Profile the Kernel Using CUpti profiling API

sequenceDiagram
    participant Device
    participant Session
    participant ConfigurationImage as Configuration Image
    participant CounterDataImage as Counter Data Image
    participant Event
    participant Counter
    participant Metric
    participant Pass
    participant Range



    note over ConfigurationImage: Configures Contain desire metric
    note over CounterDataImage: Store the collected metrics

    rect rgb(200, 150, 255) 
        note right of Device: Multiple Sessions
        Device->>Session: Allocates resources
    end


    Session->>CounterDataImage: Contains
    Session->>ConfigurationImage: Contains

    Event->>Counter: Counts occurrences
    Counter->>Metric: Used to calculate

    note over Range: Multiple kernels
    Pass->>Range: Consists of
    loop Multiple Replay
        Replay->>CounterDataImage: Per unique range-stack
    end

    Session->>Replay: Runs series of

CUDA_INJECTION64_PATH is set to a shared library
INJECTION_KERNEL_COUNT: This sets the number of kernels in a session
INJECTION_METRICS: This sets the metrics to gather, separated by space, comma,or semicolon. Default metrics are: sm__cycles_elapsed.avg gpu__time_duration.sum

#build example
bazel build //src:reduce 

#build shared lib
bazel build --config=rules_cuda //lib:injection_shared_using_rules_cuda 

#profile
env INJECTION_KERNEL_COUNT=2 CUDA_INJECTION64_PATH=./bazel-bin/lib/libinjection.so bazel-bin/src/reduce 16777216 256 256 1

Range Name	Metric Name	Metric Value (ns)
0	gpu__time_duration.sum	73920.000000
1	gpu__time_duration.sum	1472.000000
0	sm__cycles_elapsed.avg	185964.296875
1	sm__cycles_elapsed.avg	3650.703125

Lashi0812/CUptiProfiler

Profile the Kernel Using CUpti profiling API