COMPMID-1235: Implements StridedSlice for CL
Change-Id: If2b44da31fae528c76be742b4b3a21fb0eb06b49
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/145284
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 3c92257..6f45756 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -359,6 +359,7 @@
{ "softmax_layer_max_shift_exp_sum_quantized_parallel", "softmax_layer_quantized.cl" },
{ "softmax_layer_max_shift_exp_sum_serial", "softmax_layer.cl" },
{ "softmax_layer_max_shift_exp_sum_parallel", "softmax_layer.cl" },
+ { "strided_slice", "strided_slice.cl" },
{ "suppress_non_maximum", "canny.cl" },
{ "tablelookup_U8", "tablelookup.cl" },
{ "tablelookup_S16", "tablelookup.cl" },
@@ -735,6 +736,10 @@
#include "./cl_kernels/softmax_layer_quantized.clembed"
},
{
+ "strided_slice.cl",
+#include "./cl_kernels/strided_slice.clembed"
+ },
+ {
"tablelookup.cl",
#include "./cl_kernels/tablelookup.clembed"
},
diff --git a/src/core/CL/cl_kernels/strided_slice.cl b/src/core/CL/cl_kernels/strided_slice.cl
new file mode 100644
index 0000000..7c68fb9
--- /dev/null
+++ b/src/core/CL/cl_kernels/strided_slice.cl
@@ -0,0 +1,85 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "helpers.h"
+
+/** Perform a strided slice operation on a given input.
+ *
+ * @attention Supported tensor rank: up to 4
+ *
+ * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
+ * @attention Input and output tensor dephts should be given as a preprocessor arguments using -DSRC_DEPTH=size. and -DDST_DEPTH=size
+ * @attention Absolute start coordinates for each dimension should be given as preprocessor -DSTART_index=value e.g. -DSTART_0=2
+ * @attention Strides for each dimension should be given as preprocessor -DSTRIDE_index=value e.g. -DSTRIDE_1=1
+ *
+ * @param[in] input_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
+ * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes)
+ * @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] output_stride_w Stride of the destination tensor in W dimension (in bytes)
+ * @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void strided_slice(
+ TENSOR4D_DECLARATION(input),
+ TENSOR4D_DECLARATION(output))
+{
+ // Get pixels pointer
+ Tensor4D input = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, SRC_DEPTH);
+ Tensor4D output = CONVERT_TO_TENSOR4D_STRUCT(output, DST_DEPTH);
+
+ int offset_0 = 0;
+ int offset_1 = 0;
+ int offset_2 = 0;
+ int offset_3 = 0;
+
+ // Calculate offset
+#if defined(START_0) && defined(STRIDE_0)
+ offset_0 = (int)START_0 + (int)get_global_id(0) * (int)STRIDE_0;
+#endif // defined(START_0) && defined(STRIDE_0)
+#if defined(START_1) && defined(STRIDE_1)
+ offset_1 = (int)START_1 + (int)get_global_id(1) * (int)STRIDE_1;
+#endif // defined(START_1) && defined(STRIDE_1)
+#if defined(START_2) && defined(STRIDE_2)
+ offset_2 = (int)START_2 + ((int)get_global_id(2) % (int)DST_DEPTH) * (int)STRIDE_2;
+#endif // defined(START_2) && defined(STRIDE_2)
+#if defined(START_3) && defined(STRIDE_3)
+ offset_3 = (int)START_3 + ((int)get_global_id(2) / (int)DST_DEPTH) * (int)STRIDE_3;
+#endif // defined(START_2) && defined(STRIDE_2)
+
+ // Store result
+ *((__global DATA_TYPE *)(output.ptr)) = *((__global DATA_TYPE *)tensor4D_offset(&input, offset_0, offset_1, offset_2, offset_3));
+}
diff --git a/src/core/CL/kernels/CLStridedSliceKernel.cpp b/src/core/CL/kernels/CLStridedSliceKernel.cpp
new file mode 100644
index 0000000..f07436a
--- /dev/null
+++ b/src/core/CL/kernels/CLStridedSliceKernel.cpp
@@ -0,0 +1,189 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLStridedSliceKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Window.h"
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/helpers/tensor_transform.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
+ const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+ int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+ ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
+ DataType::U8, DataType::S8, DataType::QASYMM8,
+ DataType::U16, DataType::S16,
+ DataType::U32, DataType::S32,
+ DataType::F16, DataType::F32);
+
+ ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().num_dimensions() > 4);
+ ARM_COMPUTE_RETURN_ERROR_ON(starts.num_dimensions() > input->num_dimensions());
+ ARM_COMPUTE_RETURN_ERROR_ON(ends.num_dimensions() > input->num_dimensions());
+ ARM_COMPUTE_RETURN_ERROR_ON(strides.num_dimensions() > input->num_dimensions());
+ for(unsigned int i = 0; i < strides.num_dimensions(); ++i)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON(strides[i] == 0);
+ }
+
+ // Get expected output shape
+ const TensorShape exp_output_shape = arm_compute::misc::shape_calculator::compute_strided_slice_shape(*input,
+ starts, ends, strides,
+ begin_mask, end_mask, shrink_axis_mask);
+ ARM_COMPUTE_RETURN_ERROR_ON(exp_output_shape.total_size() == 0);
+
+ // Checks output if configured
+ if(output->total_size() != 0)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape() != exp_output_shape);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+ }
+
+ return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output,
+ const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+ int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+ // Output tensor auto initialization if not yet initialized
+ const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_strided_slice_shape(*input,
+ starts, ends, strides,
+ begin_mask, end_mask, shrink_axis_mask);
+ auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape));
+
+ // Create window
+ const unsigned int num_elems_processed_per_iteration = 1;
+
+ Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+ output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+
+ return std::make_pair(Status{}, win);
+}
+} // namespace
+
+CLStridedSliceKernel::CLStridedSliceKernel()
+ : _input(nullptr), _output(nullptr)
+{
+}
+
+void CLStridedSliceKernel::configure(const ICLTensor *input, ICLTensor *output,
+ const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+ int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), starts, ends, strides, begin_mask, end_mask, shrink_axis_mask));
+
+ _input = input;
+ _output = output;
+
+ const TensorShape &input_shape = input->info()->tensor_shape();
+
+ const Coordinates final_strides = arm_compute::helpers::tensor_transform::strided_slice_strides(input_shape, strides);
+ const Coordinates starts_abs = arm_compute::helpers::tensor_transform::strided_slice_absolute_start_coords(input_shape, starts, final_strides, begin_mask);
+ const Coordinates ends_abs = arm_compute::helpers::tensor_transform::strided_slice_absolute_end_coords(input_shape, starts_abs, ends, final_strides, end_mask, shrink_axis_mask);
+
+ // Configure kernel window
+ auto win_config = validate_and_configure_window(input->info(), output->info(), starts, ends, strides, begin_mask, end_mask, shrink_axis_mask);
+ ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+ ICLKernel::configure_internal(win_config.second);
+
+ // Create build options
+ CLBuildOptions build_opts;
+ build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+ for(unsigned int i = 0; i < input_shape.num_dimensions(); ++i)
+ {
+ build_opts.add_option("-DSTART_" + support::cpp11::to_string(i) + "=" + support::cpp11::to_string(starts_abs[i]));
+ build_opts.add_option("-DSTRIDE_" + support::cpp11::to_string(i) + "=" + support::cpp11::to_string(final_strides[i]));
+ }
+ build_opts.add_option_if_else(input_shape.num_dimensions() > 2,
+ "-DSRC_DEPTH=" + support::cpp11::to_string(input_shape.z()),
+ "-DSRC_DEPTH=1");
+ build_opts.add_option_if_else(_output->info()->num_dimensions() > 2,
+ "-DDST_DEPTH=" + support::cpp11::to_string(_output->info()->tensor_shape().z()),
+ "-DDST_DEPTH=1");
+
+ // Create kernel
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("strided_slice", build_opts.options()));
+
+ // Set config_id for enabling LWS tuning
+ _config_id = "strided_slice";
+ _config_id += "_";
+ _config_id += lower_string(string_from_data_type(input->info()->data_type()));
+ for(unsigned int i = 0; i < input_shape.num_dimensions(); ++i)
+ {
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(input->info()->dimension(i));
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(starts_abs[i]);
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(ends_abs[i]);
+ _config_id += "_";
+ _config_id += support::cpp11::to_string(final_strides[i]);
+ }
+}
+
+Status CLStridedSliceKernel::validate(const ITensorInfo *input, const ITensorInfo *output,
+ const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+ int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, starts, ends, strides, begin_mask, end_mask, shrink_axis_mask));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(),
+ starts, ends, strides, begin_mask, end_mask, shrink_axis_mask)
+ .first);
+
+ return Status{};
+}
+
+void CLStridedSliceKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+ ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+ ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+ Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+ Window slice = window_collapsed.first_slice_window_4D();
+
+ do
+ {
+ unsigned int idx = 0;
+ add_4D_tensor_argument(idx, _input, slice);
+ add_4D_tensor_argument(idx, _output, slice);
+ enqueue(queue, *this, slice, lws_hint());
+ }
+ while(window_collapsed.slide_window_slice_4D(slice));
+}
+} // namespace arm_compute