1 #ifndef STAN_MATH_OPENCL_PRIM_NORMAL_LPDF_HPP
2 #define STAN_MATH_OPENCL_PRIM_NORMAL_LPDF_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  * The log of the normal density for the specified scalar(s) given
18  * the specified mean(s) and deviation(s). y, mu, or sigma can
19  * each be either a scalar or a vector matrix_cl. Any vector inputs
20  * must be the same length.
21  *
22  * <p>The result log probability is defined to be the sum of the
23  * log probabilities for each observation/mean/deviation triple.
24  *
25  * @tparam T_y_cl type of scalar
26  * @tparam T_loc_cl type of location parameter
27  * @tparam T_scale_cl type of scale parameter
28  * @param y (Sequence of) scalar(s).
29  * @param mu (Sequence of) location parameter(s)
30  * for the normal distribution.
31  * @param sigma (Sequence of) scale parameters for the normal distribution.
32  * @return The log of the product of the densities.
33  * @throw std::domain_error if the scale is not positive.
34  */
35 template <
36     bool propto, typename T_y_cl, typename T_loc_cl, typename T_scale_cl,
37     require_all_prim_or_rev_kernel_expression_t<T_y_cl, T_loc_cl,
38                                                 T_scale_cl>* = nullptr,
39     require_any_not_stan_scalar_t<T_y_cl, T_loc_cl, T_scale_cl>* = nullptr>
normal_lpdf(const T_y_cl & y,const T_loc_cl & mu,const T_scale_cl & sigma)40 inline return_type_t<T_y_cl, T_loc_cl, T_scale_cl> normal_lpdf(
41     const T_y_cl& y, const T_loc_cl& mu, const T_scale_cl& sigma) {
42   static const char* function = "normal_lpdf(OpenCL)";
43   using T_partials_return = partials_return_t<T_y_cl, T_loc_cl, T_scale_cl>;
44   using std::isfinite;
45   using std::isnan;
46 
47   check_consistent_sizes(function, "Random variable", y, "Location parameter",
48                          mu, "Scale parameter", sigma);
49   const size_t N = max_size(y, mu, sigma);
50   if (N == 0) {
51     return 0.0;
52   }
53   if (!include_summand<propto, T_y_cl, T_loc_cl, T_scale_cl>::value) {
54     return 0.0;
55   }
56 
57   const auto& y_col = as_column_vector_or_scalar(y);
58   const auto& mu_col = as_column_vector_or_scalar(mu);
59   const auto& sigma_col = as_column_vector_or_scalar(sigma);
60 
61   const auto& y_val = value_of(y_col);
62   const auto& mu_val = value_of(mu_col);
63   const auto& sigma_val = value_of(sigma_col);
64 
65   auto check_y_not_nan
66       = check_cl(function, "Random variable", y_val, "not NaN");
67   auto y_not_nan = !isnan(y_val);
68   auto check_mu_finite
69       = check_cl(function, "Location parameter", mu_val, "finite");
70   auto mu_finite = isfinite(mu_val);
71   auto check_sigma_positive
72       = check_cl(function, "Scale parameter", sigma_val, "positive");
73   auto sigma_positive = 0 < sigma_val;
74 
75   auto inv_sigma = elt_divide(1., sigma_val);
76   auto y_scaled = elt_multiply((y_val - mu_val), inv_sigma);
77   auto y_scaled_sq = elt_multiply(y_scaled, y_scaled);
78 
79   auto logp1 = -0.5 * y_scaled_sq;
80   auto logp_expr
81       = colwise_sum(static_select<include_summand<propto, T_scale_cl>::value>(
82           logp1 - log(sigma_val), logp1));
83 
84   auto scaled_diff = elt_multiply(inv_sigma, y_scaled);
85   auto sigma_deriv = elt_multiply(inv_sigma, y_scaled_sq) - inv_sigma;
86 
87   matrix_cl<double> logp_cl;
88   matrix_cl<double> mu_deriv_cl;
89   matrix_cl<double> y_deriv_cl;
90   matrix_cl<double> sigma_deriv_cl;
91 
92   results(check_y_not_nan, check_mu_finite, check_sigma_positive, logp_cl,
93           y_deriv_cl, mu_deriv_cl, sigma_deriv_cl)
94       = expressions(y_not_nan, mu_finite, sigma_positive, logp_expr,
95                     calc_if<!is_constant<T_y_cl>::value>(-scaled_diff),
96                     calc_if<!is_constant<T_loc_cl>::value>(scaled_diff),
97                     calc_if<!is_constant<T_scale_cl>::value>(sigma_deriv));
98 
99   T_partials_return logp = sum(from_matrix_cl(logp_cl));
100 
101   if (include_summand<propto>::value) {
102     logp += NEG_LOG_SQRT_TWO_PI * N;
103   }
104 
105   operands_and_partials<decltype(y_col), decltype(mu_col), decltype(sigma_col)>
106       ops_partials(y_col, mu_col, sigma_col);
107 
108   if (!is_constant<T_y_cl>::value) {
109     ops_partials.edge1_.partials_ = std::move(y_deriv_cl);
110   }
111   if (!is_constant<T_loc_cl>::value) {
112     ops_partials.edge2_.partials_ = std::move(mu_deriv_cl);
113   }
114   if (!is_constant<T_scale_cl>::value) {
115     ops_partials.edge3_.partials_ = std::move(sigma_deriv_cl);
116   }
117   return ops_partials.build(logp);
118 }
119 
120 }  // namespace math
121 }  // namespace stan
122 #endif
123 #endif
124