Compute Library
 20.08
CLQLSTMLayerNormalizationKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 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  */
26 #include "arm_compute/core/Error.h"
28 #include "support/StringSupport.h"
29 
30 namespace arm_compute
31 {
32 namespace
33 {
34 QuantizationInfo compute_output_qinfo()
35 {
36  return QuantizationInfo(1.f / 4096);
37 }
38 
39 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
40 {
42  // Output auto inizialitation if not yet initialized
43  auto_init_if_empty(*output, *input);
44  output->set_quantization_info(compute_output_qinfo());
45 
46  const uint32_t temp_num_elems_processed_per_iteration = max_cl_vector_width / input->element_size();
47  /* If width is less then step, then make step same as width to avoid global size being step instead of actual width. */
48  /* Or we should fix in arm_compute::enqueue() or arm_compute::calculate_max_window(). */
49  const uint32_t num_elems_processed_per_iteration = (input->dimension(0) < temp_num_elems_processed_per_iteration) ? input->dimension(0) : temp_num_elems_processed_per_iteration;
50 
51  // This kernel doesn't need padding
53  output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
54 
55  return std::make_pair(Status{}, win);
56 }
57 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias)
58 {
60 
61  ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 2, "Input tensor cannot have more than 2 dimensions");
62  ARM_COMPUTE_RETURN_ERROR_ON_MSG(weight->num_dimensions() > 1, "Weight tensor cannot have more than 1 dimensions");
63  ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->num_dimensions() > 1, "Bias tensor cannot have more than 1 dimensions");
64 
68 
69  ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().x() != weight->tensor_shape().x());
71 
72  // Checks performed when output is configured
73  if(output->total_size() != 0)
74  {
77  }
78  return Status{};
79 }
80 } // namespace
81 
83  : _input(nullptr), _weight(nullptr), _bias(nullptr), _output(nullptr)
84 {
85 }
86 
87 void CLQLSTMLayerNormalizationKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const ICLTensor *weight, const ICLTensor *bias)
88 {
89  ARM_COMPUTE_ERROR_ON_NULLPTR(input, weight, bias, output);
90 
91  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), weight->info(), bias->info()));
92 
93  _input = input;
94  _weight = weight;
95  _bias = bias;
96  _output = output;
97 
98  const uint32_t num_elems_processed_per_iteration = max_cl_vector_width / input->info()->element_size();
99 
100  int32_t output_multiplier{};
101  int32_t output_shift{};
102  const UniformQuantizationInfo quan_info = _weight->info()->quantization_info().uniform();
103  const Status status = quantization::calculate_quantized_multiplier(quan_info.scale, &output_multiplier, &output_shift);
104  output_shift *= -1;
105 
106  // Set build options
107  CLBuildOptions build_opts;
108  build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
110  build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
111  build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
112  build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
113  build_opts.add_option("-DMIN_BOUND=" + support::cpp11::to_string(std::get<0>(quantization::get_min_max_values_from_quantized_data_type(input->info()->data_type()))));
114  build_opts.add_option("-DMAX_BOUND=" + support::cpp11::to_string(std::get<1>(quantization::get_min_max_values_from_quantized_data_type(input->info()->data_type()))));
115 
116  // Create kernel
117  _kernel = create_kernel(compile_context, "qlstm_layer_normalization", build_opts.options());
118 
119  // Configure kernel window
120  auto win_config = validate_and_configure_window(input->info(), output->info());
121  ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
122  ICLKernel::configure_internal(win_config.second);
123 
124  // Set config_id for enabling LWS tuning
125  _config_id = "qlstm_layer_normalization_";
126  _config_id += lower_string(string_from_data_type(input->info()->data_type()));
127  _config_id += "_";
128  _config_id += support::cpp11::to_string(input->info()->dimension(0));
129  _config_id += "_";
130  _config_id += support::cpp11::to_string(input->info()->dimension(1));
131 }
132 
134 {
135  configure(CLKernelLibrary::get().get_compile_context(), input, output, weight, bias);
136 }
137 
139 {
141  ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
142  return Status{};
143 }
144 
145 void CLQLSTMLayerNormalizationKernel::run(const Window &window, cl::CommandQueue &queue)
146 {
149 
151  // Set slice step equal to width to force gws[0] to 1, as each thread normalizes across all rows
152  slice.set_dimension_step(Window::DimX, _input->info()->dimension(0));
153 
154  Window weight_window;
155  Window weight_slice;
156 
157  weight_window.use_tensor_dimensions(_weight->info()->tensor_shape());
158  weight_slice = weight_window.first_slice_window_1D();
159 
160  do
161  {
162  unsigned int idx = 0;
163  add_2D_tensor_argument(idx, _input, slice);
164  add_1D_tensor_argument(idx, _weight, weight_slice);
165  add_1D_tensor_argument(idx, _bias, weight_slice);
166  add_2D_tensor_argument(idx, _output, slice);
167 
168  enqueue(queue, *this, slice, lws_hint());
169  }
171 }
172 } // namespace arm_compute
Window first_slice_window_2D() const
First 2D slice of the window.
Definition: Window.h:281
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
quantized, symmetric fixed-point 16-bit number
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:39
const StringSet & options() const
Gets the current options list set.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
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
Quantization info when assuming per layer quantization.
Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift, bool ignore_epsilon=false)
Calculate quantized representation of multiplier.
Status class.
Definition: Error.h:52
std::string lower_string(const std::string &val)
Lower a given string.
Definition: Utils.cpp:326
#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 use_tensor_dimensions(const TensorShape &shape, size_t first_dimension=Window::DimX)
Use the tensor's dimensions to fill the window dimensions.
Definition: Window.inl:276
void configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *weight, const ICLTensor *bias)
Initialise the kernel's input and outputs.
bool slide_window_slice_2D(Window &slice) const
Slide the passed 2D window slice.
Definition: Window.h:321
Copyright (c) 2017-2020 Arm Limited.
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...
Definition: Helpers.inl:207
ITensorInfo * info() const override
Interface to be implemented by the child class to return the tensor's metadata.
Definition: Tensor.cpp:33
1 channel, 1 S32 per channel
void add_option(std::string option)
Adds option to the existing build option list.
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias)
Static function to check if given info will lead to a valid configuration of CLQLSTMLayerNormalizatio...
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
const std::string & string_from_data_type(DataType dt)
Convert a data type identity into a string.
Definition: Utils.cpp:135
std::pair< int, int > get_min_max_values_from_quantized_data_type(DataType data_type)
Get minimum and maximum values for the input quantized data type.
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:443
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
UniformQuantizationInfo uniform() const
Return per layer quantization info.
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
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.
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
CLCompileContext class.
void run(const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue.
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:163
void add_2D_tensor_argument(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:135
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
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 add_1D_tensor_argument(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:111
#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
Window first_slice_window_1D() const
First 1D slice of the window.
Definition: Window.h:273
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)