COMPMID-1727 - CL: Implement Gather
Change-Id: I3d859da09a4de1019bb8c2046725eab942247927
Reviewed-on: https://review.mlplatform.org/386
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 470a50f..2bc2d06 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -262,6 +262,7 @@
{ "finalize", "optical_flow_pyramid_lk.cl" },
{ "fuse_batchnormalization_layer", "batchnormalization_layer.cl" },
{ "floor_layer", "floor.cl" },
+ { "gather", "gather.cl" },
{ "gaussian1x5_sub_x", "gaussian_pyramid.cl" },
{ "gaussian5x1_sub_y", "gaussian_pyramid.cl" },
{ "gemm_accumulate_biases", "gemm.cl" },
@@ -684,6 +685,10 @@
#include "./cl_kernels/floor.clembed"
},
{
+ "gather.cl",
+#include "./cl_kernels/gather.clembed"
+ },
+ {
"gaussian_pyramid.cl",
#include "./cl_kernels/gaussian_pyramid.clembed"
},
diff --git a/src/core/CL/cl_kernels/gather.cl b/src/core/CL/cl_kernels/gather.cl
new file mode 100644
index 0000000..34593ef
--- /dev/null
+++ b/src/core/CL/cl_kernels/gather.cl
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2018-2019 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"
+
+#if defined(DATA_TYPE) && defined(AXIS)
+
+/** Performs the Gather operation along the chosen axis
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
+ * @note Axis should be given as a preprocessor argument using -DAXIS=axis. e.g. -DAXIS=1
+ * @attention Output tensor depth should be given as a preprocessor argument using -DOUTPUT_DIM_Z=size. e.g. -DOUTPUT_DIM_Z=16
+ * @attention Input tensor depth should be given as a preprocessor argument using -DINPUT_DIM_Z=size. e.g. -DINPUT_DIM_Z=16
+ *
+ *
+ * @param[in] input_ptr Pointer to the source tensor. Supported data types: U8/S8/U16/S16/U32/S32/F16/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 work item (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 work item (in bytes)
+ * @param[in] input_stride_z Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] input_step_z input_stride_z * number of elements along Z processed per work item (in bytes)
+ * @param[in] input_stride_w Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] input_step_w input_stride_w * number of elements along W processed per work item (in bytes)
+ * @param[in] input_offset_first_element_in_bytes Offset of the first element in the source tensor
+ * @param[in] indices_ptr Pointer to the indices vector. Supported data types: U32.
+ * @param[in] indices_stride_x Stride of the indices vector in X dimension (in bytes)
+ * @param[in] indices_step_x input_stride_x * number of elements along X processed per work item (in bytes)
+ * @param[in] indices_offset_first_element_in_bytes Offset of the first element in the indices vector
+ * @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 work item (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 work item (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 work item (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 work item (in bytes)
+ * @param[in] output_offset_first_element_in_bytes Offset of the first element in the destination tensor
+ */
+__kernel void gather(
+ TENSOR4D_DECLARATION(input),
+ VECTOR_DECLARATION(indices),
+ TENSOR4D_DECLARATION(output))
+{
+ const int px = get_global_id(0);
+ const int py = get_global_id(1);
+ const int pz = get_global_id(2) % OUTPUT_DIM_Z;
+ const int pw = get_global_id(2) / OUTPUT_DIM_Z;
+
+ const Tensor4D input = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, INPUT_DIM_Z);
+ const Vector indices = CONVERT_TO_VECTOR_STRUCT_NO_STEP(indices);
+ Tensor4D output = CONVERT_TO_TENSOR4D_STRUCT(output, OUTPUT_DIM_Z);
+
+#if AXIS == 0
+ const uint index = *(__global const uint *)vector_offset(&indices, px);
+ __global const uchar *input_addr = tensor4D_offset(&input, index, py, pz, pw);
+#elif AXIS == 1
+ const uint index = *(__global const uint *)vector_offset(&indices, py);
+ __global const uchar *input_addr = tensor4D_offset(&input, px, index, pz, pw);
+#elif AXIS == 2
+ const uint index = *(__global const uint *)vector_offset(&indices, pz);
+ __global const uchar *input_addr = tensor4D_offset(&input, px, py, index, pw);
+#elif AXIS == 3
+ const uint index = *(__global const uint *)vector_offset(&indices, pw);
+ __global const uchar *input_addr = tensor4D_offset(&input, px, py, pz, index);
+#endif //AXIS
+
+ *(__global DATA_TYPE *)output.ptr = *((__global const DATA_TYPE *)input_addr);
+}
+
+#endif //defined(DATA_TYPE) && defined(AXIS)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLGatherKernel.cpp b/src/core/CL/kernels/CLGatherKernel.cpp
new file mode 100644
index 0000000..006e755
--- /dev/null
+++ b/src/core/CL/kernels/CLGatherKernel.cpp
@@ -0,0 +1,135 @@
+/*
+ * Copyright (c) 2018-2019 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/CLGatherKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.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/CL/OpenCL.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include <string>
+
+namespace arm_compute
+{
+namespace
+{
+inline Status validate_arguments(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis)
+{
+ const uint32_t actual_axis = wrap_around(axis, static_cast<int>(input->num_dimensions()));
+ ARM_COMPUTE_RETURN_ERROR_ON(indices->num_dimensions() > 1);
+ ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+ ARM_COMPUTE_RETURN_ERROR_ON(actual_axis >= input->num_dimensions());
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+ 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);
+
+ if(output->total_size() != 0)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
+ TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->tensor_shape(), indices->tensor_shape(), actual_axis);
+ ARM_COMPUTE_RETURN_ERROR_ON(output_shape.total_size() != output->tensor_shape().total_size());
+ }
+
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(indices, 1, DataType::U32);
+
+ return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *indices, ITensorInfo *output, int axis)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, indices);
+ const uint32_t actual_axis = wrap_around(axis, static_cast<int>(input->num_dimensions()));
+ // Output auto initialization if not yet initialized
+ TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->tensor_shape(), indices->tensor_shape(), actual_axis);
+ auto_init_if_empty((*output), output_shape, 1, input->data_type());
+
+ // Create window
+ Window win = calculate_max_window(*output, Steps());
+ output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+
+ return std::make_pair(Status{}, win);
+}
+
+} // namespace
+
+CLGatherKernel::CLGatherKernel()
+ : _input(nullptr), _indices(nullptr), _output(nullptr), _axis(0)
+{
+}
+
+void CLGatherKernel::configure(const ICLTensor *input, const ICLTensor *indices, ICLTensor *output, int axis)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, indices);
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), indices->info(), output->info(), axis));
+
+ // Configure kernel window
+ auto win_config = validate_and_configure_window(input->info(), indices->info(), output->info(), axis);
+ ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+
+ _input = input;
+ _output = output;
+ _indices = indices;
+ _axis = wrap_around(axis, static_cast<int>(input->info()->num_dimensions()));
+
+ // Set build options
+ CLBuildOptions build_opts;
+ build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+ build_opts.add_option("-DOUTPUT_DIM_Z=" + support::cpp11::to_string(output->info()->dimension(2)));
+ build_opts.add_option("-DINPUT_DIM_Z=" + support::cpp11::to_string(input->info()->dimension(2)));
+ build_opts.add_option("-DAXIS=" + support::cpp11::to_string(_axis));
+
+ // Create kernel
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gather", build_opts.options()));
+ ICLKernel::configure_internal(win_config.second);
+}
+
+Status CLGatherKernel::validate(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis)
+{
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, indices, output, axis));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), indices->clone().get(), output->clone().get(), axis).first);
+ return Status{};
+}
+
+void CLGatherKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+ ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+ ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+ Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+ unsigned int idx = 0;
+ add_4D_tensor_argument(idx, _input, window_collapsed);
+ add_1D_tensor_argument(idx, _indices, window_collapsed);
+ add_4D_tensor_argument(idx, _output, window_collapsed);
+ enqueue(queue, *this, window_collapsed);
+}
+} // namespace arm_compute