• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2017 The Gemmlowp 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 // dispatch_gemm_shape.h: dispatch GEMM calls according to their shape
16 
17 #ifndef GEMMLOWP_INTERNAL_DISPATCH_GEMM_SHAPE_H_
18 #define GEMMLOWP_INTERNAL_DISPATCH_GEMM_SHAPE_H_
19 
20 #include "../internal/kernel_default.h"
21 #include "../public/map.h"
22 #include "../public/output_stages.h"
23 #include "multi_thread_gemm.h"
24 
25 namespace gemmlowp {
26 
27 template <typename T>
28 struct TransposeImpl {
29   typedef T DstType;
RunTransposeImpl30   static T Run(const T& t) { return t; }
31 };
32 
33 template <typename T>
34 using TransposeType = typename TransposeImpl<T>::DstType;
35 
36 template <typename T>
Transpose(const T & t)37 TransposeType<T> Transpose(const T& t) {
38   return TransposeImpl<T>::Run(t);
39 }
40 
41 template <MapOrder Order>
42 struct TransposeMapOrder {
43   static constexpr MapOrder Value =
44       Order == MapOrder::RowMajor ? MapOrder::ColMajor : MapOrder::RowMajor;
45 };
46 
47 template <VectorShape Shape>
48 struct TransposeVectorShape {
49   static constexpr VectorShape Value =
50       Shape == VectorShape::Row ? VectorShape::Col : VectorShape::Row;
51 };
52 
53 template <typename Scalar, VectorShape Shape>
54 struct TransposeImpl<VectorMap<Scalar, Shape>> {
55   typedef VectorMap<Scalar, Shape> SrcType;
56   static constexpr VectorShape TransposedShape =
57       TransposeVectorShape<Shape>::Value;
58   typedef VectorMap<Scalar, TransposedShape> DstType;
59   static DstType Run(const SrcType& src) {
60     return DstType(src.data(), src.size());
61   }
62 };
63 
64 template <typename Scalar, MapOrder Order>
65 struct TransposeImpl<MatrixMap<Scalar, Order>> {
66   typedef MatrixMap<Scalar, Order> SrcType;
67   static constexpr MapOrder TransposedOrder = TransposeMapOrder<Order>::Value;
68   typedef MatrixMap<Scalar, TransposedOrder> DstType;
69   static DstType Run(const SrcType& src) {
70     return DstType(src.data(), src.cols(), src.rows(), src.stride());
71   }
72 };
73 
74 template <VectorShape Shape>
75 struct TransposeImpl<OutputStageQuantizeDownInt32ToUint8ScalePC<Shape>> {
76   typedef OutputStageQuantizeDownInt32ToUint8ScalePC<Shape> SrcType;
77   static const VectorShape TransposedShape = TransposeVectorShape<Shape>::Value;
78   typedef OutputStageQuantizeDownInt32ToUint8ScalePC<TransposedShape> DstType;
79   static DstType Run(const SrcType& src) {
80     DstType dst;
81     dst.result_shift = src.result_shift;
82     dst.result_offset = Transpose(src.result_offset);
83     dst.result_mult_int = Transpose(src.result_mult_int);
84     return dst;
85   }
86 };
87 
88 template <typename VectorMapType>
89 struct TransposeImpl<OutputStageBiasAddition<VectorMapType>> {
90   typedef OutputStageBiasAddition<VectorMapType> SrcType;
91   typedef TransposeType<VectorMapType> TransposedVectorMapType;
92   typedef OutputStageBiasAddition<TransposedVectorMapType> DstType;
93   static DstType Run(const SrcType& src) {
94     DstType dst;
95     dst.bias_vector = Transpose(src.bias_vector);
96     return dst;
97   }
98 };
99 
100 // TODO(benoitjacob) - does anyone understand C++ variadic templates?
101 // How to use them to implement TransposeTuple? Note: there are lots
102 // of answers on StackOverflow but they seem to all involve either
103 // C++14/C++17 (we can only use C++11) or lots of abstract nonsense.
104 inline std::tuple<> TransposeTuple(const std::tuple<>& t) { return t; }
105 
106 template <typename T0>
107 std::tuple<TransposeType<T0>> TransposeTuple(const std::tuple<T0>& t) {
108   return std::make_tuple(Transpose(std::get<0>(t)));
109 }
110 
111 template <typename T0, typename T1>
112 std::tuple<TransposeType<T0>, TransposeType<T1>> TransposeTuple(
113     const std::tuple<T0, T1>& t) {
114   return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)));
115 }
116 
117 template <typename T0, typename T1, typename T2>
118 std::tuple<TransposeType<T0>, TransposeType<T1>, TransposeType<T2>>
119 TransposeTuple(const std::tuple<T0, T1, T2>& t) {
120   return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)),
121                          Transpose(std::get<2>(t)));
122 }
123 
124 template <typename T0, typename T1, typename T2, typename T3>
125 std::tuple<TransposeType<T0>, TransposeType<T1>, TransposeType<T2>,
126            TransposeType<T3>>
127 TransposeTuple(const std::tuple<T0, T1, T2, T3>& t) {
128   return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)),
129                          Transpose(std::get<2>(t)), Transpose(std::get<3>(t)));
130 }
131 
132 template <typename T0, typename T1, typename T2, typename T3, typename T4>
133 std::tuple<TransposeType<T0>, TransposeType<T1>, TransposeType<T2>,
134            TransposeType<T3>, TransposeType<T4>>
135 TransposeTuple(const std::tuple<T0, T1, T2, T3, T4>& t) {
136   return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)),
137                          Transpose(std::get<2>(t)), Transpose(std::get<3>(t)),
138                          Transpose(std::get<4>(t)));
139 }
140 
141 template <typename T0, typename T1, typename T2, typename T3, typename T4,
142           typename T5>
143 std::tuple<TransposeType<T0>, TransposeType<T1>, TransposeType<T2>,
144            TransposeType<T3>, TransposeType<T4>, TransposeType<T5>>
145 TransposeTuple(const std::tuple<T0, T1, T2, T3, T4, T5>& t) {
146   return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)),
147                          Transpose(std::get<2>(t)), Transpose(std::get<3>(t)),
148                          Transpose(std::get<4>(t)), Transpose(std::get<5>(t)));
149 }
150 
151 template <typename InputScalar, typename OutputScalar, typename BitDepthParams,
152           MapOrder LhsOrder, MapOrder RhsOrder, MapOrder ResultOrder,
153           typename LhsOffset, typename RhsOffset, typename OutputPipelineType,
154           typename GemmContextType>
155 void DispatchGemmShape(GemmContextType* context,
156                        const MatrixMap<const InputScalar, LhsOrder>& lhs,
157                        const MatrixMap<const InputScalar, RhsOrder>& rhs,
158                        MatrixMap<OutputScalar, ResultOrder>* result,
159                        const LhsOffset& lhs_offset, const RhsOffset& rhs_offset,
160                        const OutputPipelineType& output_pipeline) {
161   assert(lhs.cols() == rhs.rows());
162 
163   int rows = result->rows();
164   int cols = result->cols();
165   int depth = lhs.cols();
166 
167   if (rows == 0 || cols == 0 || depth == 0) {
168     // Vacuous GEMM, return early to avoid having to deal with
169     // zero sizes below.
170     return;
171   }
172 
173   if (rows < cols) {
174     auto transposed_result_map = Transpose(*result);
175     return DispatchGemmShape<InputScalar, OutputScalar, BitDepthParams>(
176         context, Transpose(rhs), Transpose(lhs), &transposed_result_map,
177         Transpose(rhs_offset), Transpose(lhs_offset),
178         TransposeTuple(output_pipeline));
179   }
180 
181   typedef DefaultKernel<BitDepthParams> Kernel;
182   MultiThreadGemm<typename Kernel::Format, InputScalar, OutputScalar,
183                   BitDepthParams>(context, Kernel(), lhs, rhs, result,
184                                   lhs_offset, rhs_offset, output_pipeline);
185 }
186 
187 }  // end namespace gemmlowp
188 
189 #endif  // GEMMLOWP_INTERNAL_DISPATCH_GEMM_SHAPE_H_
190