indexing_rev() [2/2]

__kernel void stan::math::opencl_kernels::indexing_rev	(	__global double *	adj,
		const __global int *	index,
		const __global double *	res,
		__local double *	adj_loc,
		int	index_size,
		int	adj_size
	)

Increments adjoint of the indexing operation argument given the indices and adjoints of the indexing result.

This kernel uses local atomics and is the fastest for medium sizes of the indexed matrix (and will not work for large sizes).

Parameters

[in,out]	adj	adjoint to increment
	index
	res	adjoint of the result of indexing
	batch_size	Number of matrices in the batch.

Note

Code is a const char* held in add_batch_kernel_code.

This kernel makes each thread build its own copy of the adjoints before combining them. It is the fastest (and only works for) small size of the indexed matrix.

Parameters

[in,out]	adj	adjoint to increment
	index	int array index
	res	adjoint of the result of indexing
	adj_loc
	index_size	number of elements in index
	adj_size

Note

Code is a const char* held in add_batch_kernel_code.

Definition at line 71 of file indexing_rev.hpp.