dfdx 0.13.0

Ergonomic auto differentiation in Rust, with pytorch like apis.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137

use crate::{
    dtypes::*,
    shapes::*,
    tensor::{launch_cfg, Cuda, Tensor, Tensorlike},
    tensor_ops::reduction_utils::*,
};

use cudarc::driver::{DeviceRepr, DeviceSlice, LaunchAsync, ValidAsZeroBits};

use std::vec::Vec;

const PTX_SRC: &str = include_str!(concat!(env!("OUT_DIR"), "/sum_to.ptx"));

trait HasCudaKernel<E> {
    const MOD: &'static str;
    const FNS: &'static [&'static str];
}

#[cfg(feature = "f16")]
impl HasCudaKernel<f16> for Cuda {
    const MOD: &'static str = "sum_f16";
    const FNS: &'static [&'static str] = &["sum_to_fwd_f16", "sum_to_bwd_f16"];
}

#[cfg(feature = "f16")]
impl HasCudaKernel<AMP<f16>> for Cuda {
    const MOD: &'static str = "sum_amp_f16";
    const FNS: &'static [&'static str] = &["sum_to_fwd_amp_f16", "sum_to_bwd_f16"];
}

impl HasCudaKernel<f32> for Cuda {
    const MOD: &'static str = "sum_f32";
    const FNS: &'static [&'static str] = &["sum_to_fwd_f32", "sum_to_bwd_f32"];
}

impl HasCudaKernel<f64> for Cuda {
    const MOD: &'static str = "sum_f64";
    const FNS: &'static [&'static str] = &["sum_to_fwd_f64", "sum_to_bwd_f64"];
}

impl<E: Dtype + ValidAsZeroBits + DeviceRepr> super::SumKernel<E> for Cuda
where
    Self: HasCudaKernel<E>,
{
    fn forward<Src: Shape, Dst: Shape, Ax: Axes>(
        &self,
        dst: Dst,
        inp: &Tensor<Src, E, Self>,
    ) -> Result<Tensor<Dst, E, Self>, Self::Err>
    where
        Src: ReduceShapeTo<Dst, Ax>,
    {
        if !self.dev.has_func(Self::MOD, Self::FNS[0]) {
            self.dev.load_ptx(PTX_SRC.into(), Self::MOD, Self::FNS)?;
        }

        let fwd_fn = self.dev.get_func(Self::MOD, Self::FNS[0]).unwrap();

        let (dims, strides) = permute_for_reductions::<_, Ax>(inp.shape.concrete(), inp.strides);
        let num_dims = dims.len();

        let mut info = Vec::with_capacity(num_dims * 2);
        info.extend(dims);
        info.extend(strides);
        let info = self.dev.htod_copy(info)?;

        let elems_per_thread = E::from_usize(reduction_elems_per_thread::<_, Src>(
            inp.shape.concrete(),
            inp.strides,
            Ax::as_array(),
        ))
        .unwrap();

        let physical_numel = inp.data.len();
        let (dst_physical_numel, dst_strides) =
            reduction_output_strides::<Ax, Src, Dst>(inp.strides, dst);
        let chunk_len = physical_numel / dst_physical_numel;

        let cfg = launch_cfg::<128>(physical_numel as u32);

        let mut storage = unsafe { self.alloc_empty::<E>(dst_physical_numel) }?;
        self.dev.memset_zeros(&mut storage)?;
        let params = (
            physical_numel,    // const size_t numel,
            num_dims,          // const size_t num_dims,
            elems_per_thread,  // const float elems_per_thread,
            chunk_len,         // const size_t chunk_len,
            &info,             // const size_t *info,
            inp.data.as_ref(), // const float *inp,
            &mut storage,      // float *out
        );
        unsafe { fwd_fn.launch(cfg, params) }?;
        Ok(self.build_tensor(dst, dst_strides, storage))
    }

    fn backward<Src: Shape, Dst: Shape, Ax: Axes>(
        &self,
        dst: Dst,
        inp: &impl Tensorlike<Src, E, Self>,
        grad_inp: &mut Self::Vec,
        grad_out: &Self::Vec,
    ) -> Result<(), Self::Err>
    where
        Src: ReduceShapeTo<Dst, Ax>,
    {
        let bwd_fn = self.dev.get_func(Self::MOD, Self::FNS[1]).unwrap();

        let out_strides: Src::Concrete =
            BroadcastStridesTo::<Src, Ax>::broadcast_strides(&dst, dst.strides());
        let physical_numel = inp.len();
        let elems_per_thread = E::from_usize(reduction_elems_per_thread::<_, Src>(
            inp.shape().concrete(),
            inp.strides(),
            Ax::as_array(),
        ))
        .unwrap();

        let cfg = launch_cfg::<128>(physical_numel as u32);

        let mut info: Vec<usize> = Vec::with_capacity(3 * Src::NUM_DIMS);
        info.extend(inp.shape().concrete());
        info.extend(inp.strides());
        info.extend(out_strides);
        let info = self.dev.htod_copy(info)?;

        let params = (
            physical_numel,   // const size_t numel,
            Src::NUM_DIMS,    // const size_t num_dims,
            elems_per_thread, // const float elems_per_thread,
            &info,            // const size_t *info,
            grad_inp,         // float *grad_inp,
            grad_out,         // const float *grad_out,
        );
        unsafe { bwd_fn.launch(cfg, params) }?;
        Ok(())
    }
}