Compute Library
 21.02
GCSoftmaxLayerKernel.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 
32 #include "arm_compute/core/Utils.h"
38 #include "support/StringSupport.h"
39 
40 #include <set>
41 #include <string>
42 
43 using namespace arm_compute;
44 
46 {
49 
50  // Softmax across the x dimension
52  output_shape.set(0, 1);
53 
54  // Output auto initialization if not yet initialized
55  auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type());
56 
59 
60  _input = input;
61  _output = output;
62 
63  // Set build options
64  std::set<std::string> build_opts;
65  std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
66  build_opts.insert("#define " + dt_name);
67  build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1));
68  build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1));
69  build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
70  build_opts.insert("#define SOFTMAX_LAYER_MAX");
71 
72  // Tell the kernel that the width is not a multiple of 8
73  if((input->info()->dimension(0) % 8) != 0)
74  {
75  build_opts.insert("#define NON_MULTIPLE_OF_8");
76  }
77 
78  // Create kernel
79  _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("softmax_layer_max", build_opts));
80 
81  // Set fixed arguments
82  unsigned int idx = 2 * num_arguments_per_3D_tensor(); //Skip the input and output parameters
83  _kernel.set_argument(idx++, input->info()->dimension(0));
84 
85  // Configure kernel window
86  // The kernel loops over all elements in steps of 8
87  const unsigned int num_elems_processed_per_iteration = ceil_to_multiple(input->info()->dimension(0), 8);
88  unsigned int num_elems_written_per_iteration = 1;
89  if(input->info()->data_type() == DataType::F16)
90  {
91  num_elems_written_per_iteration = 2;
92  }
93 
94  Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
96  AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration);
97 
98  update_window_and_padding(win, input_access, output_access);
99 
100  output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
101 
102  IGCKernel::configure(win);
103 }
104 
106  : _input(nullptr), _max(nullptr), _output(nullptr), _sum(nullptr)
107 {
108 }
109 
111 {
113  ARM_COMPUTE_ERROR_ON_NULLPTR(max, sum, output);
114 
115  // Output auto initialization if not yet initialized
116  auto_init_if_empty(*sum->info(), max->info()->tensor_shape(), 1, input->info()->data_type());
117  auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type());
118 
119  ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, max, sum);
122 
123  _input = input;
124  _max = max;
125  _output = output;
126  _sum = sum;
127 
128  // Set build options
129  std::set<std::string> build_opts;
130  std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
131  build_opts.insert("#define " + dt_name);
132  build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1));
133  build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1));
134  build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
135  build_opts.insert("#define SOFTMAX_LAYER_SHIFT_EXP_SUM");
136 
137  // Tell the kernel that the width is not a multiple of 8
138  if((input->info()->dimension(0) % 8) != 0)
139  {
140  build_opts.insert("#define NON_MULTIPLE_OF_8");
141  }
142 
143  // Create kernel
144  _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("softmax_layer_shift_exp_sum", build_opts));
145 
146  // Set fixed arguments
147  unsigned int idx = 4 * num_arguments_per_3D_tensor(); //Skip the input and output parameters
148  _kernel.set_argument(idx++, input->info()->dimension(0));
149 
150  // Configure window
151  // The kernel loops over all elements in steps of 8
152  const unsigned int num_elems_processed_per_iteration = ceil_to_multiple(input->info()->dimension(0), 8);
153  unsigned int num_elems_written_per_iteration = 1;
154  if(input->info()->data_type() == DataType::F16)
155  {
156  num_elems_written_per_iteration = 2;
157  }
158 
159  Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
160 
162  AccessWindowHorizontal max_access(max->info(), 0, num_elems_written_per_iteration);
164  AccessWindowHorizontal sum_access(sum->info(), 0, num_elems_written_per_iteration);
165 
166  update_window_and_padding(win, input_access, max_access, output_access, sum_access);
167 
168  output_access.set_valid_region(win, input->info()->valid_region());
169  sum_access.set_valid_region(win, ValidRegion(Coordinates(), sum->info()->tensor_shape()));
170 
171  IGCKernel::configure(win);
172 }
173 
175 {
178 
179  Window window_collapsed = window.collapse_if_possible(IGCKernel::window(), Window::DimZ);
180  Window slice = window_collapsed.first_slice_window_3D();
181 
182  _kernel.use();
183 
184  do
185  {
186  unsigned int idx = 0;
187  unsigned int binding = 1; // SSBO binding starts from 1.
188  // Set inputs
189  add_3D_tensor_argument(idx, _input, binding++, slice);
190  add_3D_tensor_argument(idx, _max, binding++, slice);
191  add_3D_tensor_argument(idx, _output, binding++, slice);
192  add_3D_tensor_argument(idx, _sum, binding++, slice);
193  _kernel.update_shader_params();
194  enqueue(*this, slice);
195  }
196  while(window_collapsed.slide_window_slice_3D(slice));
197 }
198 
200  : _input(nullptr), _sum(nullptr), _output(nullptr)
201 {
202 }
203 
205 {
207  ARM_COMPUTE_ERROR_ON_NULLPTR(sum, output);
208 
209  // Output auto initialization if not yet initialized
210  auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type());
211 
214 
215  _input = input;
216  _sum = sum;
217  _output = output;
218 
219  // Set build options
220  std::set<std::string> build_opts;
221  std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
222  build_opts.insert("#define " + dt_name);
223  build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1));
224  build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1));
225  build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
226  build_opts.insert("#define SOFTMAX_LAYER_NORM");
227 
228  // Create kernel
229  _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("softmax_layer_norm", build_opts));
230 
231  // Configure window
232  constexpr unsigned int num_elems_processed_per_iteration = 8;
233  unsigned int num_elems_written_per_iteration = 1;
234  if(input->info()->data_type() == DataType::F16)
235  {
236  num_elems_written_per_iteration = 2;
237  }
238 
239  Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
240 
242  AccessWindowStatic sum_access(sum->info(), 0, 0, num_elems_written_per_iteration, sum->info()->dimension(1));
244 
245  update_window_and_padding(win, input_access, sum_access, output_access);
246 
247  output_access.set_valid_region(win, input->info()->valid_region());
248 
249  IGCKernel::configure(win);
250 }
251 
253 {
256 
257  Window window_collapsed = window.collapse_if_possible(IGCKernel::window(), Window::DimZ);
258  Window slice = window_collapsed.first_slice_window_3D();
259 
260  _kernel.use();
261 
262  do
263  {
264  Window sum_slice = slice;
265  sum_slice.set(Window::DimX, Window::Dimension(0, 1, 1));
266 
267  unsigned int idx = 0;
268  unsigned int binding = 1; // SSBO binding starts from 1.
269  // Set inputs
270  add_3D_tensor_argument(idx, _input, binding++, slice);
271  add_3D_tensor_argument(idx, _sum, binding++, slice);
272  add_3D_tensor_argument(idx, _output, binding++, slice);
273 
274  _kernel.update_shader_params();
275  enqueue(*this, slice);
276  }
277  while(window_collapsed.slide_window_slice_3D(slice));
278 }
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
Shape of a tensor.
Definition: TensorShape.h:39
void add_3D_tensor_argument(unsigned int &idx, const IGCTensor *tensor, const unsigned int binding_point, 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: IGCKernel.cpp:132
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.
DATA_TYPE sum(__global const DATA_TYPE *input)
Calculate sum of a vector.
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.
1 channel, 1 F32 per channel
Interface for GLES Compute tensor.
Definition: IGCTensor.h:35
Describe one of the image&#39;s dimensions with a start, end and step.
Definition: Window.h:77
GCKernel class.
unsigned int num_arguments_per_3D_tensor() const
Returns the number of arguments enqueued per 3D tensor object.
Definition: IGCKernel.cpp:147
Copyright (c) 2017-2021 Arm Limited.
1 channel, 1 F16 per channel
virtual ValidRegion valid_region() const =0
Valid region of the tensor.
Implementation of a static rectangular access pattern.
void run(const Window &window) override
Enqueue the OpenGL ES shader to process the given window.
void run(const Window &window) override
Enqueue the OpenGL ES shader to process the given window.
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: WindowHelpers.h:46
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.
#define ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:543
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:71
Manages all the GLES kernels compilation and caching, provides accessors for the GLES Context...
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.
#define ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:441
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 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
#define ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:790
Wrapper to configure the Khronos EGL and OpenGL ES C header.
static GCKernelLibrary & get()
Get the static instance of GCKernelLibrary.
#define ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(...)
Definition: Validate.h:286
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Definition: Window.h:47
GCKernel create_kernel(const std::string &shader_name, const StringSet &build_options_set={}) const
Creates a kernel from the kernel library.
unsigned int num_elems_processed_per_iteration
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Container for valid region of a window.
Definition: Types.h:188
void configure(const IGCTensor *input, IGCTensor *output)
Set the input and output tensors.
void configure(const IGCTensor *input, const IGCTensor *max, IGCTensor *output, IGCTensor *sum)
Set the input and output tensors.
Window first_slice_window_3D() const
First 3D slice of the window.
Definition: Window.h:291
void configure(const IGCTensor *input, const IGCTensor *sum, IGCTensor *output)
Set the input and output tensors.
Describe a multidimensional execution window.
Definition: Window.h:39
TensorShape & set(size_t dimension, size_t value, bool apply_dim_correction=true, bool increase_dim_unit=true)
Accessor to set the value of one of the dimensions.
Definition: TensorShape.h:79
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:205
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)