Question

We will mostly foucs on the use of CUDA Python via the numbapro compiler. Low level Python code using the numbapro.cuda module is similar to CUDA C, and will compile to the same machine code, but with the benefits of integerating into Python for use of numpy arrays, convenient I/O, graphics etc.

Optionally, CUDA Python can provide

• Automatic memory transfer
  • NumPy arrays are automatically transferred
  • CPU -> GPU
  • GPU -> CPU
• Automatic work scheduling
  • The work is distributed the across all threads on the GPU
  • The GPU hardware handles the scheduling
• Automatic GPU memory management
  • GPU memory is tied to object lifetime
  • freed automatically

but these can be over-riden with explicit control instructions if desired. Source

Python CUDA also provides syntactic sugar for obtaining thread identity. For example,

tx = cuda.threadIdx.x
ty = cuda.threadIdx.y
bx = cuda.blockIdx.x
by = cuda.blockIdx.y
bw = cuda.blockDim.x
bh = cuda.blockDim.y
x = tx + bx * bw
y = ty + by * bh
array[x, y] = something(x, y)

can be abbreivated to

x, y = cuda.grid(2)
array[x, y] = something(x, y)

Decorators are also provided for quick GPU parallelization, and it may be sufficient to use the high-level decorators jit , autojit , vectorize and guvectorize for running functoins on the GPU. When we need fine control, we can always drop back to CUDA Python.