From 920f2b6c2070f6328891e26538e8bcad63e2a79c Mon Sep 17 00:00:00 2001 From: Pablo Marquez Tello Date: Wed, 27 Apr 2022 11:46:31 +0100 Subject: Add support for 2d and 3d indices for axis 0 * Partially resolves COMPMID-5055 Change-Id: Id05374b8c69e6b9ab4c2790a4de93d7172063b71 Signed-off-by: Pablo Marquez Tello Change-Id: Ic6e2c2d1d34abbf6222c8d56859514e267447266 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7488 Tested-by: Arm Jenkins Reviewed-by: Giorgio Arena Comments-Addressed: Arm Jenkins --- src/core/NEON/kernels/NEGatherKernel.cpp | 129 ++++++++++++++++++++++++------- src/core/NEON/kernels/NEGatherKernel.h | 15 ++-- 2 files changed, 113 insertions(+), 31 deletions(-) (limited to 'src/core/NEON/kernels') diff --git a/src/core/NEON/kernels/NEGatherKernel.cpp b/src/core/NEON/kernels/NEGatherKernel.cpp index 7090da8015..8d86a22b7e 100644 --- a/src/core/NEON/kernels/NEGatherKernel.cpp +++ b/src/core/NEON/kernels/NEGatherKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,16 +47,21 @@ namespace template void validate_indices(const ITensor *indices) { - for(size_t i = 0; i < indices->info()->tensor_shape()[0]; ++i) + auto *indices_ptr = (reinterpret_cast( indices->buffer() + indices->info()->offset_first_element_in_bytes() )); + for(size_t i = 0; i < indices->info()->total_size(); ++i) { - ARM_COMPUTE_ERROR_ON(*(reinterpret_cast(indices->ptr_to_element(Coordinates(i)))) < 0); + const U index_value = indices_ptr[i]; + ARM_COMPUTE_UNUSED(index_value); + if(index_value < 0) + { + ARM_COMPUTE_ERROR_ON(index_value < 0); + } } } Status validate_arguments(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, indices, output); - ARM_COMPUTE_RETURN_ERROR_ON(indices->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); if(axis < 0) @@ -65,6 +70,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *indices, } ARM_COMPUTE_RETURN_ERROR_ON(0 > axis || axis >= static_cast(input->num_dimensions())); + ARM_COMPUTE_RETURN_ERROR_ON(axis != 0 && indices->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); if(output->total_size() != 0) @@ -86,6 +92,40 @@ NEGatherKernel::NEGatherKernel() { } +template +inline void NEGatherKernel::gather_dims_0_axis(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON(_indices->info()->num_dimensions() < 2); + validate_indices(_indices); + + Window output_window{ window }; + output_window.set(Window::DimX, Window::Dimension(0, 1, 1)); + Iterator output_it(_output, output_window); + + const uint8_t *const in_ptr_start = _input->buffer() + _input->info()->offset_first_element_in_bytes(); + const uint32_t input_stride_y = _input->info()->strides_in_bytes()[1]; + const uint32_t output_stride_y = _output->info()->strides_in_bytes()[1]; + + const U *const dex_ptr_start = reinterpret_cast(_indices->buffer() + _indices->info()->offset_first_element_in_bytes()); + execute_window_loop(output_window, [&](const Coordinates & id) + { + const auto new_index = *(dex_ptr_start + id.y() + id.z() * _output->info()->tensor_shape()[1] + id[3] * _indices->info()->tensor_shape()[1] * _indices->info()->tensor_shape()[0]); + U *out_ptr = reinterpret_cast(output_it.ptr()); + const char *const in_ptr = reinterpret_cast(in_ptr_start + new_index * input_stride_y); + memcpy(out_ptr, in_ptr, output_stride_y); + }, + output_it); +} + +template +inline void NEGatherKernel::gather_dims_n_axis(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_UNUSED(window); + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); +} + template inline void NEGatherKernel::gather_0_axis(const Window &window, const ThreadInfo &info) { @@ -147,38 +187,75 @@ void NEGatherKernel::configure(const ITensor *input, const ITensor *indices, ITe } ARM_COMPUTE_ERROR_ON(0 > _axis || _axis >= static_cast(input->info()->num_dimensions())); - if(0 == _axis) + if(indices->info()->num_dimensions() == 1u) { - switch(_indices->info()->data_type()) + if(0 == _axis) + { + switch(_indices->info()->data_type()) + { + case DataType::U32: + _func = &NEGatherKernel::gather_0_axis; + break; + case DataType::S32: + _func = &NEGatherKernel::gather_0_axis; + break; + default: + ARM_COMPUTE_ERROR("Not supported"); + break; + } + } + else { - case DataType::U32: - _func = &NEGatherKernel::gather_0_axis; - break; - case DataType::S32: - _func = &NEGatherKernel::gather_0_axis; - break; - default: - ARM_COMPUTE_ERROR("Not supported"); - break; + switch(_indices->info()->data_type()) + { + case DataType::U32: + _func = &NEGatherKernel::gather_n_axis; + break; + case DataType::S32: + _func = &NEGatherKernel::gather_n_axis; + break; + default: + ARM_COMPUTE_ERROR("Not supported"); + break; + } } } else { - switch(_indices->info()->data_type()) + if(0 == _axis) { - case DataType::U32: - _func = &NEGatherKernel::gather_n_axis; - break; - case DataType::S32: - _func = &NEGatherKernel::gather_n_axis; - break; - default: - ARM_COMPUTE_ERROR("Not supported"); - break; + switch(_indices->info()->data_type()) + { + case DataType::U32: + _func = &NEGatherKernel::gather_dims_0_axis; + break; + case DataType::S32: + _func = &NEGatherKernel::gather_dims_0_axis; + break; + default: + ARM_COMPUTE_ERROR("Not supported"); + break; + } + } + else + { + switch(_indices->info()->data_type()) + { + case DataType::U32: + _func = &NEGatherKernel::gather_dims_n_axis; + break; + case DataType::S32: + _func = &NEGatherKernel::gather_dims_n_axis; + break; + default: + ARM_COMPUTE_ERROR("Not supported"); + break; + } } } + // Output auto initialization if not yet initialized - TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->info()->tensor_shape(), indices->info()->tensor_shape(), _axis); + const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->info()->tensor_shape(), indices->info()->tensor_shape(), _axis); auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape)); // Create window diff --git a/src/core/NEON/kernels/NEGatherKernel.h b/src/core/NEON/kernels/NEGatherKernel.h index 0711f8190b..fc0e67854b 100644 --- a/src/core/NEON/kernels/NEGatherKernel.h +++ b/src/core/NEON/kernels/NEGatherKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -85,18 +85,23 @@ private: * * For gather on the 0 axis an element by element copy is performed. * - * @param[in] window Region on which to execute the kernel. (Must be a region of the window returned by window()) - * @param[in] info Info about executing thread and CPU. + * @param[in] window Region on which to run the kernel. (Must be a region of the window returned by window()) + * @param[in] info Info about running thread and CPU. */ template void gather_0_axis(const Window &window, const ThreadInfo &info); + template + void gather_dims_0_axis(const Window &window, const ThreadInfo &info); + + template + void gather_dims_n_axis(const Window &window, const ThreadInfo &info); /** Implementation of the gather operation. * * For 1<=axis a row-wise copy is taking place. * - * @param[in] window Region on which to execute the kernel. (Must be a region of the window returned by window()) - * @param[in] info Info about executing thread and CPU. + * @param[in] window Region on which to run the kernel. (Must be a region of the window returned by window()) + * @param[in] info Info about running thread and CPU. */ template void gather_n_axis(const Window &window, const ThreadInfo &info); -- cgit v1.2.1