++++++++++
> tests/meson.build    |   1 +
> 2 files changed, 177 insertions(+)
> create mode 100644 tests/intel/xe_pmu.c
>
>diff --git a/tests/intel/xe_pmu.c b/tests/intel/xe_pmu.c
>new file mode 100644
>index 000000000..61a7a7e38
>--- /dev/null
>+++ b/tests/intel/xe_pmu.c
>@@ -0,0 +1,176 @@
>+// SPDX-License-Identifier: MIT
>+/*
>+ * Copyright © 2025 Intel Corporation
>+ */
>+
>+/**
>+ * TEST: Test Xe PMU functionality
>+ * Category: Perf Monitoring Unit
>+ * Mega feature: Perf Monitoring Unit
>+ * Sub-category: Power Management

this is not accurate... power management is 1 of the possible counters
we can expose from kernel to userspace.

>+ * Functionality: Power/Perf
>+ * Test category: Functional tests
>+ */
>+
>+#include <fcntl.h>
>+#include <limits.h>
>+#include <time.h>
>+#include <errno.h>
>+#include <dirent.h>
>+#include <string.h>
>+#include <sys/time.h>
>+
>+#include "igt.h"
>+#include "igt_device.h"
>+#include "igt_power.h"
>+#include "igt_sysfs.h"
>+#include "igt_perf.h"
>+
>+#include "lib/igt_syncobj.h"
>+#include "xe/xe_ioctl.h"
>+#include "xe/xe_gt.h"
>+#include "xe/xe_query.h"
>+#include "xe/xe_spin.h"
>+#include "xe/xe_util.h"
>+
>+#define SLEEP_DURATION 2 /* in seconds */
>+const double tolerance = 0.1;
>+const unsigned long batch_duration_ns = 500e6;

unused

>+const char *no_debug_data = "\0";

no idea what this is for

>+
>+#define __assert_within_epsilon(x, ref, tol_up, tol_down, debug_data) \
>+	igt_assert_f((double)(x) <= (1.0 + (tol_up)) * (double)(ref) && \
>+		     (double)(x) >= (1.0 - (tol_down)) * (double)(ref), \
>+		     "'%s' != '%s' (%f not within +%.1f%%/-%.1f%% tolerance of %f)\n%s\n",\
>+		     #x, #ref, (double)(x), \
>+		     (tol_up) * 100.0, (tol_down) * 100.0, \
>+		     (double)(ref), debug_data)
>+
>+#define assert_within_epsilon(x, ref, tolerance) \
>+	__assert_within_epsilon(x, ref, tolerance, tolerance, no_debug_data)

git grep __assert_within_epsilon

we already have too much of these copy and paste. Please, don't add
more.

>+
>+static int open_pmu(int xe, uint64_t config)
>+{
>+	int fd;
>+
>+	fd = perf_xe_open(xe, config);
>+	igt_skip_on(fd < 0 && errno == ENODEV);
>+	igt_assert(fd >= 0);
>+
>+	return fd;
>+}
>+
>+static uint64_t __pmu_read_single(int fd, uint64_t *ts)
>+{
>+	uint64_t data[2];
>+
>+	igt_assert_eq(read(fd, data, sizeof(data)), sizeof(data));
>+	if (ts)
>+		*ts = data[1];
>+
>+	return data[0];
>+}
>+
>+/**
>+ * SUBTEST: gt-c6
>+ * Description: Basic residency test to validate idle residency
>+ *		measured over a time interval is within the tolerance

why is this comment on top of measured_usleep()?

>+ */
>+static unsigned int measured_usleep(unsigned int usec)
>+{
>+	struct timespec ts = { };
>+	unsigned int slept;
>+
>+	slept = igt_nsec_elapsed(&ts);

this function is another copy and paste that is wrong everywhere.


>+	igt_assert(slept == 0);
>+	do {
>+		usleep(usec - slept);

slept is in nanosec, while usec is in usec.

>+		slept = igt_nsec_elapsed(&ts) / 1000;
>+	} while (slept < usec);
>+
>+	return igt_nsec_elapsed(&ts) / 1000;

we have macros to convert usec/msec/nsec

>+}
>+
>+static unsigned long read_idle_residency(int fd, int gt)
>+{
>+	unsigned long residency = 0;
>+	int gt_fd;
>+
>+	gt_fd = xe_sysfs_gt_open(fd, gt);
>+	igt_assert(gt_fd >= 0);
>+	igt_assert(igt_sysfs_scanf(gt_fd, "gtidle/idle_residency_ms", "%lu", &residency) == 1);
>+	close(gt_fd);
>+
>+	return residency;
>+}
>+
>+static u64 get_event_config(int xe, unsigned int gt, char *event)
>+{
>+	int ret;
>+	char xe_device[100];
>+	u64 pmu_config;
>+	u32 start, end;
>+
>+	xe_perf_device(xe, xe_device, sizeof(xe_device));
>+	ret = perf_event_config(xe_device, event, &pmu_config);
>+	igt_assert(ret >= 0);
>+	ret = perf_event_format(xe_device, "gt", &start, &end);
>+	igt_assert(ret >= 0);
>+	pmu_config |= (u64) gt << start;
>+
>+	return pmu_config;
>+}
>+
>+static void test_gt_c6(int xe, unsigned int gt)
>+{
>+	int pmu_fd;
>+	u64 pmu_config;
>+	char event[100];
>+	uint64_t ts[2];
>+	unsigned long slept, start, end;
>+	uint64_t val;
>+
>+	/* Get the PMU config for the gt-c6 event */
>+	sprintf(event, "gt-c6-residency");
>+	pmu_config = get_event_config(xe, gt, event);
>+
>+	pmu_fd = open_pmu(xe, pmu_config);
>+
>+	igt_require_f(igt_wait(xe_gt_is_in_c6(xe, gt), 1000, 10), "GT %d should be in C6\n", gt);
>+
>+	/* While idle check full RC6. */
>+	start = read_idle_residency(xe, gt);
>+	val = __pmu_read_single(pmu_fd, &ts[0]);
>+	slept = measured_usleep(SLEEP_DURATION * USEC_PER_SEC) / 1000;
>+	end = read_idle_residency(xe, gt);
>+	val = __pmu_read_single(pmu_fd, &ts[1]) - val;

I actually expected 2 subtests:

1) gt_c6_idle

	we should be in rc6 for the entire time when nothing is running

2) gt_c6_full_load

	we should not be in rc6 when something is keeping the engine
	busy

For (2), we can start a work load with xe_cork


>+
>+	igt_debug("gt%u: slept=%lu, perf=%"PRIu64"\n",
>+		  gt, slept,  val);
>+
>+	igt_debug("Start res: %lu, end_res: %lu", start, end);

is this reading the value from sysfs to aid debugging?


I tested this with the kernel implementation, before merging it, and it
worked. I don't mind if this is merged and cleaned on top. I will leave
it for you and Riana to decide.

thanks
Lucas De Marchi

>+
>+	assert_within_epsilon(val,
>+			      (ts[1] - ts[0])/USEC_PER_SEC,
>+			      tolerance);
>+	close(pmu_fd);
>+}
>+
>+igt_main
>+{
>+	int fd, gt;
>+
>+	igt_fixture {
>+		fd = drm_open_driver(DRIVER_XE);
>+		igt_require(!IS_PONTEVECCHIO(xe_dev_id(fd)));
>+	}
>+
>+	igt_describe("Validate PMU gt-c6 residency counters");
>+	igt_subtest("gt-c6")
>+		xe_for_each_gt(fd, gt)
>+			test_gt_c6(fd, gt);
>+
>+	igt_fixture {
>+		close(fd);
>+	}
>+}
>diff --git a/tests/meson.build b/tests/meson.build
>index 33dffad31..d20f50766 100644
>--- a/tests/meson.build
>+++ b/tests/meson.build
>@@ -309,6 +309,7 @@ intel_xe_progs = [
> 	'xe_pat',
> 	'xe_peer2peer',
> 	'xe_pm',
>+	'xe_pmu',
> 	'xe_pm_residency',
> 	'xe_prime_self_import',
> 	'xe_query',
>-- 
>2.38.1
>
Belgaumkar, Vinay Jan. 30, 2025, 9:42 p.m. UTC | #2 
On 1/27/2025 3:56 PM, Lucas De Marchi wrote:
> On Mon, Jan 27, 2025 at 02:33:01PM -0800, Vinay Belgaumkar wrote:
>> Simple tests for validating the PMU implementation for GT C6
>> residencies.
>>
>> v2: Rename rc6-residency-* to gt-c6-residency and remove freq tests.
>> v3: Keep just gt-c6 tests, add frequency tests later.
>> v4: Review comments (Riana)
>>
>> Cc: Lucas De Marchi <lucas.demarchi@intel.com>
>> Cc: Riana Tauro <riana.tauro@intel.com>
>> Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
>> Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
>> ---
>> tests/intel/xe_pmu.c | 176 +++++++++++++++++++++++++++++++++++++++++++
>> tests/meson.build    |   1 +
>> 2 files changed, 177 insertions(+)
>> create mode 100644 tests/intel/xe_pmu.c
>>
>> diff --git a/tests/intel/xe_pmu.c b/tests/intel/xe_pmu.c
>> new file mode 100644
>> index 000000000..61a7a7e38
>> --- /dev/null
>> +++ b/tests/intel/xe_pmu.c
>> @@ -0,0 +1,176 @@
>> +// SPDX-License-Identifier: MIT
>> +/*
>> + * Copyright © 2025 Intel Corporation
>> + */
>> +
>> +/**
>> + * TEST: Test Xe PMU functionality
>> + * Category: Perf Monitoring Unit
>> + * Mega feature: Perf Monitoring Unit
>> + * Sub-category: Power Management
>
> this is not accurate... power management is 1 of the possible counters
> we can expose from kernel to userspace.
Ok, maybe Telemetry?
>
>> + * Functionality: Power/Perf
>> + * Test category: Functional tests
>> + */
>> +
>> +#include <fcntl.h>
>> +#include <limits.h>
>> +#include <time.h>
>> +#include <errno.h>
>> +#include <dirent.h>
>> +#include <string.h>
>> +#include <sys/time.h>
>> +
>> +#include "igt.h"
>> +#include "igt_device.h"
>> +#include "igt_power.h"
>> +#include "igt_sysfs.h"
>> +#include "igt_perf.h"
>> +
>> +#include "lib/igt_syncobj.h"
>> +#include "xe/xe_ioctl.h"
>> +#include "xe/xe_gt.h"
>> +#include "xe/xe_query.h"
>> +#include "xe/xe_spin.h"
>> +#include "xe/xe_util.h"
>> +
>> +#define SLEEP_DURATION 2 /* in seconds */
>> +const double tolerance = 0.1;
>> +const unsigned long batch_duration_ns = 500e6;
>
> unused
ok. Remnant from the freq tests.
>
>> +const char *no_debug_data = "\0";
>
> no idea what this is for
It is used in the assert macro below or an error message when assert 
happens.
>
>> +
>> +#define __assert_within_epsilon(x, ref, tol_up, tol_down, debug_data) \
>> +    igt_assert_f((double)(x) <= (1.0 + (tol_up)) * (double)(ref) && \
>> +             (double)(x) >= (1.0 - (tol_down)) * (double)(ref), \
>> +             "'%s' != '%s' (%f not within +%.1f%%/-%.1f%% tolerance 
>> of %f)\n%s\n",\
>> +             #x, #ref, (double)(x), \
>> +             (tol_up) * 100.0, (tol_down) * 100.0, \
>> +             (double)(ref), debug_data)
>> +
>> +#define assert_within_epsilon(x, ref, tolerance) \
>> +    __assert_within_epsilon(x, ref, tolerance, tolerance, 
>> no_debug_data)
>
> git grep __assert_within_epsilon
>
> we already have too much of these copy and paste. Please, don't add
> more.
ok, will move to a lib.
>
>> +
>> +static int open_pmu(int xe, uint64_t config)
>> +{
>> +    int fd;
>> +
>> +    fd = perf_xe_open(xe, config);
>> +    igt_skip_on(fd < 0 && errno == ENODEV);
>> +    igt_assert(fd >= 0);
>> +
>> +    return fd;
>> +}
>> +
>> +static uint64_t __pmu_read_single(int fd, uint64_t *ts)
>> +{
>> +    uint64_t data[2];
>> +
>> +    igt_assert_eq(read(fd, data, sizeof(data)), sizeof(data));
>> +    if (ts)
>> +        *ts = data[1];
>> +
>> +    return data[0];
>> +}
>> +
>> +/**
>> + * SUBTEST: gt-c6
>> + * Description: Basic residency test to validate idle residency
>> + *        measured over a time interval is within the tolerance
>
> why is this comment on top of measured_usleep()?
will move.
>
>> + */
>> +static unsigned int measured_usleep(unsigned int usec)
>> +{
>> +    struct timespec ts = { };
>> +    unsigned int slept;
>> +
>> +    slept = igt_nsec_elapsed(&ts);
>
> this function is another copy and paste that is wrong everywhere.
ok.
>
>
>> +    igt_assert(slept == 0);
>> +    do {
>> +        usleep(usec - slept);
>
> slept is in nanosec, while usec is in usec.
we are dividing by 1000, so slept is usec as well (except the first 
time, but that is asserted to be 0 anyways).
>
>> +        slept = igt_nsec_elapsed(&ts) / 1000;
>> +    } while (slept < usec);
>> +
>> +    return igt_nsec_elapsed(&ts) / 1000;
>
> we have macros to convert usec/msec/nsec
Yup, I will add a separate patch for moving this to a lib.
>
>> +}
>> +
>> +static unsigned long read_idle_residency(int fd, int gt)
>> +{
>> +    unsigned long residency = 0;
>> +    int gt_fd;
>> +
>> +    gt_fd = xe_sysfs_gt_open(fd, gt);
>> +    igt_assert(gt_fd >= 0);
>> +    igt_assert(igt_sysfs_scanf(gt_fd, "gtidle/idle_residency_ms", 
>> "%lu", &residency) == 1);
>> +    close(gt_fd);
>> +
>> +    return residency;
>> +}
>> +
>> +static u64 get_event_config(int xe, unsigned int gt, char *event)
>> +{
>> +    int ret;
>> +    char xe_device[100];
>> +    u64 pmu_config;
>> +    u32 start, end;
>> +
>> +    xe_perf_device(xe, xe_device, sizeof(xe_device));
>> +    ret = perf_event_config(xe_device, event, &pmu_config);
>> +    igt_assert(ret >= 0);
>> +    ret = perf_event_format(xe_device, "gt", &start, &end);
>> +    igt_assert(ret >= 0);
>> +    pmu_config |= (u64) gt << start;
>> +
>> +    return pmu_config;
>> +}
>> +
>> +static void test_gt_c6(int xe, unsigned int gt)
>> +{
>> +    int pmu_fd;
>> +    u64 pmu_config;
>> +    char event[100];
>> +    uint64_t ts[2];
>> +    unsigned long slept, start, end;
>> +    uint64_t val;
>> +
>> +    /* Get the PMU config for the gt-c6 event */
>> +    sprintf(event, "gt-c6-residency");
>> +    pmu_config = get_event_config(xe, gt, event);
>> +
>> +    pmu_fd = open_pmu(xe, pmu_config);
>> +
>> +    igt_require_f(igt_wait(xe_gt_is_in_c6(xe, gt), 1000, 10), "GT %d 
>> should be in C6\n", gt);
>> +
>> +    /* While idle check full RC6. */
>> +    start = read_idle_residency(xe, gt);
>> +    val = __pmu_read_single(pmu_fd, &ts[0]);
>> +    slept = measured_usleep(SLEEP_DURATION * USEC_PER_SEC) / 1000;
>> +    end = read_idle_residency(xe, gt);
>> +    val = __pmu_read_single(pmu_fd, &ts[1]) - val;
>
> I actually expected 2 subtests:
>
> 1) gt_c6_idle
>
>     we should be in rc6 for the entire time when nothing is running
>
> 2) gt_c6_full_load
>
>     we should not be in rc6 when something is keeping the engine
>     busy
>
> For (2), we can start a work load with xe_cork
I can add a busy test later on, since I will be adding a workload 
related freq subtest after this.
>
>
>> +
>> +    igt_debug("gt%u: slept=%lu, perf=%"PRIu64"\n",
>> +          gt, slept,  val);
>> +
>> +    igt_debug("Start res: %lu, end_res: %lu", start, end);
>
> is this reading the value from sysfs to aid debugging?
yes.
>
>
> I tested this with the kernel implementation, before merging it, and it
> worked. I don't mind if this is merged and cleaned on top. I will leave
> it for you and Riana to decide.

Sent another version with most of the above comments addressed.

Thanks,

Vinay.

>
> thanks
> Lucas De Marchi
>
>> +
>> +    assert_within_epsilon(val,
>> +                  (ts[1] - ts[0])/USEC_PER_SEC,
>> +                  tolerance);
>> +    close(pmu_fd);
>> +}
>> +
>> +igt_main
>> +{
>> +    int fd, gt;
>> +
>> +    igt_fixture {
>> +        fd = drm_open_driver(DRIVER_XE);
>> +        igt_require(!IS_PONTEVECCHIO(xe_dev_id(fd)));
>> +    }
>> +
>> +    igt_describe("Validate PMU gt-c6 residency counters");
>> +    igt_subtest("gt-c6")
>> +        xe_for_each_gt(fd, gt)
>> +            test_gt_c6(fd, gt);
>> +
>> +    igt_fixture {
>> +        close(fd);
>> +    }
>> +}
>> diff --git a/tests/meson.build b/tests/meson.build
>> index 33dffad31..d20f50766 100644
>> --- a/tests/meson.build
>> +++ b/tests/meson.build
>> @@ -309,6 +309,7 @@ intel_xe_progs = [
>>     'xe_pat',
>>     'xe_peer2peer',
>>     'xe_pm',
>> +    'xe_pmu',
>>     'xe_pm_residency',
>>     'xe_prime_self_import',
>>     'xe_query',
>> -- 
>> 2.38.1
>>
diff mbox series

Patch

diff --git a/tests/intel/xe_pmu.c b/tests/intel/xe_pmu.c
new file mode 100644
index 000000000..61a7a7e38
--- /dev/null
+++ b/tests/intel/xe_pmu.c
@@ -0,0 +1,176 @@ 
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+/**
+ * TEST: Test Xe PMU functionality
+ * Category: Perf Monitoring Unit
+ * Mega feature: Perf Monitoring Unit
+ * Sub-category: Power Management
+ * Functionality: Power/Perf
+ * Test category: Functional tests
+ */
+
+#include <fcntl.h>
+#include <limits.h>
+#include <time.h>
+#include <errno.h>
+#include <dirent.h>
+#include <string.h>
+#include <sys/time.h>
+
+#include "igt.h"
+#include "igt_device.h"
+#include "igt_power.h"
+#include "igt_sysfs.h"
+#include "igt_perf.h"
+
+#include "lib/igt_syncobj.h"
+#include "xe/xe_ioctl.h"
+#include "xe/xe_gt.h"
+#include "xe/xe_query.h"
+#include "xe/xe_spin.h"
+#include "xe/xe_util.h"
+
+#define SLEEP_DURATION 2 /* in seconds */
+const double tolerance = 0.1;
+const unsigned long batch_duration_ns = 500e6;
+const char *no_debug_data = "\0";
+
+#define __assert_within_epsilon(x, ref, tol_up, tol_down, debug_data) \
+	igt_assert_f((double)(x) <= (1.0 + (tol_up)) * (double)(ref) && \
+		     (double)(x) >= (1.0 - (tol_down)) * (double)(ref), \
+		     "'%s' != '%s' (%f not within +%.1f%%/-%.1f%% tolerance of %f)\n%s\n",\
+		     #x, #ref, (double)(x), \
+		     (tol_up) * 100.0, (tol_down) * 100.0, \
+		     (double)(ref), debug_data)
+
+#define assert_within_epsilon(x, ref, tolerance) \
+	__assert_within_epsilon(x, ref, tolerance, tolerance, no_debug_data)
+
+static int open_pmu(int xe, uint64_t config)
+{
+	int fd;
+
+	fd = perf_xe_open(xe, config);
+	igt_skip_on(fd < 0 && errno == ENODEV);
+	igt_assert(fd >= 0);
+
+	return fd;
+}
+
+static uint64_t __pmu_read_single(int fd, uint64_t *ts)
+{
+	uint64_t data[2];
+
+	igt_assert_eq(read(fd, data, sizeof(data)), sizeof(data));
+	if (ts)
+		*ts = data[1];
+
+	return data[0];
+}
+
+/**
+ * SUBTEST: gt-c6
+ * Description: Basic residency test to validate idle residency
+ *		measured over a time interval is within the tolerance
+ */
+static unsigned int measured_usleep(unsigned int usec)
+{
+	struct timespec ts = { };
+	unsigned int slept;
+
+	slept = igt_nsec_elapsed(&ts);
+	igt_assert(slept == 0);
+	do {
+		usleep(usec - slept);
+		slept = igt_nsec_elapsed(&ts) / 1000;
+	} while (slept < usec);
+
+	return igt_nsec_elapsed(&ts) / 1000;
+}
+
+static unsigned long read_idle_residency(int fd, int gt)
+{
+	unsigned long residency = 0;
+	int gt_fd;
+
+	gt_fd = xe_sysfs_gt_open(fd, gt);
+	igt_assert(gt_fd >= 0);
+	igt_assert(igt_sysfs_scanf(gt_fd, "gtidle/idle_residency_ms", "%lu", &residency) == 1);
+	close(gt_fd);
+
+	return residency;
+}
+
+static u64 get_event_config(int xe, unsigned int gt, char *event)
+{
+	int ret;
+	char xe_device[100];
+	u64 pmu_config;
+	u32 start, end;
+
+	xe_perf_device(xe, xe_device, sizeof(xe_device));
+	ret = perf_event_config(xe_device, event, &pmu_config);
+	igt_assert(ret >= 0);
+	ret = perf_event_format(xe_device, "gt", &start, &end);
+	igt_assert(ret >= 0);
+	pmu_config |= (u64) gt << start;
+
+	return pmu_config;
+}
+
+static void test_gt_c6(int xe, unsigned int gt)
+{
+	int pmu_fd;
+	u64 pmu_config;
+	char event[100];
+	uint64_t ts[2];
+	unsigned long slept, start, end;
+	uint64_t val;
+
+	/* Get the PMU config for the gt-c6 event */
+	sprintf(event, "gt-c6-residency");
+	pmu_config = get_event_config(xe, gt, event);
+
+	pmu_fd = open_pmu(xe, pmu_config);
+
+	igt_require_f(igt_wait(xe_gt_is_in_c6(xe, gt), 1000, 10), "GT %d should be in C6\n", gt);
+
+	/* While idle check full RC6. */
+	start = read_idle_residency(xe, gt);
+	val = __pmu_read_single(pmu_fd, &ts[0]);
+	slept = measured_usleep(SLEEP_DURATION * USEC_PER_SEC) / 1000;
+	end = read_idle_residency(xe, gt);
+	val = __pmu_read_single(pmu_fd, &ts[1]) - val;
+
+	igt_debug("gt%u: slept=%lu, perf=%"PRIu64"\n",
+		  gt, slept,  val);
+
+	igt_debug("Start res: %lu, end_res: %lu", start, end);
+
+	assert_within_epsilon(val,
+			      (ts[1] - ts[0])/USEC_PER_SEC,
+			      tolerance);
+	close(pmu_fd);
+}
+
+igt_main
+{
+	int fd, gt;
+
+	igt_fixture {
+		fd = drm_open_driver(DRIVER_XE);
+		igt_require(!IS_PONTEVECCHIO(xe_dev_id(fd)));
+	}
+
+	igt_describe("Validate PMU gt-c6 residency counters");
+	igt_subtest("gt-c6")
+		xe_for_each_gt(fd, gt)
+			test_gt_c6(fd, gt);
+
+	igt_fixture {
+		close(fd);
+	}
+}
diff --git a/tests/meson.build b/tests/meson.build
index 33dffad31..d20f50766 100644
--- a/tests/meson.build
+++ b/tests/meson.build
@@ -309,6 +309,7 @@  intel_xe_progs = [
 	'xe_pat',
 	'xe_peer2peer',
 	'xe_pm',
+	'xe_pmu',
 	'xe_pm_residency',
 	'xe_prime_self_import',
 	'xe_query',