Skip to content

perfetto: Add InstrumentedSamplingConfig to GpuCounterConfig #5358

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
dreveman wants to merge 1 commit into
base: main
Choose a base branch
from
Open

perfetto: Add InstrumentedSamplingConfig to GpuCounterConfig #5358

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
56 changes: 56 additions & 0 deletions docs/data-sources/gpu.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,6 +87,12 @@ data_sources: {

`counter_period_ns` sets the desired sampling interval.

Alternatively, counters can be selected by name using `counter_names`. Use one
or the other, not both. Not all producers support this — check
`supports_counter_names` in the `GpuCounterDescriptor` data source descriptor.
Some producers may also support glob patterns in `counter_names` for matching
multiple counters but this is not guaranteed.

### GPU memory

Total GPU memory usage per process is collected via ftrace:
Expand Down Expand Up @@ -167,6 +173,56 @@ data_sources: {
}
```

For more control over which GPU activities are instrumented, use
`instrumented_sampling_config` instead of the `instrumented_sampling` bool.
This enables a pipeline of filters applied in the following order:

1. **Activity name filtering**: If `activity_name_filters` is non-empty, the
activity must match at least one filter. Each filter requires a `name_glob`
pattern and an optional `name_base` (defaults to `MANGLED_KERNEL_NAME` if
not specified). If empty, all activities pass this step.

2. **TX range filtering**: If `activity_tx_include_globs` is non-empty, the
activity must fall within a TX range (e.g. NVTX range for CUDA) matching
one of the include globs. Activities in TX ranges matching
`activity_tx_exclude_globs` are excluded (excludes take precedence over
includes). TX ranges can be nested, and an activity matches if any range
in its nesting hierarchy matches. If both are empty, all activities pass
this step.

3. **Range-based sampling**: If `activity_ranges` is non-empty, only
activities within the specified skip/count ranges are instrumented.
`skip` defaults to 0 and `count` defaults to UINT32\_MAX (all remaining
activities) when not specified. If empty, all activities that passed the
previous steps are instrumented.

Example configuration that instruments only activities with demangled kernel
names matching `"myKernel*"` within TX ranges matching `"training*"`,
skipping the first 10 matching activities and then instrumenting 5:

```
data_sources: {
config {
name: "gpu.counters"
gpu_counter_config {
counter_names: "sm__cycles_elapsed.avg"
counter_names: "sm__cycles_active.avg"
instrumented_sampling_config {
activity_name_filters {
name_glob: "myKernel*"
name_base: DEMANGLED_KERNEL_NAME
}
activity_tx_include_globs: "training*"
activity_ranges {
skip: 10
count: 5
}
}
}
}
}
```

Counter descriptor mode 2 is recommended for GPGPU use-cases: the producer
emits an `InternedGpuCounterDescriptor` referenced by IID, giving each
trusted sequence its own scoped counter IDs. This avoids the global
Expand Down
5 changes: 5 additions & 0 deletions protos/perfetto/common/gpu_counter_descriptor.proto
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,6 +82,11 @@ message GpuCounterDescriptor {
// command buffer.
optional bool supports_instrumented_sampling = 5;

// optional. The producer supports selecting counters by name via
// GpuCounterConfig.counter_names. Not all producers support this; Android
// GPU producers typically do not.
optional bool supports_counter_names = 6;

// next id: 41
enum MeasureUnit {
NONE = 0;
Expand Down
88 changes: 85 additions & 3 deletions protos/perfetto/config/gpu/gpu_counter_config.proto
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,12 +22,94 @@ message GpuCounterConfig {
// Desired sampling interval for counters.
optional uint64 counter_period_ns = 1;

// List of counters to be sampled. Counter IDs correspond to the ones
// described in GpuCounterSpec in the data source descriptor.
// Selects which counters to sample. Use either counter_ids or counter_names,
// not both. Counter IDs and names correspond to the ones described in
// GpuCounterSpec in the data source descriptor.

// List of counter IDs to be sampled.
repeated uint32 counter_ids = 2;

// List of counter names to be sampled. Requires producer support; check
// GpuCounterDescriptor.supports_counter_names in the data source descriptor.
// Some producers may also support glob patterns for matching multiple
// counters by name but this is not guaranteed.
repeated string counter_names = 6;

// Configuration for sampling counters by instrumenting command buffers.
//
// When instrumented_sampling_config is used (instead of the
// instrumented_sampling bool), the following steps determine whether
// instrumented counters are enabled for a given GPU activity:
//
// 1. Activity name filtering: If activity_name_filters is non-empty, the
// activity must match at least one filter. If empty, all activities
// pass this step.
// 2. TX range filtering: If activity_tx_include_globs is non-empty, the
// activity must fall within a matching TX range. Activities in TX
// ranges matching activity_tx_exclude_globs are excluded (excludes
// take precedence over includes). If both are empty, all activities
// pass this step.
// 3. Range-based sampling: If activity_ranges is non-empty, only
// activities within the specified skip/count ranges are instrumented.
// If empty, all activities that passed the previous steps are
// instrumented.
message InstrumentedSamplingConfig {
// Filters GPU activities by name. Each filter specifies a glob pattern
// and the basis for matching (mangled or demangled kernel name).
message ActivityNameFilter {
enum NameBase {
MANGLED_KERNEL_NAME = 0;
DEMANGLED_KERNEL_NAME = 1;
}

// required. Glob pattern to use for GPU activity name filtering.
optional string name_glob = 1;

// Basis for name filtering. Defaults to MANGLED_KERNEL_NAME if not
// specified.
optional NameBase name_base = 2;
}

// GPU activity name filters. An activity matches if it matches any filter.
repeated ActivityNameFilter activity_name_filters = 3;

// Glob patterns to use for including GPU activities in TX ranges. TX
// ranges are in-process annotations that mark different sections of GPU
// work (e.g. NVTX ranges for CUDA). TX ranges can be nested, and an
// activity is included if any range in its nesting hierarchy matches.
// Only activities that fall within a matching TX range will be
// instrumented.
repeated string activity_tx_include_globs = 6;

// Glob patterns to use for excluding GPU activities from TX ranges.
// TX ranges can be nested, and an activity is excluded if any range
// in its nesting hierarchy matches. Excludes take precedence over
// includes.
repeated string activity_tx_exclude_globs = 7;

// Defines a range of GPU activities to instrument.
message ActivityRange {
// Number of GPU activities to skip before starting to instrument
// command buffers. Defaults to 0 if not specified.
optional uint32 skip = 1;

// Limit for the number of GPU activities to sample counters for by
// instrumenting command buffers. Defaults to UINT32_MAX (all
// remaining activities) if not specified.
optional uint32 count = 2;
}

// Ranges of GPU activities to instrument. Applied after activity name
// and TX range filters. If empty, all activities that passed the
// previous filters are instrumented.
repeated ActivityRange activity_ranges = 5;
}

// Sample counters by instrumenting command buffers.
optional bool instrumented_sampling = 3;
oneof instrumented_sampling_mode {
bool instrumented_sampling = 3;
InstrumentedSamplingConfig instrumented_sampling_config = 5;
}

// Fix gpu clock rate during trace session.
optional bool fix_gpu_clock = 4;
Expand Down
93 changes: 90 additions & 3 deletions protos/perfetto/config/perfetto_config.proto
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,6 +95,11 @@ message GpuCounterDescriptor {
// command buffer.
optional bool supports_instrumented_sampling = 5;

// optional. The producer supports selecting counters by name via
// GpuCounterConfig.counter_names. Not all producers support this; Android
// GPU producers typically do not.
optional bool supports_counter_names = 6;

// next id: 41
enum MeasureUnit {
NONE = 0;
Expand Down Expand Up @@ -1651,12 +1656,94 @@ message GpuCounterConfig {
// Desired sampling interval for counters.
optional uint64 counter_period_ns = 1;

// List of counters to be sampled. Counter IDs correspond to the ones
// described in GpuCounterSpec in the data source descriptor.
// Selects which counters to sample. Use either counter_ids or counter_names,
// not both. Counter IDs and names correspond to the ones described in
// GpuCounterSpec in the data source descriptor.

// List of counter IDs to be sampled.
repeated uint32 counter_ids = 2;

// List of counter names to be sampled. Requires producer support; check
// GpuCounterDescriptor.supports_counter_names in the data source descriptor.
// Some producers may also support glob patterns for matching multiple
// counters by name but this is not guaranteed.
repeated string counter_names = 6;

// Configuration for sampling counters by instrumenting command buffers.
//
// When instrumented_sampling_config is used (instead of the
// instrumented_sampling bool), the following steps determine whether
// instrumented counters are enabled for a given GPU activity:
//
// 1. Activity name filtering: If activity_name_filters is non-empty, the
// activity must match at least one filter. If empty, all activities
// pass this step.
// 2. TX range filtering: If activity_tx_include_globs is non-empty, the
// activity must fall within a matching TX range. Activities in TX
// ranges matching activity_tx_exclude_globs are excluded (excludes
// take precedence over includes). If both are empty, all activities
// pass this step.
// 3. Range-based sampling: If activity_ranges is non-empty, only
// activities within the specified skip/count ranges are instrumented.
// If empty, all activities that passed the previous steps are
// instrumented.
message InstrumentedSamplingConfig {
// Filters GPU activities by name. Each filter specifies a glob pattern
// and the basis for matching (mangled or demangled kernel name).
message ActivityNameFilter {
enum NameBase {
MANGLED_KERNEL_NAME = 0;
DEMANGLED_KERNEL_NAME = 1;
}

// required. Glob pattern to use for GPU activity name filtering.
optional string name_glob = 1;

// Basis for name filtering. Defaults to MANGLED_KERNEL_NAME if not
// specified.
optional NameBase name_base = 2;
}

// GPU activity name filters. An activity matches if it matches any filter.
repeated ActivityNameFilter activity_name_filters = 3;

// Glob patterns to use for including GPU activities in TX ranges. TX
// ranges are in-process annotations that mark different sections of GPU
// work (e.g. NVTX ranges for CUDA). TX ranges can be nested, and an
// activity is included if any range in its nesting hierarchy matches.
// Only activities that fall within a matching TX range will be
// instrumented.
repeated string activity_tx_include_globs = 6;

// Glob patterns to use for excluding GPU activities from TX ranges.
// TX ranges can be nested, and an activity is excluded if any range
// in its nesting hierarchy matches. Excludes take precedence over
// includes.
repeated string activity_tx_exclude_globs = 7;

// Defines a range of GPU activities to instrument.
message ActivityRange {
// Number of GPU activities to skip before starting to instrument
// command buffers. Defaults to 0 if not specified.
optional uint32 skip = 1;

// Limit for the number of GPU activities to sample counters for by
// instrumenting command buffers. Defaults to UINT32_MAX (all
// remaining activities) if not specified.
optional uint32 count = 2;
}

// Ranges of GPU activities to instrument. Applied after activity name
// and TX range filters. If empty, all activities that passed the
// previous filters are instrumented.
repeated ActivityRange activity_ranges = 5;
}

// Sample counters by instrumenting command buffers.
optional bool instrumented_sampling = 3;
oneof instrumented_sampling_mode {
bool instrumented_sampling = 3;
InstrumentedSamplingConfig instrumented_sampling_config = 5;
}

// Fix gpu clock rate during trace session.
optional bool fix_gpu_clock = 4;
Expand Down
93 changes: 90 additions & 3 deletions protos/perfetto/trace/perfetto_trace.proto
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,6 +95,11 @@ message GpuCounterDescriptor {
// command buffer.
optional bool supports_instrumented_sampling = 5;

// optional. The producer supports selecting counters by name via
// GpuCounterConfig.counter_names. Not all producers support this; Android
// GPU producers typically do not.
optional bool supports_counter_names = 6;

// next id: 41
enum MeasureUnit {
NONE = 0;
Expand Down Expand Up @@ -1651,12 +1656,94 @@ message GpuCounterConfig {
// Desired sampling interval for counters.
optional uint64 counter_period_ns = 1;

// List of counters to be sampled. Counter IDs correspond to the ones
// described in GpuCounterSpec in the data source descriptor.
// Selects which counters to sample. Use either counter_ids or counter_names,
// not both. Counter IDs and names correspond to the ones described in
// GpuCounterSpec in the data source descriptor.

// List of counter IDs to be sampled.
repeated uint32 counter_ids = 2;

// List of counter names to be sampled. Requires producer support; check
// GpuCounterDescriptor.supports_counter_names in the data source descriptor.
// Some producers may also support glob patterns for matching multiple
// counters by name but this is not guaranteed.
repeated string counter_names = 6;

// Configuration for sampling counters by instrumenting command buffers.
//
// When instrumented_sampling_config is used (instead of the
// instrumented_sampling bool), the following steps determine whether
// instrumented counters are enabled for a given GPU activity:
//
// 1. Activity name filtering: If activity_name_filters is non-empty, the
// activity must match at least one filter. If empty, all activities
// pass this step.
// 2. TX range filtering: If activity_tx_include_globs is non-empty, the
// activity must fall within a matching TX range. Activities in TX
// ranges matching activity_tx_exclude_globs are excluded (excludes
// take precedence over includes). If both are empty, all activities
// pass this step.
// 3. Range-based sampling: If activity_ranges is non-empty, only
// activities within the specified skip/count ranges are instrumented.
// If empty, all activities that passed the previous steps are
// instrumented.
message InstrumentedSamplingConfig {
// Filters GPU activities by name. Each filter specifies a glob pattern
// and the basis for matching (mangled or demangled kernel name).
message ActivityNameFilter {
enum NameBase {
MANGLED_KERNEL_NAME = 0;
DEMANGLED_KERNEL_NAME = 1;
}

// required. Glob pattern to use for GPU activity name filtering.
optional string name_glob = 1;

// Basis for name filtering. Defaults to MANGLED_KERNEL_NAME if not
// specified.
optional NameBase name_base = 2;
}

// GPU activity name filters. An activity matches if it matches any filter.
repeated ActivityNameFilter activity_name_filters = 3;

// Glob patterns to use for including GPU activities in TX ranges. TX
// ranges are in-process annotations that mark different sections of GPU
// work (e.g. NVTX ranges for CUDA). TX ranges can be nested, and an
// activity is included if any range in its nesting hierarchy matches.
// Only activities that fall within a matching TX range will be
// instrumented.
repeated string activity_tx_include_globs = 6;

// Glob patterns to use for excluding GPU activities from TX ranges.
// TX ranges can be nested, and an activity is excluded if any range
// in its nesting hierarchy matches. Excludes take precedence over
// includes.
repeated string activity_tx_exclude_globs = 7;

// Defines a range of GPU activities to instrument.
message ActivityRange {
// Number of GPU activities to skip before starting to instrument
// command buffers. Defaults to 0 if not specified.
optional uint32 skip = 1;

// Limit for the number of GPU activities to sample counters for by
// instrumenting command buffers. Defaults to UINT32_MAX (all
// remaining activities) if not specified.
optional uint32 count = 2;
}

// Ranges of GPU activities to instrument. Applied after activity name
// and TX range filters. If empty, all activities that passed the
// previous filters are instrumented.
repeated ActivityRange activity_ranges = 5;
}

// Sample counters by instrumenting command buffers.
optional bool instrumented_sampling = 3;
oneof instrumented_sampling_mode {
bool instrumented_sampling = 3;
InstrumentedSamplingConfig instrumented_sampling_config = 5;
}

// Fix gpu clock rate during trace session.
optional bool fix_gpu_clock = 4;
Expand Down
Loading