xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2022 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #include "generic.hpp"
26 
27 #include <functional>
28 
29 namespace arm_conv {
30 namespace depthwise {
31 namespace interleaves {
32 
PackingArguments(unsigned int kernel_rows,unsigned int kernel_cols,size_t weight_element_size,bool include_bias,size_t bias_element_size,arm_gemm::VLType vl_type,size_t accumulator_element_size,unsigned int accumulator_depth_vl,std::function<bool (unsigned int,unsigned int &,unsigned int &)> get_weight_pos)33 PackingArguments::PackingArguments(
34   unsigned int kernel_rows, unsigned int kernel_cols, size_t weight_element_size,
35   bool include_bias, size_t bias_element_size,
36   arm_gemm::VLType vl_type, size_t accumulator_element_size, unsigned int accumulator_depth_vl,
37   std::function<bool(unsigned int, unsigned int &, unsigned int &)> get_weight_pos
38 ) : kernel_rows(kernel_rows), kernel_cols(kernel_cols), weight_element_size(weight_element_size),
39     include_bias(include_bias), bias_element_size(bias_element_size),
40     vl_type(vl_type), accumulator_element_size(accumulator_element_size), accumulator_depth_vl(accumulator_depth_vl),
41     get_weight_pos(get_weight_pos)
42 {
43 }
44 
get_storage_size_generic(const PackingArguments & packing_args,const DepthwiseArgs & args)45 size_t get_storage_size_generic(const PackingArguments &packing_args, const DepthwiseArgs &args)
46 {
47   // If the channel multiplier is greater than one, then we treat this as a
48   // repeated packing of `channel_multiplier`-sized problems.
49   if (args.channel_multiplier > 1)
50   {
51     DepthwiseArgs args_per_input_channel(args);
52     args_per_input_channel.input_channels = args.channel_multiplier;
53     args_per_input_channel.channel_multiplier = 1;
54 
55     return args.input_channels * get_storage_size_generic(packing_args, args_per_input_channel);
56   }
57 
58   const unsigned int vl =
59     packing_args.accumulator_depth_vl *
60     arm_gemm::utils::get_vector_length<uint8_t>(packing_args.vl_type) / packing_args.accumulator_element_size;
61   const unsigned int n_packs = arm_gemm::iceildiv(args.input_channels, vl);
62   const auto pack_size = (packing_args.include_bias ? packing_args.bias_element_size : 0) +
63                          packing_args.kernel_points() * packing_args.weight_element_size;
64   return n_packs * pack_size * vl;
65 }
66 
pack_parameters_generic(const PackingArguments & packing_args,const DepthwiseArgs & args,void * buffer_raw,const void * biases_raw,const void * weights_raw,size_t ld_weight_col,size_t ld_weight_row)67 void pack_parameters_generic(
68   const PackingArguments &packing_args,
69   const DepthwiseArgs &args,
70   void *buffer_raw,
71   const void *biases_raw,
72   const void *weights_raw,
73   size_t ld_weight_col,
74   size_t ld_weight_row
75 )
76 {
77   // Cast the pointers to byte sizes
78   auto *buffer = static_cast<uint8_t *>(buffer_raw);
79   auto *biases = static_cast<const uint8_t *>(biases_raw);
80   auto *weights = static_cast<const uint8_t *>(weights_raw);
81 
82   // If the channel multiplier is greater than one, then we treat this as a
83   // repeated packing of `channel_multiplier`-sized problems.
84   if (args.channel_multiplier > 1)
85   {
86     // Get a modified copy of the depthwise arguments
87     DepthwiseArgs args_per_input_channel(args);
88     args_per_input_channel.input_channels = args.channel_multiplier;
89     args_per_input_channel.channel_multiplier = 1;
90 
91     // Resolve the strides here
92     ld_weight_col = ld_weight_col ? ld_weight_col : args.input_channels * args.channel_multiplier;
93     ld_weight_row = ld_weight_row ? ld_weight_row : ld_weight_col * packing_args.kernel_cols;
94 
95     auto per_input_channel_size = get_storage_size_generic(packing_args, args_per_input_channel);
96 
97     for (unsigned int c = 0; c < args.input_channels; c++)
98     {
99       pack_parameters_generic(
100         packing_args, args_per_input_channel, buffer, biases, weights, ld_weight_col, ld_weight_row);
101 
102       // Update the pointers
103       buffer += per_input_channel_size;
104       biases += (biases == nullptr) ? 0 : packing_args.bias_element_size * args.channel_multiplier;
105       weights += packing_args.weight_element_size * args.channel_multiplier;
106     }
107     return;
108   }
109 
110   // Finalise the weight strides
111   ld_weight_col = (ld_weight_col == 0) ? args.input_channels : ld_weight_col;
112   ld_weight_row = (ld_weight_row == 0) ? packing_args.kernel_cols * ld_weight_col : ld_weight_row;
113 
114   const unsigned int vl =
115     packing_args.accumulator_depth_vl *
116     arm_gemm::utils::get_vector_length<uint8_t>(packing_args.vl_type) / packing_args.accumulator_element_size;
117 
118   for (unsigned int n = 0; n < args.input_channels; n += vl)
119   {
120     const unsigned int todo = std::min(vl, args.input_channels - n);
121 
122     if (packing_args.include_bias)
123     {
124       if (biases != nullptr)
125       {
126         memcpy(buffer, biases, todo * packing_args.bias_element_size);
127         biases += todo * packing_args.bias_element_size;
128       }
129       else
130       {
131         memset(buffer, 0, vl * packing_args.bias_element_size);
132       }
133 
134       buffer += vl * packing_args.bias_element_size;
135     }
136 
137     // Copy each of the weights in turn
138     unsigned int kx, ky;
139     for (int kindex = 0; packing_args.get_weight_pos(kindex, kx, ky); kindex++)
140     {
141       const auto src_ptr = weights + (kx*ld_weight_row + ky*ld_weight_col + n) * packing_args.weight_element_size;
142       memcpy(buffer, src_ptr, todo * packing_args.weight_element_size);
143       buffer += vl * packing_args.weight_element_size;
144     }
145   }
146 }
147 
148 }  // namespace interleaves
149 }  // namespace depthwise
150 }  // namespace arm_conv
151