• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2015 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 // block_params.h: Logic to choose L1 and L2 block sizes
16 // to optimize cache-friendliness.
17 
18 #ifndef GEMMLOWP_INTERNAL_BLOCK_PARAMS_H_
19 #define GEMMLOWP_INTERNAL_BLOCK_PARAMS_H_
20 
21 #include "common.h"
22 
23 namespace gemmlowp {
24 
25 // A BlockParams instance contains a full description of all the block size
26 // parameters to be used by a Gemm.
27 // There are two nested levels of block subdivisions: first a subdivision
28 // into large blocks that should fit in last-level cache (what we call L2 here)
29 // and then another subdivision into smaller blocks that should fit in
30 // L1 cache. There is then actually a third level of subdivision to fit
31 // in registers, but we are not concerned with that here.
32 struct BlockParams {
33   // L1 block parameters determine the size of small blocks that should
34   // fit in L1 cache.
35   int l1_rows;
36   int l1_cols;
37   int l1_depth;
38 
39   // L2 block parameters determine the size of larger blocks that should
40   // fit in L2 cache.
41   int l2_rows;
42   int l2_cols;
43   int l2_depth;
44 
45   template <typename KernelFormat>
InitBlockParams46   void Init(int rows, int cols, int depth, int num_threads, int l1_bytes_to_use,
47             int l2_bytes_to_use, float l2_rhs_factor) {
48     FindL2BlockSizes<KernelFormat>(rows, cols, depth, num_threads,
49                                    l2_bytes_to_use, l2_rhs_factor, &l2_rows,
50                                    &l2_cols, &l2_depth);
51     FindL1BlockSizes<KernelFormat>(l2_rows, l2_cols, l2_depth, l1_bytes_to_use,
52                                    &l1_rows, &l1_cols, &l1_depth);
53   }
54 
55   template <typename KernelFormat>
FindL2BlockSizesBlockParams56   static void FindL2BlockSizes(int rows, int cols, int depth, int num_threads,
57                                int l2_bytes_to_use, float l2_rhs_factor,
58                                int* out_l2_rows, int* out_l2_cols,
59                                int* out_l2_depth) {
60     int l2_rows = 0;
61     int l2_cols = 0;
62     int l2_depth = 0;
63 
64     int per_thread_rows =
65         std::max(1, RoundUp<KernelFormat::kRows>(rows) / num_threads);
66 
67     // No L2 blocking in the depth dimension at the moment.
68     // Too much loss of accuracy due to storing intermediate results in
69     // low precision.
70     // However, we still want to round l2_depth up to the next multiple
71     // of register size, so as to avoid having to special-case unaligned depths.
72     l2_depth = RoundUp<kRegisterSize>(depth);
73 
74     {
75       int max_cache_friendly_l2_cols = std::max(
76           1, static_cast<int>(l2_rhs_factor * (l2_bytes_to_use / l2_depth)));
77       int min_l2_cols_blocks =
78           std::max(1, CeilQuotient(cols, max_cache_friendly_l2_cols));
79       l2_cols =
80           RoundUp<KernelFormat::kCols>(CeilQuotient(cols, min_l2_cols_blocks));
81     }
82 
83     // No L2 blocking in the row dimension if l2_rhs_factor is 1.0 as the row
84     // dimension concerns only the LHS. Blocking only RHS matrix for L2 enhances
85     // the performance on x86.
86     if (l2_rhs_factor == 1.0f) {
87       l2_rows = RoundUp<KernelFormat::kRows>(per_thread_rows);
88     } else {
89       int max_cache_friendly_l2_rows =
90           std::max(1, (l2_bytes_to_use - l2_depth * l2_cols) /
91                           (num_threads * (l2_depth + 4 * l2_cols)));
92       int min_l2_rows_blocks = std::max(
93           1, CeilQuotient(per_thread_rows, max_cache_friendly_l2_rows));
94       l2_rows = RoundUp<KernelFormat::kRows>(
95           CeilQuotient(per_thread_rows, min_l2_rows_blocks));
96     }
97 
98     *out_l2_rows = l2_rows;
99     *out_l2_cols = l2_cols;
100     *out_l2_depth = l2_depth;
101   }
102 
103   template <typename KernelFormat>
FindL1BlockSizesBlockParams104   static void FindL1BlockSizes(int rows, int cols, int depth,
105                                int l1_bytes_to_use, int* out_l1_rows,
106                                int* out_l1_cols, int* out_l1_depth) {
107     int l1_rows = 0;
108     int l1_cols = 0;
109     int l1_depth = 0;
110 
111     // L2 block sizes should already be multiples of kernel block sizes.
112     assert(rows % KernelFormat::kRows == 0);
113     assert(cols % KernelFormat::kCols == 0);
114     assert(depth % KernelFormat::kDepth == 0);
115 
116     // No L1 blocking in the columns dimension at the moment.
117     // Thought not to be needed. Similar to Eigen.
118     l1_cols = cols;
119 
120     {
121       int max_cache_friendly_l1_depth = std::max(
122           1, (l1_bytes_to_use - 4 * KernelFormat::kRows * KernelFormat::kCols) /
123                  (KernelFormat::kRows + KernelFormat::kCols));
124       int min_l1_depth_blocks =
125           std::max(1, CeilQuotient(depth, max_cache_friendly_l1_depth));
126       l1_depth =
127           RoundUp<kRegisterSize>(CeilQuotient(depth, min_l1_depth_blocks));
128     }
129 
130     {
131       int max_cache_friendly_l1_rows =
132           std::max(1, l1_bytes_to_use / (l1_depth + 4 * l1_cols));
133       int min_l1_rows_blocks =
134           std::max(1, CeilQuotient(rows, max_cache_friendly_l1_rows));
135       l1_rows =
136           RoundUp<KernelFormat::kRows>(CeilQuotient(rows, min_l1_rows_blocks));
137     }
138 
139     *out_l1_rows = l1_rows;
140     *out_l1_cols = l1_cols;
141     *out_l1_depth = l1_depth;
142   }
143 };
144 
145 // A SideBlockParams instance contains only the block params relevant to
146 // one side (LHS or RHS), expressed in terms of 'width' instead of
147 // rows/colums. See the explanation in kernel.h: in the LHS, 'width' means
148 // the number of rows, while in the RHS, 'width' means the number of columns.
149 // That allows us to write generic code that applies to either LHS or RHS.
150 struct SideBlockParams {
151   // L1 block parameters determine the size of small blocks that should
152   // fit in L1 cache.
153   int l1_width;
154   int l1_depth;
155 
156   // L2 block parameters determine the size of larger blocks that should
157   // fit in L2 cache.
158   int l2_width;
159   int l2_depth;
160 };
161 
162 enum class Side { Lhs, Rhs };
163 
GetSideBlockParams(Side side,SideBlockParams * side_block_params,const BlockParams & block_params)164 inline void GetSideBlockParams(Side side, SideBlockParams* side_block_params,
165                                const BlockParams& block_params) {
166   side_block_params->l1_width =
167       side == Side::Lhs ? block_params.l1_rows : block_params.l1_cols;
168   side_block_params->l2_width =
169       side == Side::Lhs ? block_params.l2_rows : block_params.l2_cols;
170 
171   side_block_params->l1_depth = block_params.l1_depth;
172   side_block_params->l2_depth = block_params.l2_depth;
173 }
174 
175 }  // namespace gemmlowp
176 
177 #endif  // GEMMLOWP_INTERNAL_BLOCK_PARAMS_H_
178