Compute Follows Data

Compute follows data means computation on the device where data is placed. Assume we placed input data for an algorithm we want to run into GPU memory. The algorithm will be ran on GPU and resulting data will be placed into GPU memory as well as input data.

Actually data knows which device it is located on and which execution queue associated with the device. Based on the queue an algorithm is ran on the respective device.

Actually we don’t need to know execution queue. We just need to know device where we want to compute an algorithm. DPNP provides array constructors like dpnp.array that have such parameters as device, usm_type and sycl_queue. These parameters allow us to specify where to place data. Also DPNP provides list of functions that don’t have such parameters, e.g. dpnp.matmul. These functions take the parameters from input data.

Сonsider an example where we create two arrays in GPU memory and compute matrix product of the arrays on GPU.

1import dpnp
2
3x = dpnp.array([[1, 1], [1, 1]], device="gpu")
4y = dpnp.array([[1, 1], [1, 1]], device="gpu")
5
6res = dpnp.matmul(x, y)

Resulting array is placed in GPU memory as well as input arrays.

7res_device = res.get_array().sycl_device
8
9res_device.filter_string  # 'opencl:gpu:0'

Compute follows data prevents computation on data located on different devices. In other words compute follows data prevents implicit cross-device data copying. Copying makes computation expensive.

Consider an example where we create two arrays on different devices and try to compute matrix product of the arrays. As a result ValueError is raised.

10x = dpnp.array([[1, 1], [1, 1]], device="gpu")
11y = dpnp.array([[1, 1], [1, 1]], device="cpu")
12
13res = dpnp.matmul(x, y)  # ValueError: execution queue could not be determined ...

Execution queue couldn’t be determined, because x and y are located on different devices. To avoid ValueError we can do explicit copying of y to GPU memory before the function call.

14y = dpnp.asarray(y, device="gpu")
15
16res = dpnp.matmul(x, y)

For more information about available devices please refer to DPCtl’s devices documentation.