src/core/CL/kernels/CLIm2ColKernel.cpp

   1 /*
   2  * Copyright (c) 2017-2018 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  */
  24 #include "arm_compute/core/CL/kernels/CLIm2ColKernel.h"
  25
  26 #include "arm_compute/core/CL/CLHelpers.h"
  27 #include "arm_compute/core/CL/CLKernelLibrary.h"
  28 #include "arm_compute/core/CL/ICLTensor.h"
  29 #include "arm_compute/core/CL/OpenCL.h"
  30 #include "arm_compute/core/Error.h"
  31 #include "arm_compute/core/Helpers.h"
  32 #include "arm_compute/core/Size2D.h"
  33 #include "arm_compute/core/Types.h"
  34 #include "arm_compute/core/Validate.h"
  35 #include "support/ToolchainSupport.h"
  36
  37 #include <cmath>
  38 #include <tuple>
  39
  40 using namespace arm_compute;
  41
  42 namespace
  43 {
  44 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, bool has_bias, const Size2D &dilation)
  45 {
  46     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
  47     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::QASYMM8 && has_bias);
  48     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
  49     ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
  50
  51     // Checks performed when output is configured
  52     if(output->total_size() != 0)
  53     {
  54         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
  55         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
  56     }
  57
  58     return Status{};
  59 }
  60 } // namespace
  61
  62 CLIm2ColKernel::CLIm2ColKernel()
  63     : _input(nullptr), _output(nullptr), _convolved_dims(), _num_elems_processed_per_iteration(1), _run_func(nullptr), _kernel_dims()
  64 {
  65 }
  66
  67 void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation)
  68 {
  69     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
  70
  71     // Perform validation step
  72     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), has_bias, dilation));
  73
  74     _input       = input;
  75     _output      = output;
  76     _kernel_dims = kernel_dims;
  77
  78     const DataType  data_type  = input->info()->data_type();
  79     const GPUTarget gpu_target = get_target();
  80
  81     // Create kernel
  82     CLBuildOptions build_opts;
  83     build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)));
  84     build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input->info()->element_size()));
  85     build_opts.add_option_if(has_bias, "-DHAS_BIAS");
  86     build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()));
  87
  88     int stride_x = 0;
  89     int stride_y = 0;
  90
  91     std::tie(stride_x, stride_y) = conv_info.stride();
  92
  93     const bool run_img2col_reduced = (output->info()->dimension(0) == (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))) && (TensorShape::num_max_dimensions >= 4)
  94                                      && (std::equal(input->info()->tensor_shape().cbegin() + 3,
  95                                                     input->info()->tensor_shape().cend(),
  96                                                     output->info()->tensor_shape().cbegin() + 1))
  97                                      && ((stride_x == 1) && (stride_y == 1) && !conv_info.has_padding());
  98
  99     bool is_optimized_path = false;
 100
 101     _num_elems_processed_per_iteration = 1;
 102
 103     std::string kernel_name;
 104     if(!run_img2col_reduced)
 105     {
 106         // Default kernel name
 107         kernel_name = "im2col_generic_dchw";
 108
 109         _convolved_dims = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1),
 110                                             kernel_dims.width, kernel_dims.height,
 111                                             conv_info, dilation);
 112
 113         build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
 114         build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
 115         build_opts.add_option("-DKERNEL_DEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
 116         build_opts.add_option("-DCONVOLVED_WIDTH=" + support::cpp11::to_string(_convolved_dims.first));
 117         build_opts.add_option("-DCONVOLVED_HEIGHT=" + support::cpp11::to_string(_convolved_dims.second));
 118         build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
 119         build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
 120         build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
 121         build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
 122         build_opts.add_option("-DPAD_RIGHT=" + support::cpp11::to_string(conv_info.pad_right()));
 123         build_opts.add_option("-DPAD_BOTTOM=" + support::cpp11::to_string(conv_info.pad_bottom()));
 124         build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
 125         build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
 126         build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
 127         build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
 128         build_opts.add_option_if_else(is_data_type_quantized(data_type), "-DPAD_VALUE=" + support::cpp11::to_string(input->info()->quantization_info().offset), "-DPAD_VALUE=0");
 129
 130         const bool squared_im2col = kernel_dims.width == kernel_dims.height;
 131
 132         if(dilation == Size2D(1U, 1U))
 133         {
 134             if(squared_im2col && !is_data_type_fixed_point(data_type))
 135             {
 136                 // Check if we can run an optimized im2col
 137                 switch(kernel_dims.width)
 138                 {
 139                     case 1:
 140                         // Optimized im2col1x1 if stride_x = 1 and conv_info.has_padding() = false
 141                         if(conv_info.stride().first == 1 && !conv_info.has_padding())
 142                         {
 143                             // Set hint for LWS
 144                             _lws_hint                          = cl::NDRange(1, 1, 8);
 145                             _num_elems_processed_per_iteration = 4;
 146                             is_optimized_path                  = true;
 147                             kernel_name                        = "im2col1x1_stridex1_dchw";
 148                         }
 149                         break;
 150                     case 3:
 151                         _lws_hint                          = cl::NDRange(1, 1, 8);
 152                         _num_elems_processed_per_iteration = 1;
 153                         is_optimized_path                  = true;
 154                         kernel_name                        = "im2col3x3_dchw";
 155                         break;
 156                     case 5:
 157                         _num_elems_processed_per_iteration = 1;
 158                         is_optimized_path                  = true;
 159                         kernel_name                        = "im2col5x5_dchw";
 160                         break;
 161                     case 11:
 162                         // Optimized im2col11x11 if pad_x = pad_y = 0
 163                         if(!conv_info.has_padding())
 164                         {
 165                             _num_elems_processed_per_iteration = 1;
 166                             is_optimized_path                  = true;
 167                             kernel_name                        = "im2col11x11_padx0_pady0_dchw";
 168                         }
 169                         break;
 170                     default:
 171                         is_optimized_path = false;
 172                         break;
 173                 }
 174             }
 175             else if(kernel_dims.width > 1 && !conv_info.has_padding())
 176             {
 177                 _num_elems_processed_per_iteration = 1;
 178                 kernel_name                        = "im2col_generic_padx0_pady0_dchw";
 179
 180                 // Optimized im2col is performed using one or more vector operations with the specified vector size
 181                 // and a remainder. For example, for 5x5 convolutions, im2col is performed using vectors of size 4
 182                 // and scalars; for 7x7 convolutions, using vectors of size 4 and vectors of size 3.
 183                 // Using the vector size of 4 is always safe since OpenCL supports vectors of size 2 and 3.
 184                 // Using the vector size of 8, however, may be faster.
 185                 size_t vector_size = 4;
 186                 // For 2x2 convolutions, use vectors of size 2. (For 3x3 convolutions, im2col_kernel3x3_padx0_pady0
 187                 // is used instead.)
 188                 if(kernel_dims.width < vector_size)
 189                 {
 190                     vector_size = kernel_dims.width;
 191                 }
 192                 // Local work size and vector size optimized for the 11x11 AlexNet convolution on Bifrost.
 193                 if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::TNOX) && kernel_dims.width == 11)
 194                 {
 195                     _lws_hint   = cl::NDRange(1, 1, 1);
 196                     vector_size = 8;
 197                 }
 198                 const size_t width_mod_vector_size = kernel_dims.width % vector_size;
 199                 build_opts.add_option("-DVECTOR_SIZE=" + support::cpp11::to_string(vector_size));
 200                 build_opts.add_option("-DWIDTH_MOD_VECTOR_SIZE=" + support::cpp11::to_string(width_mod_vector_size));
 201             }
 202         }
 203         _run_func = &CLIm2ColKernel::run_generic;
 204     }
 205     else
 206     {
 207         _num_elems_processed_per_iteration = 1;
 208         kernel_name                        = "im2col_reduced_dchw";
 209         _run_func                          = &CLIm2ColKernel::run_reduced;
 210     }
 211
 212     // Create kernel
 213     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 214
 215     // Configure kernel window
 216     Window win;
 217     if(is_optimized_path)
 218     {
 219         win = calculate_max_window(*input->info(),
 220                                    Steps(_num_elems_processed_per_iteration),
 221                                    false,
 222                                    BorderSize(conv_info.pad_top(), conv_info.pad_right(), conv_info.pad_bottom(), conv_info.pad_left()));
 223
 224         const int x = -conv_info.pad_left();
 225         const int y = -conv_info.pad_top();
 226         const int w = kernel_dims.width * _num_elems_processed_per_iteration;
 227         const int h = kernel_dims.height;
 228
 229         AccessWindowRectangle input_access(input->info(), x, y, w, h);
 230
 231         update_window_and_padding(win, input_access);
 232     }
 233     else
 234     {
 235         // For the generic case, CLIm2ColKernel doesn't need padding (we do not read out-of-bounds elements) so
 236         // update_window_and_padding() can be skipped
 237         win = calculate_max_window(*input->info(), Steps());
 238     }
 239
 240     output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
 241     if(!run_img2col_reduced)
 242     {
 243         // set the Z dimension's step same size as the whole dimension so that one can't split across the Z dimension
 244         win.set_dimension_step(Window::DimZ, win[Window::DimZ].end() - win[Window::DimZ].start());
 245     }
 246
 247     ICLKernel::configure(win);
 248
 249     // Set config_id for enabling LWS tuning
 250     _config_id = kernel_name;
 251     _config_id += "_";
 252     _config_id += lower_string(string_from_data_type(input->info()->data_type()));
 253     _config_id += "_";
 254     _config_id += support::cpp11::to_string(output->info()->dimension(0));
 255     _config_id += "_";
 256     _config_id += support::cpp11::to_string(output->info()->dimension(1));
 257 }
 258
 259 Status CLIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation)
 260 {
 261     ARM_COMPUTE_UNUSED(kernel_dims);
 262     ARM_COMPUTE_UNUSED(conv_info);
 263     ARM_COMPUTE_UNUSED(has_bias);
 264     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, has_bias, dilation));
 265     return Status{};
 266 }
 267
 268 void CLIm2ColKernel::run(const Window &window, cl::CommandQueue &queue)
 269 {
 270     ARM_COMPUTE_ERROR_ON(_run_func == nullptr);
 271     (this->*_run_func)(window, queue);
 272 }
 273
 274 void CLIm2ColKernel::run_generic(const Window &window, cl::CommandQueue &queue)
 275 {
 276     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
 277     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
 278
 279     // Get initial windows
 280     Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
 281     // Change the Z dimension's step back to 1
 282     window_collapsed.set_dimension_step(Window::DimZ, 1);
 283
 284     Window slice     = window_collapsed.first_slice_window_3D();
 285     Window slice_in  = window_collapsed.first_slice_window_3D();
 286     Window slice_out = window_collapsed.first_slice_window_3D();
 287
 288     // Setup slice if stride_x != 0 or stride_y != 0
 289     if(_convolved_dims.first != _input->info()->dimension(0) || _convolved_dims.second != _input->info()->dimension(1))
 290     {
 291         // If the stride_x or stride_y are not 1, the output tensor of matrix multiply (Convolved tensor) will not
 292         // have the same shape of the im2col input tensor
 293         // In this case we need to re-compute the window using the shape of the tensor after matrix multiply (convolved_dims)
 294         slice.set(Window::DimX, Window::Dimension(0, static_cast<int>(_convolved_dims.first), 1));
 295         slice.set(Window::DimY, Window::Dimension(0, static_cast<int>(_convolved_dims.second), 1));
 296     }
 297
 298     // Setup input slice
 299     // The first three dimensions of the input are increased by the inner loops
 300     slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
 301     slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
 302     slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
 303
 304     // Setup output slice
 305     slice_out.set(Window::DimX, Window::Dimension(0, _output->info()->dimension(0), _kernel_dims.area()));
 306     slice_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 1));
 307     slice_out.set(Window::DimZ, Window::Dimension(0, 1, 1));
 308
 309     do
 310     {
 311         unsigned int idx = 0;
 312         add_3D_tensor_argument(idx, _input, slice_in);
 313         add_2D_tensor_argument(idx, _output, slice_out);
 314         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input->info()->strides_in_bytes()[3]));
 315         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[3]));
 316         enqueue(queue, *this, slice, _lws_hint);
 317     }
 318     while(window_collapsed.slide_window_slice_3D(slice) && window_collapsed.slide_window_slice_3D(slice_out) && window_collapsed.slide_window_slice_3D(slice_in));
 319 }
 320
 321 void CLIm2ColKernel::run_reduced(const Window &window, cl::CommandQueue &queue)
 322 {
 323     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
 324     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
 325
 326     Window out_window;
 327     out_window.use_tensor_dimensions(_output->info()->tensor_shape());
 328
 329     Window out_slice = out_window.first_slice_window_1D();
 330     Window in_slice  = window.first_slice_window_3D();
 331
 332     // Run kernel
 333     do
 334     {
 335         // Set arguments
 336         unsigned int idx = 0;
 337         add_3D_tensor_argument(idx, _input, in_slice);
 338         add_1D_tensor_argument(idx, _output, out_slice);
 339
 340         _kernel.setArg<cl_uint>(idx++, _input->info()->dimension(0));
 341         _kernel.setArg<cl_uint>(idx++, _input->info()->dimension(1));
 342         enqueue(queue, *this, in_slice, _lws_hint);
 343     }
 344     while(window.slide_window_slice_3D(in_slice) && out_window.slide_window_slice_1D(out_slice));
 345 }