Compute Library
 21.11
ClGemmLowpOffsetContributionKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017-2021 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  */
25 
29 #include "arm_compute/core/Utils.h"
31 
33 
34 #include "support/Cast.h"
35 #include "support/StringSupport.h"
36 
37 namespace arm_compute
38 {
39 namespace opencl
40 {
41 namespace kernels
42 {
43 namespace
44 {
45 Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
46  int32_t a_offset, int32_t b_offset)
47 {
49 
50  if(bias != nullptr)
51  {
53  ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
54  ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != bias->dimension(0));
55  }
56 
57  // If a_offset == 0, vector_sum_col can be a nullptr
58  if(a_offset != 0)
59  {
61  ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != mm_result->dimension(0));
62  }
63 
64  // If b_offset == 0, vector_sum_row can be a nullptr
65  if(b_offset != 0)
66  {
68 
69  // Check if input is a 3D reinterpretation
70  const bool reinterpret_as_3d = mm_result->num_dimensions() > 1 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x();
71 
72  // Validate input
73  ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (mm_result->dimension(1) * mm_result->dimension(2)));
74  ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != mm_result->dimension(1));
75 
76  TensorShape output_shape = mm_result->tensor_shape();
77  if(output_shape.num_dimensions() > 1)
78  {
79  const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2;
80 
81  TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
82  vector_sum_row_shape.collapse_from(1);
83  output_shape.collapse_from(output_batch_idx);
84 
85  ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[output_batch_idx],
86  "mm_result tensor must have the same number of batches of output tensor");
87 
88  if(a_offset != 0)
89  {
90  TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape();
91  vector_sum_col_shape.collapse_from(1);
92 
93  ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1],
94  "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");
95  }
96  }
97  }
98 
99  return Status{};
100 }
101 } // namespace
102 
104 {
106 }
107 
109  const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
110  int32_t k, int32_t a_offset, int32_t b_offset)
111 {
112  // Perform validate step
113  ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result);
114  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, a_offset, b_offset));
115 
116  auto padding_info = get_padding_info({ mm_result, vector_sum_col, vector_sum_row, bias });
117 
118  // Check if input is a 3D reinterpretation
119  const bool reinterpret_as_3d = vector_sum_row != nullptr
120  && mm_result->num_dimensions() > 1
121  && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x();
122 
123  const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, mm_result->dimension(0));
124 
125  // Set the arguments to pass at compile time
126  CLBuildOptions build_opts;
127  build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
128  build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(mm_result->dimension(0) % num_elems_processed_per_iteration));
129 
130  // If a_offset == 0, vector_sum_col can be a nullptr
131  if(a_offset != 0)
132  {
133  build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset));
134  build_opts.add_option_if(vector_sum_col->tensor_shape().num_dimensions() > 1, "-DSUM_COL_HAS_BATCHES");
135  }
136  // If b_offset == 0, vector_sum_row can be a nullptr
137  build_opts.add_option_if(b_offset != 0, "-DB_OFFSET=" + support::cpp11::to_string(b_offset));
138  build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(a_offset * b_offset * k));
139  build_opts.add_option_if(reinterpret_as_3d, "-DHEIGHT_INPUT3D=" + support::cpp11::to_string(mm_result->dimension(1)));
140  build_opts.add_option_if(reinterpret_as_3d, "-DDEPTH_INPUT3D=" + support::cpp11::to_string(mm_result->dimension(2)));
141  build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
142 
143  std::string kernel_name("gemmlowp_offset_contribution");
144 
145  // Create kernel
146  _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
147 
148  // Configure kernel window
149  Window win = calculate_max_window(*mm_result, Steps(num_elems_processed_per_iteration));
150  IClKernel::configure_internal(win);
151 
152  // Set config_id for enabling LWS tuning
153  _config_id = kernel_name + "_";
154  _config_id += support::cpp11::to_string(mm_result->dimension(0));
155  _config_id += "_";
156  _config_id += support::cpp11::to_string(mm_result->dimension(1));
157  _config_id += "_";
158  _config_id += support::cpp11::to_string(mm_result->dimension(2));
159 
161 }
162 
163 Status ClGemmLowpOffsetContributionKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
164  int32_t a_offset, int32_t b_offset)
165 {
166  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, a_offset, b_offset));
167  return Status{};
168 }
169 
170 void ClGemmLowpOffsetContributionKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
171 {
174 
175  const auto vector_sum_col = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_VEC_COL_SUM));
176  const auto vector_sum_row = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_VEC_ROW_SUM));
177  const auto bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_BIAS));
178  const auto mm_result = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_SRC_DST));
179 
181  Window slice = collapsed.first_slice_window_3D();
182 
183  // Set window for vector_sum_col
184  Window win_vector_sum_col = slice;
185  win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
186  win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
187 
188  // Set window for vector_sum_row
189  Window win_vector_sum_row = slice;
190  win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
191  win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
192  win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
193 
194  Window biases_slice = slice;
195  biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
196  biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
197 
198  do
199  {
200  unsigned int idx = 0;
201  add_3D_tensor_argument(idx, mm_result, slice);
202  add_2D_tensor_argument_if((vector_sum_col != nullptr), idx, vector_sum_col, win_vector_sum_col);
203  add_2D_tensor_argument_if((vector_sum_row != nullptr), idx, vector_sum_row, win_vector_sum_row);
204  add_1D_tensor_argument_if((bias != nullptr), idx, bias, biases_slice);
205 
206  enqueue(queue, *this, slice, lws_hint());
207  }
208  while(collapsed.slide_window_slice_3D(slice));
209 }
210 } // namespace kernels
211 } // namespace opencl
212 } // namespace arm_compute
void add_1D_tensor_argument_if(bool cond, unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 1D tensor&#39;s parameters to the object&#39;s kernel&#39;s arguments starting from the index idx ...
Definition: ICLKernel.h:177
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
void add_2D_tensor_argument_if(bool cond, unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 2D tensor&#39;s parameters to the object&#39;s kernel&#39;s arguments starting from the index idx ...
Definition: ICLKernel.h:201
void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue...
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
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:32
cl::NDRange lws_hint() const
Return the Local-Workgroup-Size hint.
Definition: ICLKernel.h:318
#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_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Describe one of the image&#39;s dimensions with a start, end and step.
Definition: Window.h:77
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
void add_3D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 3D tensor&#39;s parameters to the object&#39;s kernel&#39;s arguments starting from the index idx...
Definition: ICLKernel.h:214
Copyright (c) 2017-2021 Arm Limited.
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.
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:87
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
Definition: ITensorPack.cpp:54
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:391
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
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.
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
unsigned int num_elems_processed_per_iteration
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
Elementeise CL kernel type.
Definition: CLTypes.h:84
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:915
bool has_padding_changed(const std::unordered_map< const ITensorInfo *, PaddingSize > &padding_map)
Check if the previously stored padding info has changed after configuring a kernel.
Definition: Utils.cpp:533
CLCompileContext class.
void configure(const CLCompileContext &compile_context, const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, int32_t k, int32_t a_offset, int32_t b_offset)
Initialise the kernel&#39;s input and output.
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
ITensor * get_tensor(int id)
Get tensor of a given id from the pac.
Definition: ITensorPack.cpp:64
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:143
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:788
std::unordered_map< const ITensorInfo *, PaddingSize > get_padding_info(std::initializer_list< const ITensorInfo *> infos)
Stores padding information before configuring a kernel.
Definition: Utils.cpp:518
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
Tensor packing service.
Definition: ITensorPack.h:39
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157
unsigned int adjust_vec_size(unsigned int vec_size, size_t dim0)
Returns the adjusted vector size in case it is less than the input&#39;s first dimension, getting rounded down to its closest valid vector size.
Definition: Utils.h:1171
T y() const
Alias to access the size of the second dimension.
Definition: Dimensions.h:92
std::string kernel_name
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:201
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)