The OpenCL backend test for ConvolveStrided in convolve.cpp fails for half (fp16) data type due to NaN appearing in the array which is a result of the matmul operation done upstream. The library used for matrix multiplication is clblast which may have this related issue.

Description

• Built Arrayfire commit eaa49ca
• Which backend is experiencing this issue? OpenCL
• Do you have a workaround?: Forcing internatlly cast to float before the matrix operation resolves it
• Can the bug be reproduced reliably on your system? Yes

Reproducible Code and/or Steps

Run test_convolve_opencl. Reproducible with CPU, Nvidia, and Intel GPUs