• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #ifndef TENSORFLOW_COMPILER_XLA_CLIENT_LIB_MATRIX_H_
17 #define TENSORFLOW_COMPILER_XLA_CLIENT_LIB_MATRIX_H_
18 
19 #include <array>
20 #include <vector>
21 
22 #include "absl/strings/string_view.h"
23 #include "absl/types/span.h"
24 #include "tensorflow/compiler/xla/client/xla_builder.h"
25 #include "tensorflow/compiler/xla/statusor.h"
26 #include "tensorflow/compiler/xla/types.h"
27 #include "tensorflow/compiler/xla/xla_data.pb.h"
28 
29 namespace xla {
30 
31 // Returns an m x n matrix with 1s on the diagonal elements, zeros everywhere
32 // else.
33 XlaOp IdentityMatrix(XlaBuilder* builder, PrimitiveType type, int64 m, int64 n);
34 
35 // Returns a mask where the 'diagonal'-th diagonal is true and everything else
36 // is false.
37 XlaOp GetDiagonalMask(XlaOp x, int diagonal = 0);
38 
39 // Get the diagonals of the last two dimensions. Use k>0 for diagonals above the
40 // main diagonal, and k<0 for diagonals below the main diagonal.
41 //
42 // If 'x' has shape [..., M, N]
43 //  If k >= 0: then the output has shape [..., min(M, N - k)], containing the
44 //            diagonal elements (i.e., with indices [..., i, i + k]).
45 //  If k < 0: then the output has shape [..., min(M + k, N)], containing the
46 //            diagonal elements (i.e., with indices [..., i - k, i]).
47 XlaOp GetMatrixDiagonal(XlaOp x, int k = 0);
48 XlaOp GetMatrixDiagonalViaGather(XlaOp x, int k = 0);
49 
50 // Places diag along the kth diagonal of target.
51 XlaOp SetMatrixDiagonal(XlaOp matrix, XlaOp diag, int k = 0);
52 
53 // Returns a lower-triangular mask, i.e., true below the `diagonal`-th diagonal
54 // and false above that diagonal.
55 XlaOp TriangleMask(XlaOp x, int diagonal);
56 
57 // Get the upper or lower triangle part of the last two dimensions
58 XlaOp Triangle(XlaOp x, bool lower);
59 
60 // Get the upper triangle part of the last two dimensions
61 XlaOp UpperTriangle(XlaOp x);
62 
63 // Get the lower triangle part of the last two dimensions
64 XlaOp LowerTriangle(XlaOp x);
65 
66 // Multiplies slices of two tensors in batches.
67 
68 // Multiplies all slices of `Tensor` `x` and `y` (each slice can be
69 // viewed as an element of a batch), and arranges the individual results
70 // in a single output tensor of the same batch size.
71 //
72 // The input tensors `x` and `y` are 2-D or higher with shape `[..., r_x, c_x]`
73 // and `[..., r_y, c_y]`.
74 //
75 // The output tensor is 2-D or higher with shape `[..., r_o, c_o]`, where:
76 //
77 //     r_o = c_x if transpose_x else r_x
78 //     c_o = r_y if transpose_y else c_y
79 //
80 // It is computed as:
81 //
82 //     output[..., :, :] = matrix(x[..., :, :]) * matrix(y[..., :, :])
83 xla::XlaOp BatchDot(
84     xla::XlaOp x, xla::XlaOp y,
85     xla::PrecisionConfig::Precision precision = xla::PrecisionConfig::DEFAULT);
86 xla::XlaOp BatchDot(
87     xla::XlaOp x, bool transpose_x, xla::XlaOp y, bool transpose_y,
88     xla::PrecisionConfig::Precision precision = xla::PrecisionConfig::DEFAULT);
89 
90 // Parse an einsum string into dimension numbers:
91 //   "ab,cb->ac"
92 // becomes:
93 //   {{0, 1},{2, 1},{0, 2}}
94 //
95 // Each occurrence of ellipsis ("...") occurring in the input is replaced with
96 // the same numeric dimensions. The number of such dimensions is inferred from
97 // x_rank and y_rank. For example:
98 //   einsum_config: "...ab,...bcd->...acd"
99 //   x_rank: 4
100 //   y_rank: 5
101 // becomes:
102 //   {{0, 1, 2, 3},{0, 1, 3, 4, 5},{0, 1, 2, 4, 5}}
103 //
104 // NOTE: This function is meant for testing, there is no need to call it
105 // directly.
106 
107 StatusOr<std::array<std::vector<int64>, 3>> ParseEinsumString(
108     absl::string_view einsum_config, int64 x_rank, int64 y_rank);
109 
110 // If an einsum config does not contain an -> one will be added and the output
111 // config will be the sorted characters with any ellipsis at the beginning.
112 // Returns an empty string if the einsum string already has an ->.
113 std::string NormalizeEinsumString(absl::string_view einsum_config);
114 
115 // Supports two operand einsum notation like "ab,cb->ac".
116 xla::XlaOp Einsum(
117     xla::XlaOp x, xla::XlaOp y, absl::string_view einsum_config,
118     xla::PrecisionConfig::Precision precision = xla::PrecisionConfig::DEFAULT);
119 xla::XlaOp Einsum(
120     xla::XlaOp x, absl::string_view einsum_config,
121     xla::PrecisionConfig::Precision precision = xla::PrecisionConfig::DEFAULT);
122 
123 
124 // Same as above but supporting numeric labels on dimensions. So "ab,cb->ac"
125 // becomes:
126 //   x_config = {0, 1}
127 //   y_config = {2, 1}
128 //   output_config = {0, 2}
129 xla::XlaOp Einsum(
130     xla::XlaOp x, absl::Span<const int64> x_config, xla::XlaOp y,
131     absl::Span<const int64> y_config, absl::Span<const int64> output_config,
132     xla::PrecisionConfig::Precision precision = xla::PrecisionConfig::DEFAULT);
133 
134 // Transposes a stack of matrices `x` by swapping the last two dimensions.
135 xla::XlaOp TransposeInMinorDims(xla::XlaOp x);
136 
137 // Transposes `x` in its minor dimensions if `transpose` is true, otherwise
138 // returns `x` unchanged.
139 xla::XlaOp MaybeTransposeInMinorDims(xla::XlaOp x, bool transpose);
140 
141 }  // namespace xla
142 
143 #endif  // TENSORFLOW_COMPILER_XLA_CLIENT_LIB_MATRIX_H_
144