From 995f55252dfaaa5b64075b96272ae11e23d6ca9b Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Mon, 29 Jan 2018 13:43:35 +0000 Subject: COMPMID-864 Window::collapse_if_possible() is misused in several CL kernels Removed unnecessary collapse_if_possible() calls. Change-Id: I6f3434bc4a26470c4de5bac4e3d90b4b019c2c9c Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/117993 Reviewed-by: Michalis Spyrou Tested-by: Jenkins Reviewed-by: Georgios Pinitas --- arm_compute/core/CL/kernels/CLQuantizationLayerKernel.h | 6 +++--- arm_compute/core/Window.h | 10 ++++++++++ arm_compute/core/Window.inl | 10 ++++++++++ src/core/CL/kernels/CLDequantizationLayerKernel.cpp | 4 ++-- src/core/CL/kernels/CLMinMaxLayerKernel.cpp | 15 ++++----------- src/core/CL/kernels/CLQuantizationLayerKernel.cpp | 14 +++++--------- src/core/NEON/kernels/NEDequantizationLayerKernel.cpp | 4 +--- src/core/NEON/kernels/NEMinMaxLayerKernel.cpp | 3 +-- src/core/NEON/kernels/NEQuantizationLayerKernel.cpp | 4 +--- src/core/NEON/kernels/NEReshapeLayerKernel.cpp | 3 +-- 10 files changed, 38 insertions(+), 35 deletions(-) diff --git a/arm_compute/core/CL/kernels/CLQuantizationLayerKernel.h b/arm_compute/core/CL/kernels/CLQuantizationLayerKernel.h index 13092400c7..a24ddb1a3a 100644 --- a/arm_compute/core/CL/kernels/CLQuantizationLayerKernel.h +++ b/arm_compute/core/CL/kernels/CLQuantizationLayerKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -51,8 +51,8 @@ public: ~CLQuantizationLayerKernel() = default; /** Set the input, output, min and max. * - * @param[in] input Source tensor. Data types supported: F32. - * @param[out] output Destination tensor. Data types supported: U8. + * @param[in] input Source tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: F32. + * @param[out] output Destination tensor with the same dimensions of input. Output data type must be U8. * @param[in] min_max Pointer to the tensor with shape [2, batches] which stores the minimum and maximum value for each 3D input tensor. * The dimensions over the second must match the batched dimensions of the input tensor. Data type supported: F32. */ diff --git a/arm_compute/core/Window.h b/arm_compute/core/Window.h index cca12c9efe..5ca210a112 100644 --- a/arm_compute/core/Window.h +++ b/arm_compute/core/Window.h @@ -177,6 +177,16 @@ public: */ void shift(size_t dimension, int shift_value); + /** Shift down all the dimensions of a window + * + * i.e new_dims[n] = old_dims[n+shift_value]. + * + * @param[in] shift_value Number of dimensions to shift the window by. + * + * @return The window with the shifted dimensions. + */ + Window shift_dimensions(unsigned int shift_value) const; + /** Adjust the start or end of a given dimension by the given value * * @param[in] dimension The dimension to adjust diff --git a/arm_compute/core/Window.inl b/arm_compute/core/Window.inl index 23b2a8e322..18d454a0cb 100644 --- a/arm_compute/core/Window.inl +++ b/arm_compute/core/Window.inl @@ -77,6 +77,16 @@ inline Window Window::collapse_if_possible(const Window &full_window, const size return collapsed; } +inline Window Window::shift_dimensions(unsigned int shift_value) const +{ + Window shifted_window; + for(size_t n = 0; n < (Coordinates::num_max_dimensions - shift_value); n++) + { + shifted_window.set(n, _dims[n + shift_value]); + } + return shifted_window; +} + inline Window Window::collapse(const Window &full_window, const size_t first, const size_t last) const { bool has_collapsed = false; diff --git a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp index 216fa2757e..4efdb764bd 100644 --- a/src/core/CL/kernels/CLDequantizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLDequantizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -97,5 +97,5 @@ void CLDequantizationLayerKernel::run(const Window &window, cl::CommandQueue &qu add_1D_tensor_argument(idx, _min_max, min_max_slice); enqueue(queue, *this, slice); } - while(window.slide_window_slice_3D(slice) && min_max_window.slide_window_slice_1D(min_max_slice)); + while(window_collapsed.slide_window_slice_3D(slice) && min_max_window.slide_window_slice_1D(min_max_slice)); } diff --git a/src/core/CL/kernels/CLMinMaxLayerKernel.cpp b/src/core/CL/kernels/CLMinMaxLayerKernel.cpp index 9b4533bd8d..8ba1f776a1 100644 --- a/src/core/CL/kernels/CLMinMaxLayerKernel.cpp +++ b/src/core/CL/kernels/CLMinMaxLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -89,7 +89,6 @@ void CLMinMaxLayerKernel::reset(cl::CommandQueue &queue) Window window_output; window_output.use_tensor_dimensions(_output->info()->tensor_shape()); window_output.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_output.collapse_if_possible(ICLKernel::window(), 1); Iterator output(_output, window_output); @@ -110,27 +109,21 @@ void CLMinMaxLayerKernel::run(const Window &window, cl::CommandQueue &queue) ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); - // Collapse min/max batches Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), 3); Window slice = window_collapsed.first_slice_window_3D(); slice.set(Window::DimX, Window::Dimension(0, 1, 1)); slice.set(Window::DimY, Window::Dimension(0, 1, 1)); slice.set(Window::DimZ, Window::Dimension(0, 1, 1)); - Window window_output; - window_output.use_tensor_dimensions(_output->info()->tensor_shape()); - window_output.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_output.collapse_if_possible(ICLKernel::window(), 1); - - Window output_slice = window_output.first_slice_window_1D(); - do { + Window output_slice = slice.shift_dimensions(2); + unsigned int idx = 0; // Set inputs add_3D_tensor_argument(idx, _input, slice); add_1D_tensor_argument(idx, _output, output_slice); enqueue(queue, *this, slice); } - while(window.slide_window_slice_3D(slice) && window_output.slide_window_slice_1D(output_slice)); + while(window_collapsed.slide_window_slice_3D(slice)); } diff --git a/src/core/CL/kernels/CLQuantizationLayerKernel.cpp b/src/core/CL/kernels/CLQuantizationLayerKernel.cpp index 47564436a9..8b082a8704 100644 --- a/src/core/CL/kernels/CLQuantizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLQuantizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -82,20 +82,16 @@ void CLQuantizationLayerKernel::run(const Window &window, cl::CommandQueue &queu Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), 3); Window slice = window_collapsed.first_slice_window_3D(); - Window window_min_max; - window_min_max.use_tensor_dimensions(_min_max->info()->tensor_shape()); - window_min_max.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_min_max.collapse_if_possible(ICLKernel::window(), 1); - - Window slice_min_max = window_min_max.first_slice_window_1D(); - do { + Window slice_min_max = slice.shift_dimensions(2); + slice_min_max.set(Window::DimX, Window::Dimension(0, 1, 1)); + unsigned int idx = 0; add_3D_tensor_argument(idx, _input, slice); add_3D_tensor_argument(idx, _output, slice); add_1D_tensor_argument(idx, _min_max, slice_min_max); enqueue(queue, *this, slice); } - while(window.slide_window_slice_3D(slice) && window_min_max.slide_window_slice_1D(slice_min_max)); + while(window_collapsed.slide_window_slice_3D(slice)); } diff --git a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp index 70984f0a75..be211b2cb2 100644 --- a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -78,13 +78,11 @@ void NEDequantizationLayerKernel::run(const Window &window, const ThreadInfo &in ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); Window window_input_output(window); - window_input_output.collapse_if_possible(INEKernel::window(), 3); window_input_output.set(3, Window::Dimension(0, 1, 1)); Window window_min_max; window_min_max.use_tensor_dimensions(_min_max->info()->tensor_shape()); window_min_max.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_min_max.collapse_if_possible(INEKernel::window(), 1); Iterator input(_input, window_input_output); Iterator output(_output, window_input_output); diff --git a/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp b/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp index a81725fe84..01be36b1c6 100644 --- a/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp +++ b/src/core/NEON/kernels/NEMinMaxLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -96,7 +96,6 @@ void NEMinMaxLayerKernel::run(const Window &window, const ThreadInfo &info) // First one will use vector operations, second one processes the left over pixels Window window_input(window); window_input.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_input.collapse_if_possible(INEKernel::window(), 3); window_input.set(3, Window::Dimension(0, 1, 1)); Iterator input(_input, window_input); diff --git a/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp b/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp index bff79f0f0c..767af08d0d 100644 --- a/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEQuantizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -77,13 +77,11 @@ void NEQuantizationLayerKernel::run(const Window &window, const ThreadInfo &info ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); Window window_input_output(window); - window_input_output.collapse_if_possible(INEKernel::window(), 3); window_input_output.set(3, Window::Dimension(0, 1, 1)); Window window_min_max; window_min_max.use_tensor_dimensions(_min_max->info()->tensor_shape()); window_min_max.set(Window::DimX, Window::Dimension(0, 1, 1)); - window_min_max.collapse_if_possible(INEKernel::window(), 1); Iterator input(_input, window_input_output); Iterator output(_output, window_input_output); diff --git a/src/core/NEON/kernels/NEReshapeLayerKernel.cpp b/src/core/NEON/kernels/NEReshapeLayerKernel.cpp index a0f324ef18..45ba68d9fa 100644 --- a/src/core/NEON/kernels/NEReshapeLayerKernel.cpp +++ b/src/core/NEON/kernels/NEReshapeLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ inline void reshape_tensor(const Window &window, const ITensor *input, ITensor * const TensorShape &output_shape = output->info()->tensor_shape(); Coordinates output_coord{}; - window.collapse_if_possible(window, 3); Iterator in(input, window); execute_window_loop(window, [&](const Coordinates & id) -- cgit v1.2.1