CosineSimilarity code Explain

[mtsol]

Hello team, I need to know the function call at line 34 in https://github.com/NVIDIA/spark-rapids-examples/blob/branch-22.12/examples/UDF-Examples/RAPIDS-accelerated-UDFs/src/main/java/com/nvidia/spark/rapids/udf/java/CosineSimilarity.java executes on GPU or CPU? And if it executes on CPU then where is the GPU code?

[jlowe]

Line 34 provides the CPU implementation of the UDF. Per the RAPIDS accelerated UDF documentation the UDF provides two implementations, the existing one for the CPU and a new one for the GPU. The GPU implementation of the UDF is provided by the evaluateColumnar interface, which starts at line 64 in that same file. This method ensures the native library is loaded and then proceeds to evaluate it on the GPU via a JNI method.

The JNI method source is at https://github.com/NVIDIA/spark-rapids-examples/blob/branch-22.12/examples/UDF-Examples/RAPIDS-accelerated-UDFs/src/main/cpp/src/CosineSimilarityJni.cpp#L60-L92 and the source of the GPU code invoked by that JNI method is at https://github.com/NVIDIA/spark-rapids-examples/blob/branch-22.12/examples/UDF-Examples/RAPIDS-accelerated-UDFs/src/main/cpp/src/cosine_similarity.cu.