Struct custos_math::custos::CUDA

pub struct CUDA {
    pub cache: RefCell<Cache<CUDA>>,
    pub kernel_cache: RefCell<KernelCacheCU>,
    pub modules: RefCell<Vec<Module, Global>>,
    pub graph: RefCell<Graph>,
    /* private fields */
}

Used to perform calculations with a CUDA capable device. To make new calculations invocable, a trait providing new operations should be implemented for [CudaDevice].

Fields

cache: RefCell<Cache<CUDA>>

kernel_cache: RefCell<KernelCacheCU>

modules: RefCell<Vec<Module, Global>>

graph: RefCell<Graph>

Implementations

impl CUDA

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

pub fn device(&self) -> &CudaIntDevice

pub fn ctx(&self) -> &Context

pub fn handle(&self) -> &CublasHandle

pub fn stream(&self) -> &Stream

Trait Implementations

impl<T: CDatatype> ActivationOps<T, (), CUDA> for CUDA

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

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

inplace

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

inplace

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

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

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

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

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

impl<T> Alloc<'_, T, ()> for CUDA

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

Allocate memory on the implemented device.

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

Allocate new memory with data

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

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

Add assign

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

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

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

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 CUDA

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

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 CUDA

type CT = RawCUBuf

fn cache(&self) -> RefMut<'_, Cache<CUDA>>

Returns a device specific Cache.

impl<T> ClearBuf<T, CUDA, ()> for CUDAwhere T: CDatatype,

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

Sets all elements of the matrix to zero.

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

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

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

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

Creates a deep copy of the specified buffer.

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

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, CUDA> for CUDAwhere R: RangeBounds<usize>,

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

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

impl Debug for CUDA

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

Formats the value using the given formatter.

impl Device for CUDA

type Ptr = CUDAPtr<U>

type Cache = Cache<CUDA>

fn new() -> Result<CUDA, 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: Copy + Default> DiagflatOp<T, CUDA> for CUDA

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

impl Drop for CUDA

fn drop(&mut self)

Executes the destructor for this type.

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

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: GenericBlas> Gemm<T, (), (), (), CUDA> for CUDA

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

impl GraphReturn for CUDA

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

impl<T: Number> MaxOps<T, CUDA> for CUDA

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

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

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

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

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

impl RawConv for CUDA

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

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

impl<T> Read<T, CUDA, ()> for CUDAwhere T: Default + Clone,

type Read = Vec<T, Global>

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

Read the data of the Buffer as type Read.

fn read_to_vec(&self, buf: &Buffer<'_, T, CUDA, ()>) -> Vec<T, Global>where T: Default + Clone,

Read the data of a buffer into a vector

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

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

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

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

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

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

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

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

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

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 CUDA

fn as_static() -> &'static CUDA

impl<T: CDatatype> SumOps<T, (), CUDA> for CUDA

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

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

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

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

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

impl<T: CudaTranspose> TransposeOp<T, (), (), CUDA> for CUDA

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

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

fn write(&self, buf: &mut Buffer<'_, T, CUDA, ()>, 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.