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
|
// Copyright Vespa.ai. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
#pragma once
#include <vespa/eval/eval/tensor_function.h>
namespace vespalib::eval {
/**
* Tensor function implementing generalized lookup of 'key' in 'map'
* with some type restrictions.
*
* 'key' may only contain the lookup dimension (called 'x' here)
* 'map' must have full mapped overlap with 'key'
*
* Both tensors must have the same cell type, which can be either
* float or double.
*
* The optimized expression looks like this: reduce(key*map,sum,x)
*
* If 'map' is also sparse, the lookup operation is a sparse dot
* product and will be optimized using SparseDotProductFunction
* instead.
*
* The best performance (simple hash lookup with a result referencing
* existing cells without having to copy them) is achieved when a
* single dense subspace in 'map' matches a cell with value 1.0 from
* 'key'. This fast-path can be ensured if this optimization is
* combined with the simple_join_count optimization:
*
* key = tensor(x{}):{my_key:1}
* map = tensor(x{},y[128])
* fallback = tensor(y[128])
*
* simple lookup with fallback:
* if(reduce(key*map,count)==128,reduce(key*map,sum,x),fallback)
**/
class MappedLookup : public tensor_function::Op2
{
public:
MappedLookup(const ValueType &res_type, const TensorFunction &key_in, const TensorFunction &map_in);
const TensorFunction &key() const { return lhs(); }
const TensorFunction &map() const { return rhs(); }
InterpretedFunction::Instruction compile_self(const ValueBuilderFactory &factory, Stash &stash) const override;
bool result_is_mutable() const override { return map().result_is_mutable(); }
static const TensorFunction &optimize(const TensorFunction &expr, Stash &stash);
};
} // namespace
|