AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

intel/perf: Extend intel_perf_stream_set_metrics_id() to syncronize metrics id changes

Browse Source 
Xe KMD added a uAPI to syncronze metrics id changes, so we can make
it wait for all previous workloads in exec_queue and all previous
metrics id changes to finish before start change it again.
This should make Vulkan queries more robust.

So this makes use of intel_bind_timeline to syncronize the metrics id
changes and xe_queue_get_syncobj_for_idle() to syncronize with
exec_queue.

As i915 and some versions of Xe KMD will not support it, this feature
will only be used then intel_bind_timeline parameter is not NULL and
timeline has a valid syncobj id.
At this patch level all callers will set it to NULL, next patch will
add and initialize timeline in ANV when supported by Xe KMD.

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: José Roberto de Souza <jose.souza@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31283>
This commit is contained in:
José Roberto de Souza
2024-08-15 10:14:49 -07:00
committed by Marge Bot
parent 3e6546f662
commit a38a98c4cb
10 changed files with 131 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
src/intel/perf/intel_perf.c
View File

@@ -1569,7 +1569,7 @@ int
intel_perf_stream_open(struct intel_perf_config *perf_config, int drm_fd,
uint32_t ctx_id, uint64_t metrics_set_id,
uint64_t period_exponent, bool hold_preemption,
bool enable)
bool enable, struct intel_bind_timeline *timeline)
{
uint64_t report_format = intel_perf_get_oa_format(perf_config);
@@ -1581,7 +1581,7 @@ intel_perf_stream_open(struct intel_perf_config *perf_config, int drm_fd,
case INTEL_KMD_TYPE_XE:
return xe_perf_stream_open(perf_config, drm_fd, ctx_id, metrics_set_id,
report_format, period_exponent,
hold_preemption, enable);
hold_preemption, enable, timeline);
default:
unreachable("missing");
return 0;
@@ -1629,13 +1629,18 @@ intel_perf_stream_set_state(struct intel_perf_config *perf_config,
int
intel_perf_stream_set_metrics_id(struct intel_perf_config *perf_config,
int perf_stream_fd, uint64_t metrics_set_id)
int drm_fd, int perf_stream_fd,
uint32_t exec_queue,
uint64_t metrics_set_id,
struct intel_bind_timeline *timeline)
{
switch (perf_config->devinfo->kmd_type) {
case INTEL_KMD_TYPE_I915:
return i915_perf_stream_set_metrics_id(perf_stream_fd, metrics_set_id);
case INTEL_KMD_TYPE_XE:
return xe_perf_stream_set_metrics_id(perf_stream_fd, metrics_set_id);
return xe_perf_stream_set_metrics_id(perf_stream_fd, drm_fd,
exec_queue, metrics_set_id,
timeline);
default:
unreachable("missing");
return -1;

8
src/intel/perf/intel_perf.h
View File

@@ -30,6 +30,7 @@
#include <string.h>
#include "compiler/glsl/list.h"
#include "common/intel_bind_timeline.h"
#include "dev/intel_device_info.h"
#include "util/bitscan.h"
#include "util/bitset.h"
@@ -585,14 +586,17 @@ void intel_perf_get_counters_passes(struct intel_perf_config *perf,
int intel_perf_stream_open(struct intel_perf_config *perf_config, int drm_fd,
uint32_t ctx_id, uint64_t metrics_set_id,
uint64_t period_exponent, bool hold_preemption,
bool enable);
bool enable, struct intel_bind_timeline *timeline);
int intel_perf_stream_read_samples(struct intel_perf_config *perf_config,
int perf_stream_fd, uint8_t *buffer,
size_t buffer_len);
int intel_perf_stream_set_state(struct intel_perf_config *perf_config,
int perf_stream_fd, bool enable);
int intel_perf_stream_set_metrics_id(struct intel_perf_config *perf_config,
int perf_stream_fd, uint64_t metrics_set_id);
int drm_fd, int perf_stream_fd,
uint32_t exec_queue,
uint64_t metrics_set_id,
struct intel_bind_timeline *timeline);
#ifdef __cplusplus
} // extern "C"

2
src/intel/perf/intel_perf_query.c
View File

@@ -359,7 +359,7 @@ intel_perf_open(struct intel_perf_context *perf_ctx,
{
int fd = intel_perf_stream_open(perf_ctx->perf, drm_fd, ctx_id,
metrics_set_id, period_exponent, false,
enable);
enable, NULL);
if (fd == -1) {
DBG("Error opening gen perf OA stream: %m\n");
return false;

81
src/intel/perf/xe/intel_perf.c
View File

@@ -11,6 +11,7 @@
#include "perf/intel_perf.h"
#include "intel_perf_common.h"
#include "intel/common/intel_gem.h"
#include "intel/common/xe/intel_queue.h"
#include "drm-uapi/xe_drm.h"
@@ -136,7 +137,8 @@ int
xe_perf_stream_open(struct intel_perf_config *perf_config, int drm_fd,
uint32_t exec_id, uint64_t metrics_set_id,
uint64_t report_format, uint64_t period_exponent,
bool hold_preemption, bool enable)
bool hold_preemption, bool enable,
struct intel_bind_timeline *timeline)
{
struct drm_xe_ext_set_property props[DRM_XE_OA_PROPERTY_NO_PREEMPT + 1] = {};
struct drm_xe_observation_param observation_param = {
@@ -144,6 +146,10 @@ xe_perf_stream_open(struct intel_perf_config *perf_config, int drm_fd,
.observation_op = DRM_XE_OBSERVATION_OP_STREAM_OPEN,
.param = (uintptr_t)&props,
};
struct drm_xe_sync sync = {
.type = DRM_XE_SYNC_TYPE_TIMELINE_SYNCOBJ,
.flags = DRM_XE_SYNC_FLAG_SIGNAL,
};
uint32_t i = 0;
int fd, flags;
@@ -157,7 +163,18 @@ xe_perf_stream_open(struct intel_perf_config *perf_config, int drm_fd,
if (hold_preemption)
oa_prop_set(props, &i, DRM_XE_OA_PROPERTY_NO_PREEMPT, hold_preemption);
fd = intel_ioctl(drm_fd, DRM_IOCTL_XE_OBSERVATION, &observation_param);
if (timeline && intel_bind_timeline_get_syncobj(timeline)) {
oa_prop_set(props, &i, DRM_XE_OA_PROPERTY_NUM_SYNCS, 1);
oa_prop_set(props, &i, DRM_XE_OA_PROPERTY_SYNCS, (uintptr_t)&sync);
sync.handle = intel_bind_timeline_get_syncobj(timeline);
sync.timeline_value = intel_bind_timeline_bind_begin(timeline);
fd = intel_ioctl(drm_fd, DRM_IOCTL_XE_OBSERVATION, &observation_param);
intel_bind_timeline_bind_end(timeline);
} else {
fd = intel_ioctl(drm_fd, DRM_IOCTL_XE_OBSERVATION, &observation_param);
}
if (fd < 0)
return fd;
@@ -181,15 +198,63 @@ xe_perf_stream_set_state(int perf_stream_fd, bool enable)
}
int
xe_perf_stream_set_metrics_id(int perf_stream_fd, uint64_t metrics_set_id)
xe_perf_stream_set_metrics_id(int perf_stream_fd, int drm_fd,
uint32_t exec_queue, uint64_t metrics_set_id,
struct intel_bind_timeline *timeline)
{
struct drm_xe_ext_set_property prop = {};
struct drm_xe_ext_set_property prop[3] = {};
uint32_t index = 0;
int ret;
oa_prop_set(&prop, &index, DRM_XE_OA_PROPERTY_OA_METRIC_SET,
metrics_set_id);
return intel_ioctl(perf_stream_fd, DRM_XE_OBSERVATION_IOCTL_CONFIG,
(void *)(uintptr_t)&prop);
oa_prop_set(prop, &index, DRM_XE_OA_PROPERTY_OA_METRIC_SET,
metrics_set_id);
if (timeline && intel_bind_timeline_get_syncobj(timeline)) {
struct drm_xe_sync xe_syncs[3] = {};
uint32_t syncobj;
int ret2;
oa_prop_set(prop, &index, DRM_XE_OA_PROPERTY_NUM_SYNCS, ARRAY_SIZE(xe_syncs));
oa_prop_set(prop, &index, DRM_XE_OA_PROPERTY_SYNCS, (uintptr_t)xe_syncs);
/* wait on all previous exec in queues */
ret = xe_queue_get_syncobj_for_idle(drm_fd, exec_queue, &syncobj);
if (ret)
return ret;
xe_syncs[0].type = DRM_XE_SYNC_TYPE_SYNCOBJ;
xe_syncs[0].flags = 0;/* wait */
xe_syncs[0].handle = syncobj;
/* wait on previous set_metrics_id to complete */
xe_syncs[1].type = DRM_XE_SYNC_TYPE_TIMELINE_SYNCOBJ;
xe_syncs[1].flags = 0;/* wait */
xe_syncs[1].handle = intel_bind_timeline_get_syncobj(timeline);
xe_syncs[1].timeline_value = intel_bind_timeline_get_last_point(timeline);
/* signal completion */
xe_syncs[2].type = DRM_XE_SYNC_TYPE_TIMELINE_SYNCOBJ;
xe_syncs[2].flags = DRM_XE_SYNC_FLAG_SIGNAL;
xe_syncs[2].handle = intel_bind_timeline_get_syncobj(timeline);
xe_syncs[2].timeline_value = intel_bind_timeline_bind_begin(timeline);
ret = intel_ioctl(perf_stream_fd, DRM_XE_OBSERVATION_IOCTL_CONFIG,
(void *)(uintptr_t)&prop);
intel_bind_timeline_bind_end(timeline);
/* Looks safe to destroy as Xe KMD should increase the ref count until
* it is using it
*/
struct drm_syncobj_destroy syncobj_destroy = {
.handle = syncobj,
};
ret2 = intel_ioctl(drm_fd, DRM_IOCTL_SYNCOBJ_DESTROY, &syncobj_destroy);
assert(ret2 == 0);
} else {
ret = intel_ioctl(perf_stream_fd, DRM_XE_OBSERVATION_IOCTL_CONFIG,
(void *)(uintptr_t)&prop);
}
return ret;
}
static int

8
src/intel/perf/xe/intel_perf.h
View File

@@ -9,6 +9,7 @@
#include <stddef.h>
#include <stdint.h>
struct intel_bind_timeline;
struct intel_perf_config;
struct intel_perf_registers;
@@ -22,8 +23,11 @@ void xe_remove_config(struct intel_perf_config *perf, int fd, uint64_t config_id
int xe_perf_stream_open(struct intel_perf_config *perf_config, int drm_fd,
uint32_t exec_id, uint64_t metrics_set_id,
uint64_t report_format, uint64_t period_exponent,
bool hold_preemption, bool enable);
bool hold_preemption, bool enable,
struct intel_bind_timeline *timeline);
int xe_perf_stream_set_state(int perf_stream_fd, bool enable);
int xe_perf_stream_set_metrics_id(int perf_stream_fd, uint64_t metrics_set_id);
int xe_perf_stream_set_metrics_id(int perf_stream_fd, int drm_fd,
uint32_t exec_queue, uint64_t metrics_set_id,
struct intel_bind_timeline *timeline);
int xe_perf_stream_read_samples(struct intel_perf_config *perf_config, int perf_stream_fd,
uint8_t *buffer, size_t buffer_len);

26
src/intel/vulkan/anv_perf.c
View File

@@ -105,12 +105,11 @@ anv_device_perf_close(struct anv_device *device)
device->perf_fd = -1;
}
static int
anv_device_perf_open(struct anv_device *device, struct anv_queue *queue, uint64_t metric_id)
static uint32_t
anv_device_perf_get_queue_context_or_exec_queue_id(struct anv_queue *queue)
{
struct anv_device *device = queue->device;
uint32_t context_or_exec_queue_id;
uint64_t period_exponent = 31; /* slowest sampling period */
int ret;
switch (device->physical->info.kmd_type) {
case INTEL_KMD_TYPE_I915:
@@ -125,9 +124,18 @@ anv_device_perf_open(struct anv_device *device, struct anv_queue *queue, uint64_
context_or_exec_queue_id = 0;
}
return context_or_exec_queue_id;
}
static int
anv_device_perf_open(struct anv_device *device, struct anv_queue *queue, uint64_t metric_id)
{
uint64_t period_exponent = 31; /* slowest sampling period */
int ret;
ret = intel_perf_stream_open(device->physical->perf, device->fd,
context_or_exec_queue_id, metric_id,
period_exponent, true, true);
anv_device_perf_get_queue_context_or_exec_queue_id(queue),
metric_id, period_exponent, true, true, NULL);
if (ret >= 0)
device->perf_queue = queue;
@@ -276,9 +284,13 @@ VkResult anv_QueueSetPerformanceConfigurationINTEL(
if (device->perf_fd < 0)
return VK_ERROR_INITIALIZATION_FAILED;
} else {
uint32_t context_or_exec_queue = anv_device_perf_get_queue_context_or_exec_queue_id(device->perf_queue);
int ret = intel_perf_stream_set_metrics_id(device->physical->perf,
device->fd,
device->perf_fd,
config->config_id);
context_or_exec_queue,
config->config_id,
NULL);
if (ret < 0)
return vk_device_set_lost(&device->vk, "i915-perf config failed: %m");
}

5
src/intel/vulkan/i915/anv_batch_chain.c
View File

@@ -928,8 +928,11 @@ i915_queue_exec_locked(struct anv_queue *queue,
(query_info->kind == INTEL_PERF_QUERY_TYPE_OA ||
query_info->kind == INTEL_PERF_QUERY_TYPE_RAW)) {
int ret = intel_perf_stream_set_metrics_id(device->physical->perf,
device->fd,
device->perf_fd,
query_info->oa_metrics_set_id);
-1,/* this parameter, exec_queue is not used in i915 */
query_info->oa_metrics_set_id,
NULL);
if (ret < 0) {
result = vk_device_set_lost(&device->vk,
"i915-perf config failed: %s",

5
src/intel/vulkan/xe/anv_batch_chain.c
View File

@@ -334,8 +334,11 @@ xe_queue_exec_locked(struct anv_queue *queue,
(query_info->kind == INTEL_PERF_QUERY_TYPE_OA ||
query_info->kind == INTEL_PERF_QUERY_TYPE_RAW)) {
int ret = intel_perf_stream_set_metrics_id(device->physical->perf,
device->fd,
device->perf_fd,
query_info->oa_metrics_set_id);
queue->exec_queue_id,
query_info->oa_metrics_set_id,
NULL);
if (ret < 0) {
result = vk_device_set_lost(&device->vk,
"intel_perf_stream_set_metrics_id failed: %s",

5
src/intel/vulkan_hasvk/anv_batch_chain.c
View File

@@ -2222,8 +2222,11 @@ anv_queue_exec_locked(struct anv_queue *queue,
(query_info->kind == INTEL_PERF_QUERY_TYPE_OA ||
query_info->kind == INTEL_PERF_QUERY_TYPE_RAW)) {
int ret = intel_perf_stream_set_metrics_id(device->physical->perf,
device->fd,
device->perf_fd,
query_info->oa_metrics_set_id);
-1,/* this parameter, exec_queue is not used in i915 */
query_info->oa_metrics_set_id,
NULL);
if (ret < 0) {
result = vk_device_set_lost(&device->vk,
"i915-perf config failed: %s",

7
src/intel/vulkan_hasvk/anv_perf.c
View File

@@ -108,7 +108,7 @@ anv_device_perf_open(struct anv_device *device, uint64_t metric_id)
return intel_perf_stream_open(device->physical->perf, device->fd,
device->context_id, metric_id,
period_exponent, true, true);
period_exponent, true, true, NULL);
}
/* VK_INTEL_performance_query */
@@ -237,8 +237,11 @@ VkResult anv_QueueSetPerformanceConfigurationINTEL(
return VK_ERROR_INITIALIZATION_FAILED;
} else {
int ret = intel_perf_stream_set_metrics_id(device->physical->perf,
device->fd,
device->perf_fd,
config->config_id);
-1,/* this parameter, exec_queue is not used in i915 */
config->config_id,
NULL);
if (ret < 0)
return vk_device_set_lost(&device->vk, "i915-perf config failed: %m");
}