Skip to main content

parpy.buffer

Functions​

sync(backend: parpy.CompileBackend)​

Synchronizes the CPU with the device of the selected backend. This function waits until all GPU kernels on the target device complete.

empty(shape: Tuple[int], dtype: parpy.ElemSize, backend: parpy.CompileBackend) -> Buffer​

Produces an empty Buffer with the specified shape shape containing values of data type dtype. The data is allocated on the backend selected using the backend argument.

empty_like(b: Buffer) -> Buffer​

Produces an empty Buffer of the same shape, data type, and backend as the provided buffer b.

zeros(shape: Tuple[int], dtype: parpy.ElemSize, backend: parpy.CompileBackend) -> Buffer​

Allocates a buffer in a similar vein to the empty function and sets all memory to zero.

zeros_like(b: Buffer) -> Buffer​

Allocates a buffer in a similar vein to the empty_like function and sets all memory to zero.

from_array(t, backend) -> Buffer​

Converts the provided array t to a Buffer of the selected backend using the array interface protocol or the CUDA array interface. If backend is set to None, the Buffer is constructed only to guide code generation and its pointer should not be used.

Classes​

Buffer​

sync(self)​

Synchronizes the contents of this buffer with the device, by delaying execution until all GPU kernels complete.

numpy(self) -> numpy.ndarray​

Converts the buffer to a NumPy array. Depending on the buffer backend, this may result in copying data from the GPU to the CPU, and the allocation of a new buffer for the resulting NumPy array.

torch_ref(self) -> torch.tensor​

Attempts to construct a PyTorch tensor referencing the data stored in the buffer. If the buffer backend is CUDA, this function produces a PyTorch tensor reusing the same data (this may be valuable for performance). Otherwise, it produces a new tensor containing a copy of the data in the original buffer.

reshape(self, *dims) -> Buffer​

Constructs a new view of the Buffer with a shape based on the provided integer dimensions dims. If the specified shape differs in size from the original shape, a ValueError is raised.

with_type(self, new_dtype) -> Buffer​

Constructs a new view of the Buffer with the provided data type dtype. This data type must either be of the internal data type parpy.DataType or one of the scalar types defined in the parpy.types module.