1 | /* |
2 | * Licensed to the Apache Software Foundation (ASF) under one |
3 | * or more contributor license agreements. See the NOTICE file |
4 | * distributed with this work for additional information |
5 | * regarding copyright ownership. The ASF licenses this file |
6 | * to you under the Apache License, Version 2.0 (the |
7 | * "License"); you may not use this file except in compliance |
8 | * with the License. You may obtain a copy of the License at |
9 | * |
10 | * http://www.apache.org/licenses/LICENSE-2.0 |
11 | * |
12 | * Unless required by applicable law or agreed to in writing, |
13 | * software distributed under the License is distributed on an |
14 | * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
15 | * KIND, either express or implied. See the License for the |
16 | * specific language governing permissions and limitations |
17 | * under the License. |
18 | */ |
19 | |
20 | /*! |
21 | * \brief Binary op constructions |
22 | * \file nn/bnn.h |
23 | */ |
24 | #ifndef TVM_TOPI_NN_BNN_H_ |
25 | #define TVM_TOPI_NN_BNN_H_ |
26 | |
27 | #include <tvm/arith/analyzer.h> |
28 | #include <tvm/te/operation.h> |
29 | #include <tvm/topi/detail/constant_utils.h> |
30 | #include <tvm/topi/tags.h> |
31 | |
32 | #include <string> |
33 | |
34 | namespace tvm { |
35 | namespace topi { |
36 | namespace nn { |
37 | |
38 | using namespace tvm::te; |
39 | |
40 | /*! |
41 | * \brief Binarization and bit-packing along a certain axis. |
42 | * |
43 | * \param data N-D tensor, can be any layout |
44 | * \param axis The axis along which to do binarization and bit-packing. This axis |
45 | * must have a size equal to an integer multiple of 32. |
46 | * \param name The name of the operation |
47 | * \param tag The tag to mark the operation |
48 | * |
49 | * \return Output tensor with dtype uint32 |
50 | */ |
51 | inline tvm::te::Tensor binarize_pack(const tvm::te::Tensor& data, int axis, |
52 | std::string name = "PackedInput" , |
53 | std::string tag = "binarize_pack" ) { |
54 | auto ishape = data->shape; |
55 | ICHECK_EQ(GetConstInt(ishape[axis]) % 32, 0) |
56 | << "binarize_pack: axis size must be a multiple of 32" ; |
57 | |
58 | arith::Analyzer analyzer; |
59 | auto n = ishape.size(); |
60 | Array<PrimExpr> oshape; |
61 | for (size_t i = 0; i < n; ++i) { |
62 | oshape.push_back(i == static_cast<size_t>(axis) ? analyzer.Simplify(indexdiv(ishape[i], 32)) |
63 | : ishape[i]); |
64 | } |
65 | |
66 | return tvm::te::compute( |
67 | oshape, |
68 | [&](const Array<Var>& indices) { |
69 | Array<PrimExpr> start_idx; |
70 | for (size_t i = 0; i < n; ++i) { |
71 | start_idx.push_back(i == static_cast<size_t>(axis) ? indices[i] * 32 |
72 | : static_cast<PrimExpr>(indices[i])); |
73 | } |
74 | auto packed = make_const(DataType::UInt(32), 0); |
75 | for (size_t j = 0; j < 32; ++j) { |
76 | Array<PrimExpr> idx; |
77 | for (size_t i = 0; i < n; ++i) { |
78 | idx.push_back(i == static_cast<size_t>(axis) ? start_idx[i] + static_cast<int>(j) |
79 | : start_idx[i]); |
80 | } |
81 | auto sign = tvm::cast(DataType::UInt(32), data(idx) >= 0); |
82 | packed = (packed | sign); |
83 | if (j == 31) { |
84 | return packed; |
85 | } |
86 | packed = packed << 1; |
87 | } |
88 | return packed; // never reached, but suppress compiler warning |
89 | }, |
90 | name, tag); |
91 | } |
92 | |
93 | /*! |
94 | * \brief Binary matrix multiplication using xor and bit-count |
95 | * |
96 | * \param data Tensor with shape [batch, in_dim], dtype is uint32 |
97 | * \param weight Tensor with shape [out_dim, in_dim], dtype is uint32 |
98 | * |
99 | * \return Tensor with shape [batch, out_dim], dtype is float32 |
100 | */ |
101 | inline tvm::te::Tensor binary_dense(const tvm::te::Tensor& data, const tvm::te::Tensor& weight) { |
102 | ICHECK_EQ(data->shape.size(), 2) << "binary_dense requires 2-D data" ; |
103 | ICHECK_EQ(weight->shape.size(), 2) << "binary_dense requires 2-D weight" ; |
104 | ICHECK_EQ(data->dtype, DataType::UInt(32)) << "binary_dense requires uint32 data" ; |
105 | ICHECK_EQ(weight->dtype, DataType::UInt(32)) << "binary_dense requires uint32 weight" ; |
106 | |
107 | auto batch = data->shape[0]; |
108 | auto in_dim = data->shape[1]; |
109 | auto out_dim = weight->shape[0]; |
110 | |
111 | auto k = tvm::te::reduce_axis(Range(0, in_dim), "k" ); |
112 | auto matmul = tvm::te::compute( |
113 | {batch, out_dim}, |
114 | [&](Var i, Var j) { return tvm::sum(popcount(data(i, k) ^ weight(j, k)), {k}); }, "tensor" , |
115 | "binary_dense" ); |
116 | |
117 | return tvm::te::compute( |
118 | {batch, out_dim}, [&](Var i, Var j) { return 32 * in_dim - 2.0f * matmul(i, j); }, "tensor" , |
119 | kElementWise); |
120 | } |
121 | |
122 | } // namespace nn |
123 | } // namespace topi |
124 | } // namespace tvm |
125 | #endif // TVM_TOPI_NN_BNN_H_ |
126 | |