|
| 1 | +#ifndef STAN_MATH_OPENCL_PRIM_DOUBLE_EXPONENTIAL_CDF_HPP |
| 2 | +#define STAN_MATH_OPENCL_PRIM_DOUBLE_EXPONENTIAL_CDF_HPP |
| 3 | +#ifdef STAN_OPENCL |
| 4 | + |
| 5 | +#include <stan/math/prim/meta.hpp> |
| 6 | +#include <stan/math/prim/err.hpp> |
| 7 | +#include <stan/math/prim/fun/constants.hpp> |
| 8 | +#include <stan/math/prim/fun/elt_divide.hpp> |
| 9 | +#include <stan/math/prim/fun/elt_multiply.hpp> |
| 10 | +#include <stan/math/opencl/kernel_generator.hpp> |
| 11 | +#include <stan/math/prim/functor/operands_and_partials.hpp> |
| 12 | + |
| 13 | +namespace stan { |
| 14 | +namespace math { |
| 15 | + |
| 16 | +/** \ingroup opencl |
| 17 | + * Returns the double exponential cumulative density function. Given |
| 18 | + * containers of matching sizes, returns the product of probabilities. |
| 19 | + * |
| 20 | + * @tparam T_y_cl type of scalar outcome |
| 21 | + * @tparam T_loc_cl type of location |
| 22 | + * @tparam T_scale_cl type of scale |
| 23 | + * @param y (Sequence of) scalar(s). |
| 24 | + * @param mu (Sequence of) location(s). |
| 25 | + * @param sigma (Sequence of) scale(s). |
| 26 | + * @return The log of the product of densities. |
| 27 | + */ |
| 28 | +template < |
| 29 | + typename T_y_cl, typename T_loc_cl, typename T_scale_cl, |
| 30 | + require_all_prim_or_rev_kernel_expression_t<T_y_cl, T_loc_cl, |
| 31 | + T_scale_cl>* = nullptr, |
| 32 | + require_any_not_stan_scalar_t<T_y_cl, T_loc_cl, T_scale_cl>* = nullptr> |
| 33 | +return_type_t<T_y_cl, T_loc_cl, T_scale_cl> double_exponential_cdf( |
| 34 | + const T_y_cl& y, const T_loc_cl& mu, const T_scale_cl& sigma) { |
| 35 | + static const char* function = "double_exponential_cdf(OpenCL)"; |
| 36 | + using T_partials_return = partials_return_t<T_y_cl, T_loc_cl, T_scale_cl>; |
| 37 | + using std::isfinite; |
| 38 | + using std::isnan; |
| 39 | + |
| 40 | + check_consistent_sizes(function, "Random variable", y, "Location parameter", |
| 41 | + mu, "Scale parameter", sigma); |
| 42 | + const size_t N = max_size(y, mu, sigma); |
| 43 | + if (N == 0) { |
| 44 | + return 1.0; |
| 45 | + } |
| 46 | + |
| 47 | + const auto& y_col = as_column_vector_or_scalar(y); |
| 48 | + const auto& mu_col = as_column_vector_or_scalar(mu); |
| 49 | + const auto& sigma_col = as_column_vector_or_scalar(sigma); |
| 50 | + |
| 51 | + const auto& y_val = value_of(y_col); |
| 52 | + const auto& mu_val = value_of(mu_col); |
| 53 | + const auto& sigma_val = value_of(sigma_col); |
| 54 | + |
| 55 | + auto check_y_not_nan |
| 56 | + = check_cl(function, "Random variable", y_val, "not NaN"); |
| 57 | + auto y_not_nan_expr = !isnan(y_val); |
| 58 | + auto check_mu_finite |
| 59 | + = check_cl(function, "Location parameter", mu_val, "finite"); |
| 60 | + auto mu_finite_expr = isfinite(mu_val); |
| 61 | + auto check_sigma_positive_finite |
| 62 | + = check_cl(function, "Scale parameter", sigma_val, "positive finite"); |
| 63 | + auto sigma_positive_finite_expr = 0 < sigma_val && isfinite(sigma_val); |
| 64 | + |
| 65 | + auto scaled_diff = elt_divide(y_val - mu_val, sigma_val); |
| 66 | + auto exp_scaled_diff = exp(scaled_diff); |
| 67 | + auto cond = y_val < mu_val; |
| 68 | + auto cdf_expr = colwise_prod(select(cond, 0.5 * exp_scaled_diff, |
| 69 | + 1.0 - elt_divide(0.5, exp_scaled_diff))); |
| 70 | + |
| 71 | + matrix_cl<double> cdf_cl; |
| 72 | + matrix_cl<double> mu_deriv_cl; // also temporarily stores exp_scaled_diff |
| 73 | + matrix_cl<double> y_deriv_cl; |
| 74 | + matrix_cl<double> sigma_deriv_cl; // also temporarily stores scaled_diff |
| 75 | + |
| 76 | + results(check_y_not_nan, check_mu_finite, check_sigma_positive_finite, cdf_cl, |
| 77 | + mu_deriv_cl, sigma_deriv_cl) |
| 78 | + = expressions( |
| 79 | + y_not_nan_expr, mu_finite_expr, sigma_positive_finite_expr, cdf_expr, |
| 80 | + calc_if<!is_constant_all<T_y_cl, T_loc_cl, T_scale_cl>::value>( |
| 81 | + exp_scaled_diff), |
| 82 | + calc_if<!is_constant<T_scale_cl>::value>(scaled_diff)); |
| 83 | + |
| 84 | + T_partials_return cdf = (from_matrix_cl(cdf_cl)).prod(); |
| 85 | + |
| 86 | + operands_and_partials<decltype(y_col), decltype(mu_col), decltype(sigma_col)> |
| 87 | + ops_partials(y_col, mu_col, sigma_col); |
| 88 | + if (!is_constant_all<T_y_cl, T_loc_cl, T_scale_cl>::value) { |
| 89 | + auto cdf_div_sigma = elt_divide(cdf, sigma_val); |
| 90 | + auto y_deriv = select(cond, cdf_div_sigma, |
| 91 | + elt_divide(cdf_div_sigma, (2.0 * mu_deriv_cl - 1.0))); |
| 92 | + auto mu_deriv = -y_deriv; |
| 93 | + auto sigma_deriv |
| 94 | + = elt_multiply(mu_deriv, static_select<is_constant<T_scale_cl>::value>( |
| 95 | + mu_deriv_cl, sigma_deriv_cl)); |
| 96 | + |
| 97 | + results(mu_deriv_cl, y_deriv_cl, sigma_deriv_cl) |
| 98 | + = expressions(calc_if<!is_constant<T_loc_cl>::value>(mu_deriv), |
| 99 | + calc_if<!is_constant<T_y_cl>::value>(y_deriv), |
| 100 | + calc_if<!is_constant<T_scale_cl>::value>(sigma_deriv)); |
| 101 | + |
| 102 | + if (!is_constant<T_y_cl>::value) { |
| 103 | + ops_partials.edge1_.partials_ = std::move(y_deriv_cl); |
| 104 | + } |
| 105 | + if (!is_constant<T_loc_cl>::value) { |
| 106 | + ops_partials.edge2_.partials_ = std::move(mu_deriv_cl); |
| 107 | + } |
| 108 | + if (!is_constant<T_scale_cl>::value) { |
| 109 | + ops_partials.edge3_.partials_ = std::move(sigma_deriv_cl); |
| 110 | + } |
| 111 | + } |
| 112 | + return ops_partials.build(cdf); |
| 113 | +} |
| 114 | + |
| 115 | +} // namespace math |
| 116 | +} // namespace stan |
| 117 | +#endif |
| 118 | +#endif |
0 commit comments