xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/directconv2d/nchw/all.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 "src/cpu/kernels/directconv2d/nhwc/neon/impl.h"
26 
27 #include "src/core/NEON/kernels/detail/NEDirectConvolutionDetail.h"
28 #include "src/core/NEON/wrapper/wrapper.h"
29 
30 #include "arm_compute/core/Error.h"
31 #include "arm_compute/core/Helpers.h"
32 #include "arm_compute/core/IAccessWindow.h"
33 #include "arm_compute/core/ITensor.h"
34 #include "arm_compute/core/Types.h"
35 #include "arm_compute/core/Utils.h"
36 #include "src/core/helpers/WindowHelpers.h"
37 
38 #include <algorithm>
39 
40 namespace arm_compute
41 {
42 namespace cpu
43 {
44 namespace kernels
45 {
46 template <typename T>
47 void convolve_nchw(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info);
48 
49 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
neon_fp16_nchw_directconv2d(const Window & window,const ITensor * src,const ITensor * weights,ITensor * dst,const PadStrideInfo & conv_info)50 void neon_fp16_nchw_directconv2d(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info)
51 {
52     convolve_nchw<float16_t>(window, src, weights, dst, conv_info);
53 }
54 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
55 
neon_fp32_nchw_directconv2d(const Window & window,const ITensor * src,const ITensor * weights,ITensor * dst,const PadStrideInfo & conv_info)56 void neon_fp32_nchw_directconv2d(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info)
57 {
58     convolve_nchw<float>(window, src, weights, dst, conv_info);
59 }
60 
61 template <typename T>
convolve_nchw(const Window & window,const ITensor * src,const ITensor * weights,ITensor * dst,const PadStrideInfo & conv_info)62 void convolve_nchw(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info)
63 {
64     ARM_COMPUTE_UNUSED(conv_info);
65 
66     // Declare useful types
67     using vtype       = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>;
68     using vector_type = typename vtype::type;
69     using tag_type    = typename vtype::tag_type;
70 
71     // Scalar quantities
72     const int element_size   = src->info()->element_size();
73     const int input_stride_w = src->info()->strides_in_bytes()[0] / element_size;
74     const int input_stride_h = src->info()->strides_in_bytes()[1] / element_size;
75     const int input_stride_c = src->info()->strides_in_bytes()[2] / element_size;
76     const int input_stride_n = src->info()->strides_in_bytes()[3] / element_size;
77 
78     const int input_dim_w = src->info()->dimension(0);
79     const int input_dim_h = src->info()->dimension(1);
80 
81     const int output_stride_c = dst->info()->strides_in_bytes()[2];
82 
83     const unsigned int kernel_stride_w = weights->info()->strides_in_bytes().x() / element_size;
84     const unsigned int kernel_stride_h = weights->info()->strides_in_bytes().y() / element_size;
85     const unsigned int kernel_stride_c = weights->info()->strides_in_bytes().z() / element_size;
86 
87     const int kernel_dim_w = weights->info()->dimension(0);
88     const int kernel_dim_h = weights->info()->dimension(1);
89 
90     const int conv_pad_top  = conv_info.pad_top();
91     const int conv_pad_left = conv_info.pad_left();
92     const int conv_stride_w = std::get<0>(conv_info.stride());
93     const int conv_stride_h = std::get<1>(conv_info.stride());
94 
95     // Setup input window for the output iterator
96     Window window_out = window;
97     window_out.set(Window::DimZ, Window::Dimension(0, 1, 1));
98 
99     // Setup input window for the weights iterator
100     Window window_w = calculate_max_window(*weights->info(), Steps());
101     window_w.set(Window::DimX, Window::Dimension(0, 1, 1));
102     window_w.set(Window::DimY, Window::Dimension(0, 1, 1));
103     window_w.set(Window::DimZ, Window::Dimension(0, 1, 1));
104 
105     Iterator out(dst, window_out);
106     Iterator wei(weights, window_w);
107 
108     constexpr int num_elems_read_per_iteration = 16 / sizeof(T);
109 
110     execute_window_loop(window_out, [&](const Coordinates & id)
111     {
112         // We are computing the theoretical starting input starting points
113         const int in_w_start_t = static_cast<int>(id.x()) * conv_stride_w - conv_pad_left;
114         const int in_h_start_t = static_cast<int>(id.y()) * conv_stride_h - conv_pad_top;
115         const int in_w_end_t   = in_w_start_t + kernel_dim_w;
116         const int in_h_end_t   = in_h_start_t + kernel_dim_h;
117 
118         // We are computing the valid initial and ending input points by checking the borders
119         const int in_w_start = std::max(in_w_start_t, 0);
120         const int in_h_start = std::max(in_h_start_t, 0);
121         const int in_w_end   = std::min(in_w_end_t, input_dim_w);
122         const int in_h_end   = std::min(in_h_end_t, input_dim_h);
123 
124         // We use the input points to select the valid weight points to use
125         const int wei_w_start = in_w_start - in_w_start_t;
126         const int wei_h_start = in_h_start - in_h_start_t;
127         const int wei_h_end   = kernel_dim_h - (in_h_end_t - in_h_end);
128 
129         const int      index_c_end  = weights->info()->dimension(2);
130         const T *const in_ptr_start = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + id[3] * input_stride_n;
131         execute_window_loop(window_w, [&](const Coordinates & id_w)
132         {
133             const T *const weights_ptr_start = reinterpret_cast<const T *>(wei.ptr());
134             uint8_t       *out_ptr           = out.ptr() + id_w[3] * output_stride_c;
135             T              out_temp          = static_cast<T>(0);
136 
137             for(int index_wei_c = 0, index_in_c = 0; index_wei_c < index_c_end; ++index_wei_c, ++index_in_c)
138             {
139                 const T *const in_ptr_row_0      = in_ptr_start + index_in_c * input_stride_c;
140                 const T *const weights_ptr_row_0 = weights_ptr_start + index_wei_c * kernel_stride_c;
141                 for(int index_wei_h = wei_h_start, index_in_h = in_h_start; index_wei_h < wei_h_end; ++index_wei_h, ++index_in_h)
142                 {
143                     const T    *in_ptr_row      = in_ptr_row_0 + index_in_h * input_stride_h;
144                     const T    *weights_ptr_row = weights_ptr_row_0 + index_wei_h * kernel_stride_h;
145                     int         index_w         = in_w_start;
146                     int         index_wei_w     = wei_w_start;
147                     vector_type out_temp_vec    = wrapper::vdup_n(static_cast<T>(0), tag_type());
148                     for(; index_w <= ((in_w_end - num_elems_read_per_iteration)); index_w += num_elems_read_per_iteration, index_wei_w += num_elems_read_per_iteration)
149                     {
150                         const auto src_vec = wrapper::vloadq(in_ptr_row + index_w * input_stride_w);
151                         const auto w_vec   = wrapper::vloadq(weights_ptr_row + index_wei_w * kernel_stride_w);
152                         out_temp_vec       = wrapper::vmla(out_temp_vec, w_vec, src_vec);
153                     }
154                     out_temp += vreduce(out_temp_vec);
155                     for(; index_w < in_w_end; ++index_w, ++index_wei_w)
156                     {
157                         const auto src_val = *(in_ptr_row + index_w * input_stride_w);
158                         const auto w_val   = *(weights_ptr_row + index_wei_w * kernel_stride_w);
159                         out_temp += src_val * w_val;
160                     }
161                 }
162             }
163             *(reinterpret_cast<T *>(out_ptr)) = out_temp;
164 
165         },
166         wei);
167     },
168     out);
169 }
170 
171 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
172 template void convolve_nchw<float16_t>(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info);
173 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
174 
175 template void convolve_nchw<float>(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info);
176 
177 } // namespace kernels
178 } // namespace cpu
179 } // namespace arm_compute
180