tenferro-internal-ops 0.1.0

Graph op vocabulary and AD rule implementations for the tenferro workspace.
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

use crate::ad::context::ShapeGuardContext;
use crate::ad::PrimitiveRuleBuilder;
use computegraph::types::{LocalValueId, OperationRole, ValueRef};
use tidu::ADRuleResult;

use crate::std_tensor_op::StdTensorOp;

pub fn linearize_extract_diag(
    builder: &mut dyn PrimitiveRuleBuilder,
    tangent_in: &[Option<LocalValueId>],
    axis_a: usize,
    axis_b: usize,
) -> ADRuleResult<Vec<Option<LocalValueId>>> {
    // TODO: ExtractDiag/EmbedDiag could be replaced by Gather/Scatter
    match tangent_in[0] {
        Some(dx) => {
            let out = builder.add_operation(
                StdTensorOp::ExtractDiag { axis_a, axis_b },
                vec![ValueRef::Local(dx)],
                OperationRole::Linearized {
                    active_mask: vec![true],
                },
            );
            Ok(vec![Some(out[0])])
        }
        None => Ok(vec![None]),
    }
}

pub fn linearize_embed_diag(
    builder: &mut dyn PrimitiveRuleBuilder,
    tangent_in: &[Option<LocalValueId>],
    axis_a: usize,
    axis_b: usize,
) -> ADRuleResult<Vec<Option<LocalValueId>>> {
    // TODO: ExtractDiag/EmbedDiag could be replaced by Gather/Scatter
    match tangent_in[0] {
        Some(dx) => {
            let out = builder.add_operation(
                StdTensorOp::EmbedDiag { axis_a, axis_b },
                vec![ValueRef::Local(dx)],
                OperationRole::Linearized {
                    active_mask: vec![true],
                },
            );
            Ok(vec![Some(out[0])])
        }
        None => Ok(vec![None]),
    }
}

pub fn transpose_extract_diag(
    builder: &mut dyn PrimitiveRuleBuilder,
    cotangent_out: &[Option<LocalValueId>],
    inputs: &[ValueRef<StdTensorOp>],
    axis_a: usize,
    axis_b: usize,
    _ctx: &mut ShapeGuardContext,
) -> ADRuleResult<Vec<Option<LocalValueId>>> {
    // TODO: ExtractDiag/EmbedDiag could be replaced by Gather/Scatter
    match cotangent_out[0] {
        Some(ct) => {
            let out = builder.add_operation(
                StdTensorOp::EmbedDiag { axis_a, axis_b },
                vec![ValueRef::Local(ct)],
                OperationRole::Linearized {
                    active_mask: vec![true],
                },
            );
            let padded_axis_a = builder.add_operation(
                StdTensorOp::PadToMatch { axis: axis_a },
                vec![ValueRef::Local(out[0]), inputs[0].clone()],
                OperationRole::Linearized {
                    active_mask: vec![true, false],
                },
            );
            let padded_axis_b = builder.add_operation(
                StdTensorOp::PadToMatch { axis: axis_b },
                vec![ValueRef::Local(padded_axis_a[0]), inputs[0].clone()],
                OperationRole::Linearized {
                    active_mask: vec![true, false],
                },
            );
            Ok(vec![Some(padded_axis_b[0])])
        }
        None => Ok(vec![None]),
    }
}

pub fn transpose_embed_diag(
    builder: &mut dyn PrimitiveRuleBuilder,
    cotangent_out: &[Option<LocalValueId>],
    inputs: &[ValueRef<StdTensorOp>],
    axis_a: usize,
    axis_b: usize,
    ctx: &mut ShapeGuardContext,
) -> ADRuleResult<Vec<Option<LocalValueId>>> {
    // TODO: ExtractDiag/EmbedDiag could be replaced by Gather/Scatter
    match cotangent_out[0] {
        Some(ct) => {
            let source_axis = if axis_b <= axis_a { axis_a + 1 } else { axis_a };
            let out = builder.add_operation(
                StdTensorOp::ExtractDiag {
                    axis_a: source_axis,
                    axis_b,
                },
                vec![ValueRef::Local(ct)],
                OperationRole::Linearized {
                    active_mask: vec![true],
                },
            );
            if axis_b < axis_a {
                let rank = ctx.rank_of(&inputs[0])?;
                let mut perm: Vec<usize> = (0..rank).collect();
                let diag_axis = perm.remove(axis_b);
                perm.insert(axis_a, diag_axis);
                let transposed = builder.add_operation(
                    StdTensorOp::Transpose { perm },
                    vec![ValueRef::Local(out[0])],
                    OperationRole::Linearized {
                        active_mask: vec![true],
                    },
                );
                return Ok(vec![Some(transposed[0])]);
            }
            Ok(vec![Some(out[0])])
        }
        None => Ok(vec![None]),
    }
}