Compute Library
 21.11
ClGemmLowpReductionKernel.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 
32 
33 #include "support/Cast.h"
34 #include "support/StringSupport.h"
35 
36 namespace arm_compute
37 {
38 namespace opencl
39 {
40 namespace kernels
41 {
42 namespace
43 {
44 Status validate_arguments_matrix_a_reduction(const ITensorInfo *src, const ITensorInfo *dst)
45 {
48 
49  if(dst->total_size() > 0)
50  {
52  ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->dimension(0) != src->dimension(1), "Output vector must have length equal to the number of rows of the input matrix");
53  }
54  return Status{};
55 }
56 
57 Status validate_arguments_matrix_b_reduction(const ITensorInfo *src, const ITensorInfo *dst)
58 {
61 
62  if(dst->total_size() > 0)
63  {
65  ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->dimension(0) != src->dimension(0), "Output vector must have length equal to the number of columns of the input matrix");
66  }
67  return Status{};
68 }
69 } // namespace
70 
72 {
74 }
75 
77 {
78  // Perform validate step
79  ARM_COMPUTE_ERROR_ON_NULLPTR(mtx_a, vector_sum_row);
80  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_matrix_a_reduction(mtx_a, vector_sum_row));
81 
82  // Output auto initialization if not yet initialized
83  auto_init_if_empty(*vector_sum_row, TensorShape(mtx_a->dimension(1)), 1, DataType::S32);
84 
85  auto padding_info = get_padding_info({ mtx_a, vector_sum_row });
86 
87  // Set the arguments to pass at compile time
88  CLBuildOptions build_opts;
89  build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(mtx_a->dimension(0)));
90  build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(mtx_a->data_type()));
91  build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(mtx_a->data_type()));
92  build_opts.add_option_if(info.mul_by_scalar, "-DSCALAR=" + support::cpp11::to_string(info.scalar));
93 
94  const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device());
95 
96  std::string kernel_name = "gemmlowp_matrix_a_reduction" + std::string(is_dot8_supported ? "_dot8" : "");
97 
98  // Create kernel
99  _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
100 
101  // Configure kernel window
102  // This kernel does not need padding
103  Window win = calculate_max_window(*vector_sum_row, Steps());
104  ICLKernel::configure_internal(win);
105 
106  _config_id = kernel_name;
107  _config_id += "_";
108  _config_id += support::cpp11::to_string(mtx_a->dimension(0));
109  _config_id += "_";
110  _config_id += support::cpp11::to_string(mtx_a->dimension(1));
111  _config_id += "_";
112  _config_id += support::cpp11::to_string(mtx_a->dimension(2));
113 
115 }
116 
118 {
119  ARM_COMPUTE_UNUSED(info);
120  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_matrix_a_reduction(mtx_a, vector_sum_row));
121 
122  return Status{};
123 }
124 
125 void ClGemmLowpMatrixAReductionKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
126 {
129 
130  const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
131  auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
132 
134  Window slice_in = collapsed.first_slice_window_2D();
135  Window slice_out = collapsed.first_slice_window_2D();
136 
137  // Setup input slice. Its dimensions are increased in the cl kernel.
138  slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
139  slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
140  slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
141 
142  do
143  {
144  unsigned int idx = 0;
145  add_3D_tensor_argument(idx, src, slice_in);
146  add_2D_tensor_argument(idx, dst, slice_out);
147  enqueue(queue, *this, slice_out, lws_hint());
148  }
149  while(collapsed.slide_window_slice_2D(slice_out));
150 }
151 
153 {
154  ARM_COMPUTE_ERROR_ON_NULLPTR(mtx_b, vector_sum_col);
155  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_matrix_b_reduction(mtx_b, vector_sum_col));
156 
157  // Output auto initialization if not yet initialized
158  auto_init_if_empty(*vector_sum_col, TensorShape(mtx_b->dimension(0)), 1, DataType::S32);
159 
160  auto padding_info = get_padding_info({ mtx_b, vector_sum_col });
161 
162  const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16, mtx_b->dimension(0));
163 
164  // Set the arguments to pass at compile time
165  CLBuildOptions build_opts;
166  build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
167  build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(mtx_b->dimension(0) % num_elems_processed_per_iteration));
168  build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(mtx_b->dimension(0)));
169  build_opts.add_option("-DROWS_B=" + support::cpp11::to_string(mtx_b->dimension(1)));
170  build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(mtx_b->data_type()));
171  build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(mtx_b->data_type()));
172  build_opts.add_option_if(info.mul_by_scalar, "-DSCALAR=" + support::cpp11::to_string(info.scalar));
173 
174  // Create kernel
175  _kernel = create_kernel(compile_context, "gemmlowp_matrix_b_reduction", build_opts.options());
176 
177  // Configure kernel window
178  Window win = calculate_max_window(*vector_sum_col, Steps(num_elems_processed_per_iteration));
179  IClKernel::configure_internal(win);
180 
182 }
183 
185 {
186  ARM_COMPUTE_UNUSED(info);
187  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_matrix_b_reduction(mtx_b, vector_sum_col));
188 
189  return Status{};
190 }
191 
192 void ClGemmLowpMatrixBReductionKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
193 {
196 
197  const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
198  auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
199 
201 
202  Window slice_out = collapsed.first_slice_window_2D();
203  Window slice_in = slice_out;
204 
205  slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
206  slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
207 
208  do
209  {
210  unsigned int idx = 0;
211  add_3D_tensor_argument(idx, src, slice_in);
212  add_2D_tensor_argument(idx, dst, slice_out);
213  enqueue(queue, *this, slice_out, lws_hint());
214  }
215  while(collapsed.slide_window_slice_2D(slice_out));
216 }
217 } // namespace kernels
218 } // namespace opencl
219 } // namespace arm_compute
Window first_slice_window_2D() const
First 2D slice of the window.
Definition: Window.h:283
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...
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
bool dot8_supported(const cl::Device &device)
Helper function to check whether the cl_arm_integer_dot_product_int8 extension is supported...
Definition: CLHelpers.cpp:241
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
std::string get_cl_dot8_acc_type_from_data_type(const DataType &dt)
Translates a tensor data type to the appropriate OpenCL dot8 accumulator type.
Definition: CLHelpers.cpp:175
#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.
#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
void configure(const CLCompileContext &compile_context, const ITensorInfo *mtx_b, ITensorInfo *vector_sum_col, const GEMMLowpReductionKernelInfo &info) override
Initialise the kernel&#39;s input and output.
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
void configure(const CLCompileContext &compile_context, const ITensorInfo *mtx_a, ITensorInfo *vector_sum_row, const GEMMLowpReductionKernelInfo &info) override
Initialise the kernel&#39;s input and output.
Status class.
Definition: Error.h:52
int32_t scalar
Scalar value to multiply each reduced column/row by.
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
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2021 Arm Limited.
static Status validate(const ITensorInfo *mtx_b, const ITensorInfo *vector_sum_col, const GEMMLowpReductionKernelInfo &info)
Static function to check if given info will lead to a valid configuration.
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:159
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 *mtx_a, const ITensorInfo *vector_sum_row, const GEMMLowpReductionKernelInfo &info)
Static function to check if given info will lead to a valid configuration.
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
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
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
quantized, asymmetric fixed-point 8-bit number unsigned
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
unsigned int num_elems_processed_per_iteration
std::string get_cl_type_from_data_type(const DataType &dt)
Translates a tensor data type to the appropriate OpenCL type.
Definition: CLHelpers.cpp:39
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...
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
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...
#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
quantized, symmetric fixed-point 8-bit number
CLCompileContext class.
quantized, symmetric per channel fixed-point 8-bit number
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
void add_2D_tensor_argument(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:190
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
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
#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
quantized, asymmetric fixed-point 8-bit number signed
std::string kernel_name
bool mul_by_scalar
True if each column/row reduction has to be multiplied by a scalar value.
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:201