xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/CpuGemmInterleave4x4Kernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1  /*
2   * Copyright (c) 2016-2021 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  #include "src/cpu/kernels/CpuGemmInterleave4x4Kernel.h"
25  
26  #include "arm_compute/core/ITensor.h"
27  #include "arm_compute/core/Validate.h"
28  #include "arm_compute/core/Window.h"
29  #include "arm_compute/core/utils/misc/ShapeCalculator.h"
30  #include "src/core/helpers/AutoConfiguration.h"
31  #include "src/core/helpers/WindowHelpers.h"
32  
33  #include <arm_neon.h>
34  
35  namespace arm_compute
36  {
37  namespace cpu
38  {
39  namespace kernels
40  {
41  using namespace arm_compute::misc::shape_calculator;
42  
configure(const ITensorInfo * src,ITensorInfo * dst)43  void CpuGemmInterleave4x4Kernel::configure(const ITensorInfo *src, ITensorInfo *dst)
44  {
45      ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
46  
47      // dst auto inizialitation if not yet initialized
48      auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_interleaved_shape(*src)));
49  
50      // Perform validate step
51      ARM_COMPUTE_ERROR_THROW_ON(CpuGemmInterleave4x4Kernel::validate(src, dst));
52  
53      Window win = calculate_max_window(*src, Steps(1, 4));
54      ICPPKernel::configure(win);
55  }
56  
validate(const ITensorInfo * src,const ITensorInfo * dst)57  Status CpuGemmInterleave4x4Kernel::validate(const ITensorInfo *src, const ITensorInfo *dst)
58  {
59      ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
60      //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src) is not needed here as this kernel doesn't use CPU FP16 instructions.
61      ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
62  
63      if(dst->total_size() != 0)
64      {
65          const TensorShape dst_shape = compute_interleaved_shape(*src);
66          ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), dst_shape);
67          ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
68          ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
69      }
70  
71      return Status{};
72  }
73  
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)74  void CpuGemmInterleave4x4Kernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
75  {
76      ARM_COMPUTE_UNUSED(info);
77      ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
78      ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
79      ARM_COMPUTE_ERROR_ON(tensors.empty());
80      /*
81      *  This kernel puts the values in a 4x4 block of Matrix A on the same row (Interleaved values)
82      *         |a00 a01 a02 a03|
83      *         |a10 a11 a12 a13|
84      *         |a20 a21 a22 a23| = | a00 a10 a20 a30 || a01 a11 a21 a31 || a02 a12 a22 a32 || a03 a13 a23 a33 |
85      *         |a30 a31 a32 a33|
86      *
87      *         After this operation, the dst matrix will have the following shape: [ height * 4, ceil(width / 4.0f) ]
88      */
89      const ITensor *src = tensors.get_const_tensor(TensorType::ACL_SRC);
90      ITensor       *dst = tensors.get_tensor(TensorType::ACL_DST);
91  
92      const size_t window_start_x = window.x().start();
93      const size_t window_end_x   = window.x().end();
94  
95      const size_t in_height = src->info()->dimension(1);
96      const size_t in_stride = src->info()->strides_in_bytes()[1];
97  
98      const size_t partial_y = in_height % 4;
99  
100      const size_t element_size = src->info()->element_size();
101  
102      // Set window for the src tensor
103      Window win = window;
104      win.set(Window::DimX, Window::Dimension(0, 1, 1));
105  
106      // Set window for the dst tensor
107      Window win_out(window);
108      win_out.set(Window::DimX, Window::Dimension(0, 1, 1));
109      win_out.scale(Window::DimY, 0.25f);
110  
111      Iterator in(src, win);
112      Iterator out(dst, win_out);
113  
114      execute_window_loop(win, [&](const Coordinates & id)
115      {
116          if(id.y() + 4 <= static_cast<int>(in_height))
117          {
118              for(size_t x = window_start_x; x < window_end_x; ++x)
119              {
120                  std::memcpy(out.ptr() + (x * 4 + 0) * element_size, (in.ptr() + 0 * in_stride) + x * element_size, element_size);
121                  std::memcpy(out.ptr() + (x * 4 + 1) * element_size, (in.ptr() + 1 * in_stride) + x * element_size, element_size);
122                  std::memcpy(out.ptr() + (x * 4 + 2) * element_size, (in.ptr() + 2 * in_stride) + x * element_size, element_size);
123                  std::memcpy(out.ptr() + (x * 4 + 3) * element_size, (in.ptr() + 3 * in_stride) + x * element_size, element_size);
124              }
125          }
126          else
127          {
128              for(size_t x = window_start_x; x < window_end_x; ++x)
129              {
130                  size_t y = 0;
131                  for(; y < partial_y; ++y)
132                  {
133                      std::memcpy(out.ptr() + (x * 4 + y) * element_size, (in.ptr() + y * in_stride) + x * element_size, element_size);
134                  }
135                  for(; y < 4; ++y)
136                  {
137                      std::memset(out.ptr() + (x * 4 + y) * element_size, 0, element_size);
138                  }
139              }
140          }
141      },
142      in, out);
143  }
144  
name() const145  const char *CpuGemmInterleave4x4Kernel::name() const
146  {
147      return "CpuGemmInterleave4x4Kernel";
148  }
149  } // namespace kernels
150  } // namespace cpu
151  } // namespace arm_compute
152