Struct TensorProfiler 

pub struct TensorProfiler { /* private fields */ }

Operation profiler for tensor computations.

Records individual operation entries and maintains running aggregates per operation name. Supports enable/disable toggling, configurable maximum entry count (oldest entries evicted on overflow), and tick-based ordering.

Implementations

impl TensorProfiler

pub fn new(max_entries: usize) -> Self

Create a new profiler with the given maximum entry capacity.

The profiler starts enabled by default.

pub fn record( &mut self, op_name: &str, duration_ns: u64, input_elements: u64, output_elements: u64, )

Record an operation.

If the profiler is disabled the call is a no-op. When the entry buffer is full the oldest entry is evicted (FIFO).

pub fn enable(&mut self)

Enable recording.

pub fn disable(&mut self)

Disable recording. Subsequent record calls become no-ops.

pub fn is_enabled(&self) -> bool

Returns true if the profiler is currently recording.

pub fn get_profile(&self, op_name: &str) -> Option<&OpProfile>

Look up the aggregate profile for a given operation name.

pub fn avg_ns(&self, op_name: &str) -> Option<f64>

Average duration per call in nanoseconds for the given op.

pub fn throughput(&self, op_name: &str) -> Option<f64>

Throughput in elements/second for the given op.

Computed as total_output_elements / (total_ns / 1e9).

pub fn hottest_ops(&self, n: usize) -> Vec<&OpProfile>

Return the top n operations by cumulative time (descending).

pub fn tick(&mut self)

Advance the internal tick counter by one.

pub fn reset(&mut self)

Clear all entries and aggregate profiles.

pub fn entry_count(&self) -> usize

Number of entries currently stored.

pub fn stats(&self) -> ProfilerStats

Return a snapshot of summary statistics.