Compute Library
 22.05
ClPool3dKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2022 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 #include "src/core/CL/CLValidate.h"
32 #include "support/Cast.h"
33 #include "utils/TypePrinter.h"
34 
35 namespace arm_compute
36 {
37 namespace opencl
38 {
39 namespace kernels
40 {
42 
43 namespace
44 {
45 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const Pooling3dLayerInfo &pool_info)
46 {
48  ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->data_layout() != DataLayout::NDHWC, "Only NDHWC layout supported");
49 
51  ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_info.stride.x() == 0 || pool_info.stride.y() == 0 || pool_info.stride.z() == 0), "Strides cannot be zero.");
53  ARM_COMPUTE_RETURN_ERROR_ON_MSG((!is_data_type_float(src->data_type())) && (!pool_info.exclude_padding
54  && (pool_info.pool_type == PoolingType::AVG)),
55  "Exclude padding is unsupported for non-float types for Avg op");
56 
57  const auto data_layout = src->data_layout();
60  const int idx_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::DEPTH);
61  const bool is_global_pooling = pool_info.is_global_pooling;
62  const unsigned int pool_size_x = is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width;
63  const unsigned int pool_size_y = is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height;
64  const unsigned int pool_size_z = is_global_pooling ? src->dimension(idx_depth) : pool_info.pool_size.depth;
65  int output_width = 0;
66  int output_height = 0;
67  int output_depth = 0;
68 
69  bool round_type_ceil_with_asymm_padding = (pool_info.round_type == DimensionRoundingType::CEIL) && (!is_symmetric(pool_info.padding));
70  ARM_COMPUTE_RETURN_ERROR_ON_MSG(round_type_ceil_with_asymm_padding, "Cannot use dimension round type CEIL when padding is asymmetric.");
71 
72  ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_pool_3d_region_entirely_outside_input(pool_info), "Pooling region that is entirely outside input tensor is unsupported");
73  std::tie(output_width, output_height, output_depth) = scaled_3d_dimensions_signed(src->tensor_shape()[idx_width], src->tensor_shape()[idx_height],
74  src->tensor_shape()[idx_depth], pool_size_x, pool_size_y,
75  pool_size_z, pool_info);
76 
77  ARM_COMPUTE_RETURN_ERROR_ON_MSG((output_width < 1 || output_height < 1 || output_depth < 1), "Calculated output dimension size is invalid");
78  // Checks performed when dst is configured
79  if(dst->total_size() != 0)
80  {
83  TensorInfo out_info(TensorInfo(compute_pool3d_shape(src->tensor_shape(), pool_info), 1, dst->data_type()));
85  }
86 
87  return Status{};
88 }
89 } // namespace
90 
92 {
93  _type = CLKernelType::POOL;
94 }
95 
96 void ClPool3dKernel::configure(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const Pooling3dLayerInfo &pool_info)
97 {
99  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, pool_info));
100  auto padding_info = get_padding_info({ src, dst });
101 
102  // Auto init if empty
103  TensorShape out_shape = compute_pool3d_shape(src->tensor_shape(), pool_info);
104  auto_init_if_empty(*dst, src->clone()->set_tensor_shape(out_shape));
105 
106  // Set instance variables
107  _pool_info = pool_info;
108  _data_layout = src->data_layout();
109 
110  _num_elems_processed_per_iteration = (dst->data_type() == DataType::F32) ? 2 : 4;
111  _num_elems_processed_per_iteration = adjust_vec_size(_num_elems_processed_per_iteration, dst->dimension(0));
112 
113  const PoolingType pool_type = pool_info.pool_type;
114  const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
115  const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
116  const int idx_depth = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::DEPTH);
117  const int idx_channel = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
118  const int idx_batch_size = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::BATCHES);
119  const int pool_size_x = pool_info.is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width;
120  const int pool_size_y = pool_info.is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height;
121  const int pool_size_z = pool_info.is_global_pooling ? src->dimension(idx_depth) : pool_info.pool_size.depth;
122  const bool exclude_padding = pool_info.exclude_padding;
123  const int pool_stride_x = pool_info.stride.x();
124  const int pool_stride_y = pool_info.stride.y();
125  const int pool_stride_z = pool_info.stride.z();
126  const int pool_pad_top = pool_info.padding.top;
127  const int pool_pad_left = pool_info.padding.left;
128  const int pool_pad_front = pool_info.padding.front;
129  const DataType data_type = src->data_type();
130 
131  // Set build options
132  CLBuildOptions build_opts;
133  build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(_num_elems_processed_per_iteration));
134  build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
135  build_opts.add_option("-DPOOL_" + string_from_pooling_type(pool_type));
136  build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x));
137  build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(pool_stride_y));
138  build_opts.add_option("-DSTRIDE_Z=" + support::cpp11::to_string(pool_stride_z));
139  build_opts.add_option("-DPAD_X=" + support::cpp11::to_string(pool_pad_left));
140  build_opts.add_option("-DPAD_Y=" + support::cpp11::to_string(pool_pad_top));
141  build_opts.add_option("-DPAD_Z=" + support::cpp11::to_string(pool_pad_front));
142  build_opts.add_option("-DPOOL_SIZE_X=" + support::cpp11::to_string(pool_size_x));
143  build_opts.add_option("-DPOOL_SIZE_Y=" + support::cpp11::to_string(pool_size_y));
144  build_opts.add_option("-DPOOL_SIZE_Z=" + support::cpp11::to_string(pool_size_z));
145  build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width)));
146  build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height)));
147  build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(idx_depth)));
148 
149  // If datatype is quantized add relevant parameters
151  {
152  const UniformQuantizationInfo iq_info = src->quantization_info().uniform();
153  const UniformQuantizationInfo oq_info = dst->quantization_info().uniform();
154 
155  build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
156  build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
157  build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
158  build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
159  }
160 
161  // Set the initial value for the pooling operation accordingly with the data type
162  if(pool_type == PoolingType::MAX)
163  {
164  if(is_data_type_quantized(data_type))
165  {
166  PixelValue type_min{};
167  std::tie(type_min, std::ignore) = get_min_max(data_type);
168  build_opts.add_option("-DINITIAL_VALUE=" + support::cpp11::to_string(type_min.get<int32_t>()));
169  }
170  else
171  {
173  }
174  }
175  else
176  {
177  // Pool AVG and Pool L2 initial value
178  build_opts.add_option("-DINITIAL_VALUE=0");
179  }
180  // Create kernel
181  // Floating point mixed precision is support on F16 only
182  const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision && pool_type != PoolingType::MAX;
183 
184  // Wider accumulation is required to avoid accuracy loss
185  // Case 1: Floating point mixed precision (fp16 src data and fp32 accumulation)
186  DataType acc_data_type = data_type;
187  if(use_fp_mixed_precision)
188  {
189  acc_data_type = DataType::F32;
190  }
191  else if(is_data_type_quantized(data_type) && pool_type != PoolingType::MAX) // Use S32 for avg pooling to allow for integer division
192  {
193  acc_data_type = DataType::S32;
194  }
195 
196  build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(acc_data_type));
197  build_opts.add_option_if(use_fp_mixed_precision, "-DFP_MIXED_PRECISION");
198  build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING");
199  build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(idx_height)));
200  build_opts.add_option("-DDST_DEPTH=" + support::cpp11::to_string(dst->dimension(idx_depth)));
201  build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst->dimension(idx_channel)));
202  build_opts.add_option("-DDST_BATCH_SIZE=" + support::cpp11::to_string(dst->dimension(idx_batch_size)));
203  build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % _num_elems_processed_per_iteration));
204 
205  // if datatype is quantized use quantized kernel function
206  std::string kernel_name = (is_data_type_quantized_asymmetric(data_type) ? "pooling_3d_layer_MxN_ndhwc_quantized" : "pooling_3d_layer_MxN_ndhwc");
207  _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
208 
209  // Configure kernel window
210  Window win = calculate_max_window(*dst, Steps(_num_elems_processed_per_iteration));
211  ICLKernel::configure_internal(win);
212 
213  // Set config_id for enabling LWS tuning
214  _config_id = "pooling_layer_3d";
215  _config_id += lower_string(string_from_data_type(data_type));
216  _config_id += "_";
217  _config_id += lower_string(string_from_data_layout(_data_layout));
218  _config_id += "_";
219  _config_id += support::cpp11::to_string(dst->dimension(idx_width));
220  _config_id += "_";
221  _config_id += support::cpp11::to_string(dst->dimension(idx_height));
222  _config_id += "_";
223  _config_id += support::cpp11::to_string(dst->dimension(idx_channel));
224  _config_id += "_";
225  _config_id += lower_string(string_from_data_layout(src->data_layout()));
226 
228 }
229 
231 {
232  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, pool_info));
233  return Status{};
234 }
235 
236 void ClPool3dKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
237 {
240 
241  const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
242  auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST_0));
243 
244  // Collapse 3D window
245  Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
246 
247  // Set CL kernel arguments
248  unsigned int idx = 0;
249  // Passing of the window not needed, as the steps are not used for the pool3d kernel
250  add_5D_tensor_argument(idx, src, window);
251  add_5D_tensor_argument(idx, dst, window);
252  enqueue(queue, *this, window_collapsed, lws_hint());
253 }
254 } // namespace kernels
255 } // namespace opencl
256 } // namespace arm_compute
bool is_data_type_quantized(DataType dt)
Check if a given data type is of quantized type.
Definition: Utils.h:1030
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)
bool is_pool_3d_region_entirely_outside_input(const Pooling3dLayerInfo &info)
Check if the 3d pool region is entirely outside the input tensor.
Definition: Utils.cpp:247
#define ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(tensor)
Definition: CLValidate.h:35
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
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(...)
Definition: Validate.h:490
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
const StringSet & options() const
Gets the current options list set.
#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.
static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const Pooling3dLayerInfo &pool_info)
Static function to check if given info will lead to a valid configuration.
1 channel, 1 F32 per channel
#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
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Quantization info when assuming per layer quantization.
TensorShape compute_pool3d_shape(const TensorShape &src, Pooling3dLayerInfo pool3d_info)
Calculate the output pool3d shape of a tensor.
Status class.
Definition: Error.h:52
std::string lower_string(const std::string &val)
Lower a given string.
Definition: Utils.cpp:351
Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const PadStrideInfo &conv_info)
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...
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2022 Arm Limited.
1 channel, 1 F16 per channel
bool is_symmetric(const Padding3D &info)
Check if the 3D padding is symmetric i.e.
Definition: Utils.h:926
#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.
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
const std::string & string_from_data_type(DataType dt)
Convert a data type identity into a string.
Definition: Utils.cpp:135
size_t front
Padding across the depth dimenstion on the front, in elements.
Definition: Types.h:806
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
std::string float_to_string_with_full_precision(float val)
Create a string with the float in full precision.
Definition: Utils.h:1124
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
void configure(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const Pooling3dLayerInfo &pool_info)
Configure kernel for a given list of arguments.
quantized, asymmetric fixed-point 8-bit number unsigned
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
size_t height
Height of the 3D shape or object.
Definition: Size3D.h:93
Pooling Layer Information struct.
Definition: Types.h:1281
Pool CL kernel type.
Definition: CLTypes.h:86
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: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...
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
size_t top
Padding across the height dimenstion on the top, in elements.
Definition: Types.h:804
void add_option_if(bool cond, std::string option)
Adds option if a given condition is true;.
size_t left
Padding across the width dimenstion on the left, in elements.
Definition: Types.h:802
size_t width
Width of the 3D shape or object.
Definition: Size3D.h:92
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
Num samples, depth, height, width, channels.
#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:601
CLCompileContext class.
bool is_data_type_quantized_asymmetric(DataType dt)
Check if a given data type is of asymmetric quantized type.
Definition: Utils.h:1052
PoolingType
Available pooling types.
Definition: Types.h:557
ITensor * get_tensor(int id)
Get tensor of a given id from the pac.
Definition: ITensorPack.cpp:64
size_t depth
Depth of the 3D shape or object.
Definition: Size3D.h:94
const std::string & string_from_data_layout(DataLayout dl)
Convert a data layout identity into a string.
Definition: Utils.cpp:123
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:439
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Definition: Window.h:47
size_t get_data_layout_dimension_index(const DataLayout &data_layout, const DataLayoutDimension &data_layout_dimension)
Get the index of the given dimension.
Definition: Helpers.inl:193
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:541
#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:586
std::tuple< int, int, int > scaled_3d_dimensions_signed(int width, int height, int depth, int kernel_width, int kernel_height, int kernel_depth, const Pooling3dLayerInfo &pool3d_info)
Returns calculated width, height and depth of output scaled tensor depending on dimensions rounding m...
Definition: Utils.cpp:490
#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
size_t z() const
Semantic accessor for depth as z.
Definition: Size3D.h:76
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:1222
quantized, asymmetric fixed-point 8-bit number signed
std::string kernel_name
DataType
Available data types.
Definition: Types.h:79
size_t x() const
Semantic accessor for width as x.
Definition: Size3D.h:58
const std::string & string_from_pooling_type(PoolingType type)
Translates a given pooling type to a string.
Definition: Utils.cpp:223
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
bool is_data_type_float(DataType dt)
Check if a given data type is of floating point type.
Definition: Utils.h:1010
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:201
virtual DataLayout data_layout() const =0
Get the data layout of the tensor.
size_t y() const
Semantic accessor for height as y.
Definition: Size3D.h:67