Compute Library
 20.08
CLGEMMLowpOffsetContributionKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017-2020 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  */
25 
28 #include "arm_compute/core/Error.h"
31 #include "arm_compute/core/Types.h"
32 #include "arm_compute/core/Utils.h"
35 #include "support/StringSupport.h"
36 
37 #include <cstddef>
38 #include <cstdint>
39 
40 using namespace arm_compute;
41 
42 namespace arm_compute
43 {
44 class Coordinates;
45 } // namespace arm_compute
46 
47 namespace
48 {
49 Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
50  int32_t a_offset, int32_t b_offset)
51 {
53 
54  if(bias != nullptr)
55  {
57  ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
58  ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != bias->dimension(0));
59  }
60 
61  // If a_offset == 0, vector_sum_col can be a nullptr
62  if(a_offset != 0)
63  {
65  ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != mm_result->dimension(0));
66  }
67 
68  // If b_offset == 0, vector_sum_row can be a nullptr
69  if(b_offset != 0)
70  {
72 
73  // Check if input is a 3D reinterpretation
74  const bool reinterpret_as_3d = mm_result->num_dimensions() > 1 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x();
75 
76  // Validate input
77  ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (mm_result->dimension(1) * mm_result->dimension(2)));
78  ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != mm_result->dimension(1));
79 
80  TensorShape output_shape = mm_result->tensor_shape();
82  {
83  const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2;
84 
85  TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
86  vector_sum_row_shape.collapse_from(1);
87  output_shape.collapse_from(output_batch_idx);
88 
89  ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[output_batch_idx],
90  "mm_result tensor must have the same number of batches of output tensor");
91 
92  if(a_offset != 0)
93  {
94  TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape();
95  vector_sum_col_shape.collapse_from(1);
96 
97  ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1],
98  "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1");
99  }
100  }
101  }
102 
103  return Status{};
104 }
105 
106 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *mm_result, ITensorInfo *vector_sum_col, ITensorInfo *vector_sum_row, ITensorInfo *bias,
107  int32_t a_offset, int32_t b_offset)
108 {
109  constexpr unsigned int num_elems_processed_per_iteration = 4;
110  bool window_changed = false;
111 
112  // Configure kernel window
114 
115  AccessWindowHorizontal mm_result_access(mm_result, 0, num_elems_processed_per_iteration);
116  window_changed = window_changed || update_window_and_padding(win, mm_result_access);
117 
118  if(a_offset != 0)
119  {
120  AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration);
121  window_changed = window_changed || update_window_and_padding(win, vector_sum_col_access);
122  }
123  if(b_offset != 0)
124  {
125  AccessWindowStatic vector_sum_row_access(vector_sum_row, 0, 0, vector_sum_row->dimension(0), 0); // NOLINT
126  window_changed = window_changed || update_window_and_padding(win, vector_sum_row_access);
127  }
128 
129  if(bias != nullptr)
130  {
131  AccessWindowStatic bias_access(bias, 0, 0, ceil_to_multiple(bias->dimension(0), num_elems_processed_per_iteration), bias->tensor_shape()[1]);
132  window_changed = window_changed || update_window_and_padding(win, bias_access);
133  }
134 
135  Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
136  return std::make_pair(err, win);
137 }
138 } // namespace
139 
141  : _vector_sum_col(nullptr), _vector_sum_row(nullptr), _mm_result(nullptr), _bias(nullptr)
142 {
143 }
144 
145 void CLGEMMLowpOffsetContributionKernel::configure(ICLTensor *mm_result, const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias, int32_t k, int32_t a_offset,
146  int32_t b_offset)
147 {
148  configure(CLKernelLibrary::get().get_compile_context(), mm_result, vector_sum_col, vector_sum_row, bias, k, a_offset, b_offset);
149 }
150 
151 void CLGEMMLowpOffsetContributionKernel::configure(const CLCompileContext &compile_context, ICLTensor *mm_result, const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias,
152  int32_t k, int32_t a_offset,
153  int32_t b_offset)
154 {
155  // Perform validate step
156  ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result);
158  vector_sum_col != nullptr ? vector_sum_col->info() : nullptr,
159  vector_sum_row != nullptr ? vector_sum_row->info() : nullptr,
160  bias != nullptr ? bias->info() : nullptr,
161  a_offset, b_offset)); // NOLINT
162 
163  _vector_sum_col = vector_sum_col;
164  _vector_sum_row = vector_sum_row;
165  _mm_result = mm_result;
166  _bias = bias;
167 
168  // Check if input is a 3D reinterpretation
169  const bool reinterpret_as_3d = vector_sum_row != nullptr
170  && mm_result->info()->num_dimensions() > 1
171  && mm_result->info()->tensor_shape().y() != vector_sum_row->info()->tensor_shape().x();
172 
173  // Set the arguments to pass at compile time
174  CLBuildOptions build_opts;
175 
176  // If a_offset == 0, vector_sum_col can be a nullptr
177  if(a_offset != 0)
178  {
179  build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset));
180  build_opts.add_option_if(vector_sum_col->info()->tensor_shape().num_dimensions() > 1, "-DSUM_COL_HAS_BATCHES");
181  }
182  // If b_offset == 0, vector_sum_row can be a nullptr
183  build_opts.add_option_if(b_offset != 0, "-DB_OFFSET=" + support::cpp11::to_string(b_offset));
184  build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(a_offset * b_offset * k));
185  build_opts.add_option_if(reinterpret_as_3d, "-DHEIGHT_INPUT3D=" + support::cpp11::to_string(mm_result->info()->dimension(1)));
186  build_opts.add_option_if(reinterpret_as_3d, "-DDEPTH_INPUT3D=" + support::cpp11::to_string(mm_result->info()->dimension(2)));
187  build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
188 
189  std::string kernel_name("gemmlowp_offset_contribution");
190 
191  // Create kernel
192  _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
193 
194  // Configure kernel window
195  auto win_config = validate_and_configure_window(mm_result->info(),
196  vector_sum_col != nullptr ? vector_sum_col->info() : nullptr,
197  vector_sum_row != nullptr ? vector_sum_row->info() : nullptr,
198  bias != nullptr ? bias->info() : nullptr,
199  a_offset, b_offset); // NOLINT
200  ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
201  ICLKernel::configure_internal(win_config.second);
202 
203  // Set config_id for enabling LWS tuning
204  _config_id = kernel_name + "_";
205  _config_id += support::cpp11::to_string(mm_result->info()->dimension(0));
206  _config_id += "_";
207  _config_id += support::cpp11::to_string(mm_result->info()->dimension(1));
208  _config_id += "_";
209  _config_id += support::cpp11::to_string(mm_result->info()->dimension(2));
210 }
211 
212 Status CLGEMMLowpOffsetContributionKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
213  int32_t a_offset, int32_t b_offset)
214 {
215  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, a_offset, b_offset));
216  ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(mm_result->clone().get(),
217  vector_sum_col != nullptr ? vector_sum_col->clone().get() : nullptr,
218  vector_sum_row != nullptr ? vector_sum_row->clone().get() : nullptr,
219  bias != nullptr ? bias->clone().get() : nullptr,
220  a_offset, b_offset)
221  .first); // NOLINT
222 
223  return Status{};
224 }
225 
226 void CLGEMMLowpOffsetContributionKernel::run(const Window &window, cl::CommandQueue &queue)
227 {
230 
232  Window slice = collapsed.first_slice_window_3D();
233 
234  // Set window for vector_sum_col
235  Window win_vector_sum_col = slice;
236  win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
237  win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
238 
239  // Set window for vector_sum_row
240  Window win_vector_sum_row = slice;
241  win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
242  win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
243  win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
244 
245  Window biases_slice = slice;
246  biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
247  biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
248 
249  do
250  {
251  unsigned int idx = 0;
252  add_3D_tensor_argument(idx, _mm_result, slice);
253  add_2D_tensor_argument_if((_vector_sum_col != nullptr), idx, _vector_sum_col, win_vector_sum_col);
254  add_2D_tensor_argument_if((_vector_sum_row != nullptr), idx, _vector_sum_row, win_vector_sum_row);
255  add_1D_tensor_argument_if((_bias != nullptr), idx, _bias, biases_slice);
256 
257  enqueue(queue, *this, slice, lws_hint());
258  }
259  while(collapsed.slide_window_slice_3D(slice));
260 }
void add_1D_tensor_argument_if(bool cond, unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 1D tensor's parameters to the object's kernel's arguments starting from the index idx ...
Definition: ICLKernel.h:122
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
Shape of a tensor.
Definition: TensorShape.h:39
void configure(ICLTensor *mm_result, const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias, int32_t k, int32_t a_offset, int32_t b_offset)
Initialise the kernel's input and output.
void add_2D_tensor_argument_if(bool cond, unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 2D tensor's parameters to the object's kernel's arguments starting from the index idx ...
Definition: ICLKernel.h:146
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
static Status validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, int32_t a_offset, int32_t b_offset)
Static function to check if given info will lead to a valid configuration of CLGEMMLowpOffsetContribu...
void enqueue(cl::CommandQueue &queue, ICLKernel &kernel, const Window &window, const cl::NDRange &lws_hint=CLKernelLibrary::get().default_ndrange(), bool use_dummy_work_items=false)
Add the kernel to the command queue with the given window.
Definition: ICLKernel.cpp:39
const StringSet & options() const
Gets the current options list set.
cl::NDRange lws_hint() const
Return the Local-Workgroup-Size hint.
Definition: ICLKernel.h:263
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
std::string to_string(T &&value)
Convert integer and float values to string.
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:792
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
Store the tensor's metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Describe one of the image's dimensions with a start, end and step.
Definition: Window.h:75
Status class.
Definition: Error.h:52
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps=Steps(), bool skip_border=false, BorderSize border_size=BorderSize())
Calculate the maximum window for a given tensor shape and border setting.
Definition: Helpers.cpp:28
void add_3D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 3D tensor's parameters to the object's kernel's arguments starting from the index idx.
Definition: ICLKernel.h:159
Copyright (c) 2017-2020 Arm Limited.
ITensorInfo * info() const override
Interface to be implemented by the child class to return the tensor's metadata.
Definition: Tensor.cpp:33
Implementation of a static rectangular access pattern.
1 channel, 1 S32 per channel
void add_option(std::string option)
Adds option to the existing build option list.
T x() const
Alias to access the size of the first dimension.
Definition: Dimensions.h:81
cl::Kernel create_kernel(const CLCompileContext &ctx, const std::string &kernel_name, const std::set< std::string > &build_opts=std::set< std::string >())
Creates an opencl kernel using a compile context.
Definition: CLHelpers.cpp:403
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
bool update_window_and_padding(Window &win, Ts &&... patterns)
Update window and padding size for each of the access patterns.
Definition: Helpers.h:437
void collapse_from(size_t start)
Collapse dimensions starting from a given point.
Definition: Dimensions.h:162
Window collapse_if_possible(const Window &full_window, size_t first, size_t last, bool *has_collapsed=nullptr) const
Collapse the dimensions between first and last if possible.
Definition: Window.inl:68
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
auto ceil_to_multiple(S value, T divisor) -> decltype(((value+divisor - 1)/divisor) *divisor)
Computes the smallest number larger or equal to value that is a multiple of divisor.
Definition: Utils.h:67
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
Coordinates of an item.
Definition: Coordinates.h:37
Implementation of a row access pattern.
std::string kernel_name
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
void add_option_if(bool cond, std::string option)
Adds option if a given condition is true;.
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
bool slide_window_slice_3D(Window &slice) const
Slide the passed 3D window slice.
Definition: Window.h:333
CLCompileContext class.
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
#define ARM_COMPUTE_CREATE_ERROR(error_code, msg)
Creates an error with a given message.
Definition: Error.h:159
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Definition: Window.h:47
unsigned int num_dimensions() const
Returns the effective dimensionality of the tensor.
Definition: Dimensions.h:122
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
unsigned int num_elems_processed_per_iteration
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
void run(const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue.
T y() const
Alias to access the size of the second dimension.
Definition: Dimensions.h:86
Window first_slice_window_3D() const
First 3D slice of the window.
Definition: Window.h:289
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:205
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:941
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)