Compute Library
 23.05
depthwise_depthfirst_generic.hpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2021-2023 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 #pragma once
26 
27 #include "depthwise_depthfirst.hpp"
28 
29 namespace arm_conv {
30 namespace depthwise {
31 
32 template <typename TInput, typename TOutput, typename TAccum>
34 {
35  using KernelType = std::function<void(const TInput *const *const, TOutput *const *const, const void *, const void *, const unsigned int, const unsigned int, const TAccum, const TAccum)>;
36 };
37 
38 template <typename TInput, typename TOutput>
39 struct GenericDepthfirstKernelStrategyFunctionType<TInput, TOutput, int32_t>
40 {
41  using KernelType = std::function<void(const TInput *const *const, TOutput *const *const, const void *, const arm_gemm::Requantize32 &, unsigned int, unsigned int)>;
42 };
43 
44 template <typename TInput, typename TWeight, typename TOutput, typename TAccum>
46 {
47  unsigned int m_n_output_points;
48  arm_gemm::VLType m_vl_type;
49  unsigned int m_accumulator_depth_vl;
50 
51  public:
52  GenericDepthfirstKernelStrategy(unsigned int n_output_points, arm_gemm::VLType vl_type, unsigned int accumulator_depth_vl=1)
53  : m_n_output_points(n_output_points), m_vl_type(vl_type), m_accumulator_depth_vl(accumulator_depth_vl)
54  {
55  }
56 
57  virtual ~GenericDepthfirstKernelStrategy() = default;
58 
59  virtual arm_gemm::VLType get_vl_type() const { return m_vl_type; }
60  virtual unsigned int get_accumulator_depth_vl() const { return m_accumulator_depth_vl; }
61  virtual unsigned int get_n_output_points() const { return m_n_output_points; }
62 
64  virtual KernelType get_kernel(void) const = 0;
65 };
66 
67 template <typename TInput,
68  typename TWeight=TInput,
69  typename TOutput=TInput,
70  typename TAccum=typename DefaultTAccum<TInput>::Type,
71  typename OutputStage=typename DefaultOutputStage<TOutput>::Type>
72 class GenericDepthfirstStrategy : public DepthwiseDepthfirstStrategyCommon<TInput, TWeight, TOutput, TAccum, OutputStage>
73 {
74  protected:
76  std::unique_ptr<KernelStrategyType> m_strategy;
77 
78  public:
80  KernelStrategyType *strat, unsigned int n_output_rows, unsigned int n_output_cols,
81  const DepthwiseArgs &args
82  )
83  : DepthwiseDepthfirstStrategyCommon<TInput, TWeight, TOutput, TAccum, OutputStage>(
84  n_output_rows, n_output_cols,
85  args.kernel_rows, args.kernel_cols,
86  args.stride_rows, args.stride_cols
87  ),
88  m_strategy(strat)
89  {
90  }
91 
94 
95  arm_gemm::VLType get_vl_type(void) const override { return m_strategy->get_vl_type(); }
96  unsigned int get_accumulator_depth_vl(void) const override { return m_strategy->get_accumulator_depth_vl(); }
97 
98  size_t get_storage_size(const DepthwiseArgs &args) const override
99  {
100  interleaves::PackingArguments packing_args(
101  this->get_kernel_rows(), this->get_kernel_cols(), sizeof(TWeight),
102  false, sizeof(TAccum), // Don't pack the bias
103  this->get_vl_type(), sizeof(TAccum), this->get_accumulator_depth_vl(),
104  [this] (unsigned int idx, unsigned int &x, unsigned int &y) -> bool
105  { return this->get_kernel_packing_point(idx, x, y); }
106  );
107  return interleaves::get_storage_size_generic(packing_args, args);
108  }
109 
111  const DepthwiseArgs &args, void *buffer,
112  const void *biases, const OutputStage &,
113  const void *weights, size_t ld_weight_col, size_t ld_weight_row
114  ) const override
115  {
116  interleaves::PackingArguments packing_args(
117  this->get_kernel_rows(), this->get_kernel_cols(), sizeof(TWeight),
118  false, sizeof(TAccum), // Don't pack the bias
119  this->get_vl_type(), sizeof(TAccum), this->get_accumulator_depth_vl(),
120  [this] (unsigned int idx, unsigned int &x, unsigned int &y) -> bool
121  { return this->get_kernel_packing_point(idx, x, y); }
122  );
124  packing_args, args, buffer, biases, weights, ld_weight_col, ld_weight_row);
125  }
126 
127  const typename KernelStrategyType::KernelType get_kernel() const { return m_strategy->get_kernel(); }
128 };
129 
130 // Use a templated function to marshal arguments when executing the kernel.
131 template <typename OutputStage> struct DepthwiseDepthfirstGenericKernelCall;
132 
133 template <>
135 {
136  template <typename StratType, typename WorkspaceType, typename TAccum>
137  static void execute(
138  const StratType *strat, const WorkspaceType *ws, const Nothing &,
139  const TAccum *bias, const void *params,
140  const unsigned int n_kernel_points, const unsigned int n_output_channels
141  )
142  {
143  strat->get_kernel()(
144  ws->inptr_array,
145  ws->outptr_array,
146  params, bias,
147  n_kernel_points, n_output_channels,
148  ws->activation_min, ws->activation_max
149  );
150  }
151 };
152 
153 template <>
155 {
156  template <typename StratType, typename WorkspaceType>
157  static void execute(
158  const StratType *strat, const WorkspaceType *ws, const arm_gemm::Requantize32 &qp,
159  const int32_t *, const void *params,
160  const unsigned int n_kernel_points, const unsigned int n_output_channels
161  )
162  {
163  strat->get_kernel()(
164  ws->inptr_array,
165  ws->outptr_array,
166  params, qp,
167  n_kernel_points, n_output_channels
168  );
169  }
170 };
171 
172 
173 /* Workspace Element for an array of input pointers as consumed by the
174  * "Generic" depthwise kernels.
175  */
176 template <typename T>
178 {
179  public:
180  struct Workspace
181  {
182  const T **inptr_array;
183  };
184 
185  template <class OutputStage>
186  static size_t get_element_size(const WorkspaceArgs<IDepthfirstStrategy, OutputStage> &args)
187  {
188  const auto kernel_points = args.depthwise_args.kernel_rows * args.depthwise_args.kernel_cols;
189  return sizeof(T **) * args.strategy->get_input_rows() * args.strategy->get_input_cols() * kernel_points;
190  }
191 
192  template <class WorkspaceType, class OutputStage>
193  static void *initialise(WorkspaceType *ws, void *buffer, const WorkspaceArgs<IDepthfirstStrategy, OutputStage> &args)
194  {
195  ws->inptr_array = reinterpret_cast<const T**>(buffer);
196  return reinterpret_cast<char *>(buffer) + get_element_size(args);
197  }
198 };
199 
200 template <typename TInput, typename TWeight=TInput, typename TOutput=TInput,
201  typename TAccum=typename DefaultTAccum<TInput>::Type,
202  typename OutputStage=typename DefaultOutputStage<TOutput>::Type>
203 class DepthwiseDepthfirstGeneric : public DepthwiseDepthfirstCommon<TInput, TWeight, TOutput, TAccum, OutputStage>
204 {
207  using WorkspaceManager = Workspace<
208  OutputArrayElement<TOutput>,
210  InputBufferElement<TInput>,
211  ActivationsElement<TAccum, OutputStage>
212  >;
213  using WorkingSpace = typename WorkspaceManager::WorkspaceType;
214  const TAccum *m_bias = nullptr;
215 
216  public:
217  DepthwiseDepthfirstGeneric(StratType *const strat, const DepthwiseArgs &args, const OutputStage &os={})
218  : Parent(strat, args, os)
219  {
220  }
221 
224 
226  void *buffer, const void *biases,
227  const void *weights, size_t ld_weight_col, size_t ld_weight_row
228  ) override
229  {
230  Parent::pack_parameters(buffer, biases, weights, ld_weight_col, ld_weight_row);
231  m_bias = reinterpret_cast<const TAccum *>(biases); // Get a copy of the biases
232  depthwise_depthfirst::stash_bias(this->get_output_stage(), m_bias);
233  }
234 
235  size_t get_working_size_per_thread(const unsigned int n_input_channels) const override
236  {
237  DepthwiseArgs args(this->m_args);
238  args.input_channels = n_input_channels;
239  return WorkspaceManager::get_sizeof_workspace(WorkspaceArgs<IDepthfirstStrategy, OutputStage>(this->m_strat.get(), args, this->get_output_stage()));
240  }
241 
242  void initialise_working_space(void *buffer, unsigned int n_input_channels) const override
243  {
244  DepthwiseArgs args(this->m_args);
245  args.input_channels = n_input_channels;
246  return WorkspaceManager::initialise(buffer, WorkspaceArgs<IDepthfirstStrategy, OutputStage>(this->m_strat.get(), args, this->get_output_stage()));
247  }
248 
249  protected:
250  void compute_tile_padded(
251  const DepthwiseArgs &args,
252  unsigned int output_i, unsigned int output_j,
253  unsigned int channel_start, unsigned int channel_end,
255  const TensorSpec<TOutput *> &output,
256  const void *parameters,
257  void *working_space_raw
258  ) const override
259  {
260  // Get the working space
261  WorkingSpace *ws = reinterpret_cast<WorkingSpace *>(working_space_raw);
262 
263  const int ii = static_cast<int>(output_i * args.stride_rows) - args.padding.top;
264  const auto input_pad_top = static_cast<unsigned int>(ii < 0 ? -ii : 0);
265  const auto input_i = static_cast<unsigned int>(ii < 0 ? 0 : ii);
266 
267  const int ij = static_cast<int>(output_j * args.stride_cols) - args.padding.left;
268  const auto input_pad_left = static_cast<unsigned int>(ij < 0 ? -ij : 0);
269  const auto input_j = static_cast<unsigned int>(ij < 0 ? 0 : ij);
270 
271  fill_pointer_array_generic_kernel<const TInput>(
272  ws->inptr_array,
273  this->m_strat->get_output_rows(), this->m_strat->get_output_cols(),
274  args.kernel_rows, args.kernel_cols,
275  args.stride_rows, args.stride_cols,
276  input.base + input_i*input.ld_row + input_j*input.ld_col + channel_start,
277  input.ld_row, input.ld_col,
278  ws->input_buffer,
279  input_pad_top, args.input_rows - input_i,
280  input_pad_left, args.input_cols - input_j
281  );
282 
283  // Compute the output pointer array
284  fill_pointer_array<TOutput>(
285  ws->outptr_array, this->m_strat->get_output_rows(), this->m_strat->get_output_cols(),
286  output.base + output_i*output.ld_row + output_j*output.ld_col + channel_start,
287  output.ld_row, output.ld_col,
288  ws->output_buffer,
289  0, args.output_rows - output_i, // Top padding, # valid rows
290  0, args.output_cols - output_j // Left padding, # valid columns
291  );
292 
293  // Execute the kernel
295  reinterpret_cast<const StratType *>(this->m_strat.get()), ws,
296  this->get_output_stage(), m_bias, parameters,
297  args.kernel_rows * args.kernel_cols,
298  channel_end - channel_start
299  );
300  }
301 };
302 
303 } // namespace depthwise
304 } // namespace arm_conv
GenericDepthfirstStrategy(KernelStrategyType *strat, unsigned int n_output_rows, unsigned int n_output_cols, const DepthwiseArgs &args)
void initialise_working_space(void *buffer, unsigned int n_input_channels) const override
typename GenericDepthfirstKernelStrategyFunctionType< TInput, TOutput, TAccum >::KernelType KernelType
static void execute(const StratType *strat, const WorkspaceType *ws, const Nothing &, const TAccum *bias, const void *params, const unsigned int n_kernel_points, const unsigned int n_output_channels)
static void * initialise(WorkspaceType *ws, void *buffer, const WorkspaceArgs< IDepthfirstStrategy, OutputStage > &args)
size_t get_storage_size_generic(const PackingArguments &packing_args, const DepthwiseArgs &args)
Definition: generic.cpp:45
GenericDepthfirstKernelStrategy(unsigned int n_output_points, arm_gemm::VLType vl_type, unsigned int accumulator_depth_vl=1)
void pack_parameters(void *_buffer, const int32_t *biases, const T *weights, size_t ld_weight_col, size_t ld_weight_row, const DepthwiseArgs &args, const arm_gemm::Requantize32 &qp, const arm_gemm::VLType vl_type, const unsigned int accumulator_depth_vl)
DepthwiseDepthfirstGeneric(StratType *const strat, const DepthwiseArgs &args, const OutputStage &os={})
std::function< void(const TInput *const *const, TOutput *const *const, const void *, const arm_gemm::Requantize32 &, unsigned int, unsigned int)> KernelType
static size_t get_element_size(const WorkspaceArgs< IDepthfirstStrategy, OutputStage > &args)
std::unique_ptr< ParametersLibrary > parameters
Definition: Framework.cpp:46
static void execute(const StratType *strat, const WorkspaceType *ws, const arm_gemm::Requantize32 &qp, const int32_t *, const void *params, const unsigned int n_kernel_points, const unsigned int n_output_channels)
void pack_parameters(const DepthwiseArgs &args, void *buffer, const void *biases, const OutputStage &, const void *weights, size_t ld_weight_col, size_t ld_weight_row) const override
std::function< void(const TInput *const *const, TOutput *const *const, const void *, const void *, const unsigned int, const unsigned int, const TAccum, const TAccum)> KernelType
void pack_parameters(void *buffer, const void *biases, const void *weights, size_t ld_weight_col, size_t ld_weight_row) override
const KernelStrategyType::KernelType get_kernel() const
template UniqueDepthwiseCommon< float > depthwise(const DepthwiseArgs &, const Nothing &)
size_t get_working_size_per_thread(const unsigned int n_input_channels) const override
size_t get_storage_size(const DepthwiseArgs &args) const override
void 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)
Definition: generic.cpp:67
const int32_t * bias