◆ categorical_logit_glm()

__kernel void stan::math::opencl_kernels::categorical_logit_glm	(	__global double *	logp_global,
		__global double *	exp_lin_global,
		__global double *	inv_sum_exp_lin_global,
		__global double *	neg_softmax_lin_global,
		__global double *	alpha_derivative,
		const __global int *	y_global,
		const __global double *	x_beta_global,
		const __global double *	alpha_global,
		const int	N_instances,
		const int	N_attributes,
		const int	N_classes,
		const int	is_y_vector,
		const int	need_alpha_derivative,
		const int	need_neg_softmax_lin_global
	)

GPU implementation of Generalized Linear Model (GLM) with categorical distribution and logit (softmax) link function.

Must be run with at least N threads and local size equal to LOCAL_SIZE_.

Parameters

[out]	logp_global	partially summed log probability (1 value per work group)
[out]	exp_lin_global	exponentiation of sum of alpha and matrix product of x and beta
[out]	inv_sum_exp_lin_global	inverse of rowwise sum of `exp_lin_global`
[out]	neg_softmax_lin_global	negated softmax of sum of alpha and matrix product of x and beta
[out]	alpha_derivative	derivative wrt alpha
[in]	y_global	a scalar or vector of classes.
[in]	x_beta_global	product of design matrix and weight matrix
[in]	alpha_global	intercept (in log odds)
	N_instances	number of instances
	N_attributes	number of attributes
	N_classes	number of classes
	is_y_vector	0 or 1 - whether y is a vector (alternatively it is a scalar)
	need_alpha_derivative	interpreted as boolean - whether alpha_derivative needs to be computed
	need_neg_softmax_lin_global	interpreted as boolean - whether neg_softmax_lin_global needs to be computed