Struct custos_math::custos::OpenCL

pub struct OpenCL {
    pub kernel_cache: RefCell<KernelCacheCL>,
    pub cache: RefCell<Cache<OpenCL>>,
    pub inner: RefCell<CLDevice>,
    pub graph: RefCell<Graph>,
    pub cpu: CPU,
}

Used to perform calculations with an OpenCL capable device. To make new calculations invocable, a trait providing new operations should be implemented for CLDevice.

Example

use custos::{OpenCL, Read, Buffer, Error};

fn main() -> Result<(), Error> {
    let device = OpenCL::new(0)?;
     
    let a = Buffer::from((&device, [1.3; 25]));
    let out = device.read(&a);
     
    assert_eq!(out, vec![1.3; 5*5]);
    Ok(())
}

Fields

kernel_cache: RefCell<KernelCacheCL>

cache: RefCell<Cache<OpenCL>>

inner: RefCell<CLDevice>

graph: RefCell<Graph>

cpu: CPU

Implementations

impl OpenCL

pub fn new( device_idx: usize ) -> Result<OpenCL, Box<dyn Error + Sync + Send + 'static, Global>>

Returns an OpenCL at the specified device index.

Errors

• No device is found at the given device index
• some other OpenCL related errors

pub fn reset(&'static mut self)

Sets the values of the attributes cache, kernel cache, graph and CPU to their default. This cleans up any accumulated allocations.

pub fn ctx(&self) -> Ref<'_, Context>

pub fn queue(&self) -> Ref<'_, CommandQueue>

pub fn device(&self) -> CLIntDevice

pub fn global_mem_size_in_gb( &self ) -> Result<f64, Box<dyn Error + Sync + Send + 'static, Global>>

pub fn max_mem_alloc_in_gb( &self ) -> Result<f64, Box<dyn Error + Sync + Send + 'static, Global>>

pub fn name( &self ) -> Result<String, Box<dyn Error + Sync + Send + 'static, Global>>

pub fn version( &self ) -> Result<String, Box<dyn Error + Sync + Send + 'static, Global>>

pub fn unified_mem(&self) -> bool

Checks whether the device supports unified memory.

pub fn set_unified_mem(&self, unified_mem: bool)

👎Deprecated since 0.6.0: Use the environment variable ‘CUSTOS_USE_UNIFIED’ set to ‘true’, ‘false’ or ‘default’[=hardware dependent] instead.

Trait Implementations

impl<T: CDatatype + Float> ActivationOps<T, (), OpenCL> for OpenCL

fn sigmoid(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn sigmoid_grad(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn tanh(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn tanh_grad(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn relu(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn relu_mut(&self, x: &mut Matrix<'_, T, Self, ()>)

inplace

fn relu_grad(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn relu_grad_mut(&self, x: &mut Matrix<'_, T, Self, ()>)

inplace

impl<T: CDatatype> AdditionalOps<T, (), OpenCL> for OpenCL

fn adds(&self, lhs: &Matrix<'_, T, Self>, rhs: T) -> Matrix<'_, T, Self>

fn subs(&self, lhs: &Matrix<'_, T, Self, ()>, rhs: T) -> Matrix<'_, T, Self, ()>

fn muls(&self, lhs: &Matrix<'_, T, Self>, rhs: T) -> Matrix<'_, T, Self>

fn divs(&self, lhs: &Matrix<'_, T, Self>, rhs: T) -> Matrix<'_, T, Self>

impl<T, S> Alloc<'_, T, S> for OpenCLwhere S: Shape,

fn alloc(&self, len: usize, flag: AllocFlag) -> CLPtr<T>

Allocate memory on the implemented device.

fn with_slice(&self, data: &[T]) -> CLPtr<T>

Allocate new memory with data

impl<T: CDatatype> AssignOps<T, (), OpenCL> for OpenCL

fn add_assign(&self, lhs: &mut Buffer<'_, T, OpenCL>, rhs: &Buffer<'_, T, OpenCL>)

Add assign

fn sub_assign(&self, lhs: &mut Buffer<'_, T, OpenCL>, rhs: &Buffer<'_, T, OpenCL>)

fn mul_assign( &self, lhs: &mut Buffer<'_, T, Self, ()>, rhs: &Buffer<'_, T, Self, ()> )

impl<T: CDatatype> BaseOps<T, (), OpenCL> for OpenCL

fn add( &self, lhs: &Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

Element-wise addition

fn sub( &self, lhs: &Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

Element-wise subtraction

fn mul( &self, lhs: &Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

Element-wise multiplication

fn div( &self, lhs: &Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

Element-wise division

impl<'a, T> CacheBuf<'a, T, ()> for OpenCL

fn cached(&'a self, len: usize) -> Buffer<'a, T, OpenCL, ()>

Adds a buffer to the cache. Following calls will return this buffer, if the corresponding internal count matches with the id used in the cache.

impl CacheReturn for OpenCL

type CT = RawCL

fn cache(&self) -> RefMut<'_, Cache<OpenCL>>where OpenCL: RawConv,

Returns a device specific Cache.

impl<T> ClearBuf<T, OpenCL, ()> for OpenCLwhere T: CDatatype,

fn clear(&self, buf: &mut Buffer<'_, T, OpenCL, ()>)

Sets all elements of the matrix to zero.

impl<T: CDatatype> ClipOp<T, (), OpenCL> for OpenCL

fn clip(&self, x: &Matrix<'_, T, Self>, min: T, max: T) -> Matrix<'_, T, Self>

impl<'a, T> CloneBuf<'a, T, ()> for OpenCL

fn clone_buf( &'a self, buf: &Buffer<'a, T, OpenCL, ()> ) -> Buffer<'a, T, OpenCL, ()>

Creates a deep copy of the specified buffer.

impl<T: CDatatype> ColOp<T, OpenCL> for OpenCL

fn add_col( &self, lhs: &Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

fn sub_col( &self, lhs: &Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

fn div_col( &self, lhs: &Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

impl<T, R> CopySlice<T, R, OpenCL> for OpenCLwhere R: RangeBounds<usize>,

fn copy_slice( &self, buf: &Buffer<'_, T, OpenCL, ()>, range: R ) -> Buffer<'_, T, OpenCL, ()>

Copy a slice of the given buffer into a new buffer.

impl Debug for OpenCL

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Device for OpenCL

type Ptr = CLPtr<U>

type Cache = Cache<OpenCL>

fn new() -> Result<OpenCL, Box<dyn Error + Sync + Send + 'static, Global>>

fn retrieve<T, S>( &self, len: usize, add_node: impl AddGraph ) -> Buffer<'_, T, Self, S>where S: Shape, Self: for<'a> Alloc<'a, T, S>,

impl<T: CDatatype> DiagflatOp<T, OpenCL> for OpenCL

fn diagflat(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

impl<T: CDatatype> FnsOps<T, (), OpenCL> for OpenCL

fn exp(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn ln(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn neg(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn powf(&self, x: &Matrix<'_, T, Self>, rhs: T) -> Matrix<'_, T, Self>

fn powi(&self, x: &Matrix<'_, T, Self>, rhs: i32) -> Matrix<'_, T, Self>

impl<T: CDatatype> Gemm<T, (), (), (), OpenCL> for OpenCL

fn gemm( &self, lhs: &Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

impl GraphReturn for OpenCL

fn graph(&self) -> RefMut<'_, Graph>

impl<T: CDatatype> MaxOps<T, OpenCL> for OpenCL

fn max(&self, x: &Matrix<'_, T, Self>) -> T

fn max_rows(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn max_cols(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

impl<T: Float> RandOp<T, (), OpenCL> for OpenCL

fn rand(&self, x: &mut Buffer<'_, T, OpenCL>, lo: T, hi: T)

impl RawConv for OpenCL

fn construct<T, S>( ptr: &<OpenCL as Device>::Ptr<T, S>, len: usize, node: Node ) -> <OpenCL as CacheReturn>::CTwhere S: Shape,

fn destruct<T, S>( ct: &<OpenCL as CacheReturn>::CT, flag: AllocFlag ) -> (<OpenCL as Device>::Ptr<T, S>, Node)where S: Shape,

impl<T> Read<T, OpenCL, ()> for OpenCLwhere T: Clone + Default,

type Read = Vec<T, Global>

fn read<'a>( &self, buf: &'a Buffer<'_, T, OpenCL, ()> ) -> <OpenCL as Read<T, OpenCL, ()>>::Read<'a>

Read the data of the Buffer as type Read.

fn read_to_vec(&self, buf: &Buffer<'_, T, OpenCL, ()>) -> Vec<T, Global>

Read the data of a buffer into a vector

impl<T: CDatatype> RowOp<T, (), (), OpenCL> for OpenCL

fn add_row( &self, lhs: &Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

fn add_row_mut(&self, lhs: &mut Matrix<'_, T, Self>, rhs: &Matrix<'_, T, Self>)

impl<T: CDatatype> ScalarAssign<T, (), OpenCL> for OpenCL

fn adds_assign(&self, lhs: &mut Matrix<'_, T, Self>, rhs: T)

fn muls_assign(&self, lhs: &mut Matrix<'_, T, Self>, rhs: T)

fn divs_assign(&self, lhs: &mut Matrix<'_, T, Self>, rhs: T)

fn subs_assign(&self, lhs: &mut Matrix<'_, T, Self>, rhs: T)

impl<T: GenericBlas + MatrixMultiply + Float> SoftmaxOps<T, OpenCL> for OpenCL

fn softmax(&self, inputs: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

fn softmax_grad( &self, activated: &Matrix<'_, T, Self>, grads: &Matrix<'_, T, Self> ) -> Matrix<'_, T, Self>

impl StaticGPU for OpenCL

fn as_static() -> &'static OpenCL

impl<T: Number> SumOps<T, (), OpenCL> for OpenCL

fn sum(&self, x: &Matrix<'_, T, Self>) -> T

fn mean(&self, x: &Matrix<'_, T, Self>) -> T

impl<T: CDatatype> SumOverOps<T, (), (), OpenCL> for OpenCL

fn sum_rows<'a>(&'a self, x: &Matrix<'_, T, Self>) -> Matrix<'a, T, Self>

fn sum_cols(&self, x: &Matrix<'_, T, Self>) -> Matrix<'_, T, Self>

impl<T: CDatatype> TransposeOp<T, (), (), OpenCL> for OpenCL

fn transpose(&self, x: &Matrix<'_, T, OpenCL>) -> Matrix<'_, T, OpenCL>

impl<T> WriteBuf<T, OpenCL, ()> for OpenCL

fn write(&self, buf: &mut Buffer<'_, T, OpenCL, ()>, data: &[T])

Write data to the buffer.

fn write_buf( &self, _dst: &mut Buffer<'_, T, Self, S>, _src: &Buffer<'_, T, Self, S> )

Writes data from Buffer to other Buffer.