1use custos::{impl_stack, number::Number, Alloc, Buffer, Device, MainMemory, Shape, CPU};
2
3#[cfg(feature = "stack")]
4use custos::Stack;
5
6#[cfg(any(feature = "cuda", feature = "opencl"))]
7use custos::CDatatype;
8
9#[cfg(feature = "opencl")]
10use crate::cl_tew_self;
11#[cfg(feature = "opencl")]
12use custos::OpenCL;
13
14#[cfg(feature = "cuda")]
15use crate::cu_ew_self;
16use crate::{assign_to_lhs, element_wise_op_mut, Matrix};
17
18#[cfg_attr(feature = "cpu", doc = "```")]
21#[cfg_attr(not(feature = "cpu"), doc = "```ignore")]
22pub trait AssignOps<T, S: Shape = (), D: Device = Self>: Device {
36 #[cfg_attr(feature = "cpu", doc = "```")]
39 #[cfg_attr(not(feature = "cpu"), doc = "```ignore")]
40 fn add_assign(&self, lhs: &mut Buffer<T, Self, S>, rhs: &Buffer<T, D, S>);
54 fn sub_assign(&self, lhs: &mut Buffer<T, Self, S>, rhs: &Buffer<T, D, S>);
55 fn mul_assign(&self, lhs: &mut Buffer<T, Self, S>, rhs: &Buffer<T, D, S>);
56}
57
58pub fn ew_op<'a, T, F, D, S, Host>(
59 device: &'a Host,
60 lhs: &Matrix<T, D, S>,
61 rhs: &Matrix<T, D, S>,
62 f: F,
63) -> Matrix<'a, T, Host, S>
64where
65 T: Copy + Default,
66 F: Fn(T, T) -> T,
67 D: MainMemory,
68 S: Shape,
69 Host: for<'b> Alloc<'b, T, S> + MainMemory,
70{
71 let mut out = device.retrieve(lhs.size(), (lhs.node.idx, rhs.node.idx));
72 element_wise_op_mut(lhs, rhs, &mut out, f);
73 (out, lhs.dims()).into()
74}
75
76#[impl_stack]
77impl<T: Number, D: MainMemory, S: Shape> AssignOps<T, S, D> for CPU {
78 #[inline]
79 fn add_assign(&self, lhs: &mut Buffer<T, Self, S>, rhs: &Buffer<T, D, S>) {
80 assign_to_lhs(lhs, rhs, |x, y| *x += y)
81 }
82
83 #[inline]
84 fn sub_assign(&self, lhs: &mut Buffer<T, Self, S>, rhs: &Buffer<T, D, S>) {
85 assign_to_lhs(lhs, rhs, |x, y| *x -= y)
86 }
87
88 #[inline]
89 fn mul_assign(&self, lhs: &mut Buffer<T, Self, S>, rhs: &Buffer<T, D, S>) {
90 assign_to_lhs(lhs, rhs, |x, y| *x *= y)
91 }
92}
93
94#[cfg(feature = "opencl")]
95impl<T: CDatatype> AssignOps<T> for OpenCL {
96 #[inline]
97 fn add_assign(&self, lhs: &mut Buffer<T, OpenCL>, rhs: &Buffer<T, OpenCL>) {
98 cl_tew_self(self, lhs, rhs, "+").unwrap()
99 }
100
101 #[inline]
102 fn sub_assign(&self, lhs: &mut Buffer<T, OpenCL>, rhs: &Buffer<T, OpenCL>) {
103 cl_tew_self(self, lhs, rhs, "-").unwrap()
104 }
105
106 #[inline]
107 fn mul_assign(&self, lhs: &mut Buffer<T, Self, ()>, rhs: &Buffer<T, Self, ()>) {
108 cl_tew_self(self, lhs, rhs, "*").unwrap()
109 }
110}
111
112#[cfg(feature = "cuda")]
113impl<T: CDatatype> AssignOps<T> for custos::CUDA {
114 #[inline]
115 fn add_assign(&self, lhs: &mut Buffer<T, custos::CUDA>, rhs: &Buffer<T, custos::CUDA>) {
116 cu_ew_self(self, lhs, rhs, "+").unwrap();
117 }
118
119 #[inline]
120 fn sub_assign(&self, lhs: &mut Buffer<T, custos::CUDA>, rhs: &Buffer<T, custos::CUDA>) {
121 cu_ew_self(self, lhs, rhs, "-").unwrap();
122 }
123
124 #[inline]
125 fn mul_assign(&self, lhs: &mut Buffer<T, Self, ()>, rhs: &Buffer<T, Self, ()>) {
126 cu_ew_self(self, lhs, rhs, "*").unwrap();
127 }
128}