Hands-On GPU Computing with Python

Explicit kernels

You can also use Numba to write kernels explicitly in NVIDIA CUDA or AMD ROCm (HSA).

In CUDA, we use syntax that we are already familiar with:

@cuda.jitdef multiply(p, q):    # Thread id in a 1D block    tx = cuda.threadIdx.x    # Block id in a 1D grid    ty = cuda.blockIdx.x    # Block width, i.e. number of threads per block    bw = cuda.blockDim.x    # Computing flattened index inside the array    index = tx + ty * bw    if index < N: # Check array size limit        q[index]=p[index]*q[index]

In ROCm, we use the following OpenCL syntax:

@roc.jitdef multiply(p, q):    # workitem id in a 1D workgroup    tx = roc.get_local_id(0)    # workgroup id in a 1D grid    ty = roc.get_group_id(0)    # workgroup size, i.e. number of workitems per workgroup bw = roc.get_local_size(0) ...