Compute Library
 21.02
CLGEMMLowpOffsetContributionOutputStageKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2018-2020 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 
30 #include "arm_compute/core/Utils.h"
35 #include "support/StringSupport.h"
36 
37 namespace arm_compute
38 {
39 namespace
40 {
41 Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output,
42  int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
43 {
45 
46  if(bias != nullptr)
47  {
49  ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
50  ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != bias->dimension(0));
51  }
52 
54  ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
56  ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
57  if(output_stage.is_quantized_per_channel)
58  {
59  ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != output_shifts->dimension(0));
60  ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != output_multipliers->dimension(0));
61  }
62 
63  // If a_offset == 0, vector_sum_col can be a nullptr
64  if(a_offset != 0)
65  {
67  ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != mm_result->dimension(0));
68  }
69 
70  // If b_offset == 0, vector_sum_row can be a nullptr
71  if(b_offset != 0)
72  {
74 
75  // Check if input is a 3D reinterpretation
76  const bool reinterpret_as_3d = mm_result->num_dimensions() > 1 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x();
77 
78  // Validate input
79  ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (mm_result->dimension(1) * mm_result->dimension(2)));
80  ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != mm_result->dimension(1));
81 
82  TensorShape output_shape = mm_result->tensor_shape();
83  if(output_shape.num_dimensions() > 1)
84  {
85  const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2;
86 
87  TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
88  vector_sum_row_shape.collapse_from(1);
89  output_shape.collapse_from(output_batch_idx);
90 
91  ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[output_batch_idx],
92  "mm_result tensor must have the same number of batches of output tensor");
93 
94  if(a_offset != 0)
95  {
96  TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape();
97  vector_sum_col_shape.collapse_from(1);
98 
99  ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1],
100  "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");
101  }
102  }
103  }
104 
106  // Checks performed when output is configured
107  if((output != nullptr) && (output->total_size() != 0))
108  {
109  ARM_COMPUTE_RETURN_ERROR_ON(output_stage.output_data_type != output->data_type());
112  }
113 
114  ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound);
115  ARM_COMPUTE_RETURN_ERROR_ON_MSG(output_stage.gemmlowp_multipliers.size() != output_stage.gemmlowp_shifts.size(), "per channel quantization info is incorrect");
116 
117  return Status{};
118 }
119 } // namespace
120 
122  : _mm_result(nullptr),
123  _vector_sum_col(nullptr),
124  _vector_sum_row(nullptr),
125  _bias(nullptr),
126  _output(nullptr),
127  _output_multipliers(nullptr),
128  _output_shifts(nullptr),
129  _is_quantized_per_channel(false)
130 {
131 }
132 
133 void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const ICLTensor *mm_result, const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias, ICLTensor *output,
134  int32_t k, int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage,
135  const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
136 {
137  configure(CLKernelLibrary::get().get_compile_context(), mm_result, vector_sum_col, vector_sum_row, bias, output, k, a_offset, b_offset, output_stage, output_multipliers, output_shifts);
138 }
139 
140 void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const CLCompileContext &compile_context, const ICLTensor *mm_result, const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row,
141  const ICLTensor *bias, ICLTensor *output,
142  int32_t k, int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage,
143  const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
144 {
145  // Perform validate step
146  ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result, output, output_multipliers, output_shifts);
148  vector_sum_col != nullptr ? vector_sum_col->info() : nullptr,
149  vector_sum_row != nullptr ? vector_sum_row->info() : nullptr,
150  bias != nullptr ? bias->info() : nullptr,
151  output->info(),
152  a_offset, b_offset, output_stage,
153  output_multipliers->info(), output_shifts->info())); // NOLINT
154 
155  auto padding_info = get_padding_info({ mm_result, vector_sum_col, vector_sum_row, bias, output, output_multipliers, output_shifts });
156 
157  const int min = output_stage.gemmlowp_min_bound;
158  const int max = output_stage.gemmlowp_max_bound;
159 
160  _vector_sum_col = vector_sum_col;
161  _vector_sum_row = vector_sum_row;
162  _mm_result = mm_result;
163  _bias = bias;
164  _output = output;
165  _output_multipliers = output_multipliers;
166  _output_shifts = output_shifts;
167  _is_quantized_per_channel = output_stage.is_quantized_per_channel;
168 
169  // Check if input is a 3D reinterpretation
170  const bool reinterpret_as_3d = vector_sum_row != nullptr
171  && mm_result->info()->num_dimensions() > 1
172  && mm_result->info()->tensor_shape().y() != vector_sum_row->info()->tensor_shape().x();
173 
174  // Auto initialize the output
175  auto_init_if_empty(*output->info(), mm_result->info()->clone()->set_data_type(output_stage.output_data_type));
176 
177  const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, mm_result->info()->dimension(0));
178 
179  // Set the arguments to pass at compile time
180  CLBuildOptions build_opts;
181  build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
182  build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(mm_result->info()->dimension(0) % num_elems_processed_per_iteration));
183 
184  // If a_offset == 0, vector_sum_col can be a nullptr
185  if(a_offset != 0)
186  {
187  build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset));
188  build_opts.add_option_if(vector_sum_col->info()->tensor_shape().num_dimensions() > 1, "-DSUM_COL_HAS_BATCHES");
189  }
190  // If b_offset == 0, vector_sum_row can be a nullptr
191  build_opts.add_option_if(b_offset != 0, "-DB_OFFSET=" + support::cpp11::to_string(b_offset));
192  build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(a_offset * b_offset * k));
193  build_opts.add_option_if(reinterpret_as_3d, "-DHEIGHT_INPUT3D=" + support::cpp11::to_string(mm_result->info()->dimension(1)));
194  build_opts.add_option_if(reinterpret_as_3d, "-DDEPTH_INPUT3D=" + support::cpp11::to_string(mm_result->info()->dimension(2)));
195  build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
196  build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(output_stage.gemmlowp_offset));
197  build_opts.add_option("-DRESULT_MULTIPLIER=" + support::cpp11::to_string(output_stage.gemmlowp_multipliers[0]));
198  build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(output_stage.gemmlowp_shifts[0]));
199  build_opts.add_option_if(_is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION");
200  build_opts.add_option("-DOUTPUT_DATA_TYPE=" + get_cl_type_from_data_type(output->info()->data_type()));
201 
202  PixelValue min_val{};
203  PixelValue max_val{};
204  std::tie(min_val, max_val) = get_min_max(output->info()->data_type());
205  build_opts.add_option_if((min > min_val.get<int32_t>()), "-DMIN_BOUND=" + support::cpp11::to_string(min));
206  build_opts.add_option_if((max < max_val.get<int32_t>()), "-DMAX_BOUND=" + support::cpp11::to_string(max));
207 
208  std::string kernel_name("gemmlowp_offset_contribution");
209  kernel_name += "_" + string_from_gemmlowp_output_stage(output_stage.type);
210 
211  // Create kernel
212  _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
213 
214  // Configure kernel window
215  Window win = calculate_max_window(*mm_result->info(), Steps(num_elems_processed_per_iteration));
216  ICLKernel::configure_internal(win);
217 
218  // Set config_id for enabling LWS tuning
219  _config_id = kernel_name + "_";
220  _config_id += support::cpp11::to_string(mm_result->info()->dimension(0));
221  _config_id += "_";
222  _config_id += support::cpp11::to_string(mm_result->info()->dimension(1));
223  _config_id += "_";
224  _config_id += support::cpp11::to_string(mm_result->info()->dimension(2));
225 
227 }
228 
229 Status CLGEMMLowpOffsetContributionOutputStageKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
230  const ITensorInfo *output, int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage,
231  const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
232 {
233  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, output, a_offset, b_offset, output_stage, output_multipliers, output_shifts));
234  return Status{};
235 }
236 
238 {
241 
243  Window slice = collapsed.first_slice_window_3D();
244 
245  // Set window for vector_sum_col
246  Window win_vector_sum_col = slice;
247  win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
248  win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
249 
250  // Set window for vector_sum_row
251  Window win_vector_sum_row = slice;
252  win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
253  win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
254  win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
255 
256  Window biases_slice = slice;
257  biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
258  biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
259 
260  do
261  {
262  unsigned int idx = 0;
263  add_3D_tensor_argument(idx, _mm_result, slice);
264  add_2D_tensor_argument_if((_vector_sum_col != nullptr), idx, _vector_sum_col, win_vector_sum_col);
265  add_2D_tensor_argument_if((_vector_sum_row != nullptr), idx, _vector_sum_row, win_vector_sum_row);
266  add_1D_tensor_argument_if((_bias != nullptr), idx, _bias, biases_slice);
267  add_3D_tensor_argument(idx, _output, slice);
268  add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_multipliers, biases_slice);
269  add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_shifts, biases_slice);
270  enqueue(queue, *this, slice, lws_hint());
271  }
272  while(collapsed.slide_window_slice_3D(slice));
273 }
274 } // 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:135
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
Class describing the value of a pixel for any image format.
Definition: PixelValue.h:34
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:159
void enqueue(IGCKernel &kernel, const Window &window, const gles::NDRange &lws=gles::NDRange(1U, 1U, 1U))
Add the kernel to the command queue with the given window.
Definition: IGCKernel.cpp:41
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
cl::NDRange lws_hint() const
Return the Local-Workgroup-Size hint.
Definition: ICLKernel.h:276
#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.
virtual DataType data_type() const =0
Data type used for each element of the tensor.
void run(const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue...
void configure(const ICLTensor *mm_result, const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias, ICLTensor *output, int32_t k, int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage, const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
Initialise the kernel&#39;s input and output.
#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
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
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
const std::string & string_from_gemmlowp_output_stage(GEMMLowpOutputStageType output_stage)
Translates a given GEMMLowp output stage to a string.
Definition: Utils.cpp:260
#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:172
Copyright (c) 2017-2021 Arm Limited.
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
static Status validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
Static function to check if given info will lead to a valid configuration of CLGEMMLowpOffsetContribu...
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
void collapse_from(size_t start)
Collapse dimensions starting from a given point.
Definition: Dimensions.h:183
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.
quantized, asymmetric fixed-point 8-bit number unsigned
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
std::string kernel_name
std::string get_cl_type_from_data_type(const DataType &dt)
Translates a tensor data type to the appropriate OpenCL type.
Definition: CLHelpers.cpp:37
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
GEMMLowp output stage info.
Definition: Types.h:1952
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
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:335
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:941
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:528
CLCompileContext class.
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:443
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:792
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:513
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
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
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:1358
T y() const
Alias to access the size of the second dimension.
Definition: Dimensions.h:92
quantized, asymmetric fixed-point 8-bit number signed
Window first_slice_window_3D() const
First 3D slice of the window.
Definition: Window.h:291
std::tuple< PixelValue, PixelValue > get_min_max(DataType dt)
Compute the mininum and maximum values a data type can take.
Definition: Utils.h:564
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:205
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)