Skip to content

Latest commit

 

History

History

cute_vector_copy

README.md

CuTe Vector Copy

Introduction

These examples demonstrate the implementation of vector copy kernels using the CuTe. The general vector copy kernel does boundary checks and can be used for any matrix size. The vector copy vectorized kernel assumes the vector size is a multiple of certain size depending on the data type.

Usages

Run Unit Tests

$ ctest --test-dir build/ --tests-regex "TestAllVectorCopy.*" --verbose

The following tests passed:
        TestAllVectorCopy
        TestAllVectorCopyVectorized

100% tests passed, 0 tests failed out of 2

Run Performance Measurement

$ ctest --test-dir build/ --tests-regex "ProfileAllVectorCopy.*" --verbose

The following tables show the performance measurements of copying 1 GB floating point values using the vector copy kernels on NVIDIA GeForce RTX 3090.

Kernel Name Latency (ms) Effective Bandwidth (GB/s) Peak Bandwidth Percentage (%)
Vector Copy 2.81923 709.413 75.7842
Vector Copy Vectorized 2.83133 706.382 75.4604

Run Nsight Compute Profiling

for file in build/examples/cute_vector_copy/tests/profile_*; do
    filename=$(basename -- "$file")
    ncu --set full -f -o ncu_reports/"$filename" "$file"
done

Run Compute Sanitizer

for file in build/examples/cute_vector_copy/tests/test_*; do
    compute-sanitizer --leak-check full "$file"
done

References