GPU Support
When available through an NVIDIA card with CUDA support, NumPower can store and perform operations on the graphics card, this is especially interesting for linear algebra operations on large tensors.
Copy NDArray to GPU
To copy an NDArray to your video card, just use the gpu() method. This method will return a new NDArray with
the data copied to your VRAM.
<?php
use \NDArray as nd;
$a = nd::ones([10, 10]);
$a_gpu = $a->gpu();
In this example we create an NDArray with format (10, 10) full of one. Then we use the gpu() method to create
a copy of this NDArray on the GPU.
Copy NDArray to CPU
In most cases, you will want your NDArray stored in your RAM. To copy an NDArray that is stored in your VRAM (GPU)
to your RAM (CPU), just use the cpu() method.
<?php
use \NDArray as nd;
$a = nd::ones([10, 10]);
$a_gpu = $a->gpu();
// OPERATIONS
$result = $a_gpu->cpu();
GPU and CPU operations
In operations involving more than one tensor, like NDArray::add, both tensors involved must be on the same device.
Operations between arrays on different devices will raise an exception.
In this first version of NumPower, we want the user to explicitly say where they want to store their data, so automatic copies between GPU and CPU are not available.
Some GPU-incompatible operations may raise an exception. In these cases the user must copy the tensor to the CPU manually using the appropriate method.