COMPMID-462: Implement TensorReshape for NEON and CL.

Change-Id: I11b39c2ceca26ade73822e29a384ef866ae05729
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/87707
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

6 files changed, 373 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 696fcb475c..6602b4dc78 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -243,6 +243,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "reduction_operation", "reduction_operation.cl" },
     { "remap_nearest_neighbour", "remap.cl" },
     { "remap_bilinear", "remap.cl" },
+    { "reshape_layer", "reshape_layer.cl" },
     { "reshape_to_columns", "convolution_layer.cl" },
     { "RGB888_to_IYUV_bt709", "color_convert.cl" },
     { "RGB888_to_NV12_bt709", "color_convert.cl" },
@@ -504,6 +505,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/remap.clembed"
     },
     {
+        "reshape_layer.cl",
+#include "./cl_kernels/reshape_layer.clembed"
+    },
+    {
         "roi_pooling_layer.cl",
 #include "./cl_kernels/roi_pooling_layer.clembed"
     },
diff --git a/src/core/CL/cl_kernels/reshape_layer.cl b/src/core/CL/cl_kernels/reshape_layer.cl
new file mode 100644
index 0000000000..23eccbf817
--- /dev/null
+++ b/src/core/CL/cl_kernels/reshape_layer.cl
@@ -0,0 +1,70 @@
+/*
+ * Copyright (c) 2017 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 tensor reshape
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
+ *
+ * @param[in]  input_ptr                            Pointer to the first source tensor. Supported data types: U8/S8/QS8/U16/S16/QS16/U32/S32/F16/F32
+ * @param[in]  input_stride_x                       Stride of the first 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 first 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 first 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_offset_first_element_in_bytes  The offset of the first element in the first 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_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in]  input_shape                          Input spatial shape
+ * @param[in]  output_shape                         Output spatial shape
+ */
+__kernel void reshape_layer(TENSOR3D_DECLARATION(input),
+                            TENSOR3D_DECLARATION(output),
+                            int2 input_shape,
+                            int2 output_shape)
+{
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
+
+    int3 id = (int3)(get_global_id(0), get_global_id(1), get_global_id(2));
+
+    // Linearize index
+    int linear_idx = id.x + id.y * input_shape.x + id.z * input_shape.x * input_shape.y;
+
+    // Translate to output
+    int3 out_id;
+    out_id.x = linear_idx % output_shape.x;
+    out_id.y = (linear_idx / output_shape.x) % output_shape.y;
+    out_id.z = linear_idx / (output_shape.x * output_shape.y);
+
+    // Store result
+    *((__global DATA_TYPE *)tensor3D_offset(&out, out_id.x, out_id.y, out_id.z)) = *((__global DATA_TYPE *)in.ptr);
+}
diff --git a/src/core/CL/kernels/CLReshapeLayerKernel.cpp b/src/core/CL/kernels/CLReshapeLayerKernel.cpp
new file mode 100644
index 0000000000..0131bd3856
--- /dev/null
+++ b/src/core/CL/kernels/CLReshapeLayerKernel.cpp
@@ -0,0 +1,109 @@
+/*
+ * Copyright (c) 2017 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/CLReshapeLayerKernel.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/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/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <string>
+
+using namespace arm_compute;
+
+CLReshapeLayerKernel::CLReshapeLayerKernel()
+    : _input(nullptr), _output(nullptr)
+{
+}
+
+void CLReshapeLayerKernel::configure(const ICLTensor *input, ICLTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16,
+                                                  DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    ARM_COMPUTE_ERROR_ON(input->info()->tensor_shape().total_size() != output->info()->tensor_shape().total_size());
+
+    _input  = input;
+    _output = output;
+
+    constexpr unsigned int num_elems_processed_per_iteration = 1;
+
+    // Create kernel
+    std::set<std::string> build_opts = { "-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()) };
+    _kernel                          = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("reshape_layer", build_opts));
+
+    // Add static arguments
+    const cl_int2 input_shape =
+    {
+        {
+            static_cast<cl_int>(_input->info()->tensor_shape()[0]),
+            static_cast<cl_int>(_input->info()->tensor_shape()[1])
+        }
+    };
+    const cl_int2 output_shape =
+    {
+        {
+            static_cast<cl_int>(_output->info()->tensor_shape()[0]),
+            static_cast<cl_int>(_output->info()->tensor_shape()[1])
+        }
+    };
+    unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters
+    _kernel.setArg<cl_int2>(idx++, input_shape);
+    _kernel.setArg<cl_int2>(idx++, output_shape);
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
+    AccessWindowStatic     output_access(output->info(), 0, 0, output->info()->tensor_shape().x(), output->info()->tensor_shape().y());
+    update_window_and_padding(win, input_access, output_access);
+
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    ICLKernel::configure(win);
+}
+
+void CLReshapeLayerKernel::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);
+    Window slice            = window_collapsed.first_slice_window_3D();
+
+    // Set inputs
+    unsigned int idx = 0;
+    add_3D_tensor_argument(idx, _input, window_collapsed);
+    add_3D_tensor_argument(idx, _output, window_collapsed);
+    enqueue(queue, *this, slice);
+}
diff --git a/src/core/NEON/kernels/NEReshapeLayerKernel.cpp b/src/core/NEON/kernels/NEReshapeLayerKernel.cpp
new file mode 100644
index 0000000000..8e69252505
--- /dev/null
+++ b/src/core/NEON/kernels/NEReshapeLayerKernel.cpp
@@ -0,0 +1,114 @@
+/*
+ * Copyright (c) 2017 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/NEON/kernels/NEReshapeLayerKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Validate.h"
+
+#include <arm_neon.h>
+#include <cstdint>
+
+using namespace arm_compute;
+
+namespace
+{
+template <typename T>
+inline void reshape_tensor(const Window &window, const ITensor *input, ITensor *output)
+{
+    const TensorShape &input_shape  = input->info()->tensor_shape();
+    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)
+    {
+        output_coord                                                 = index2coords(output_shape, coords2index(input_shape, id));
+        *reinterpret_cast<T *>(output->ptr_to_element(output_coord)) = *reinterpret_cast<T *>(in.ptr());
+    },
+    in);
+}
+} // namespace
+
+void NEReshapeLayerKernel::configure(const ITensor *input, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16,
+                                                  DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    ARM_COMPUTE_ERROR_ON(input->info()->tensor_shape().total_size() != output->info()->tensor_shape().total_size());
+
+    _input  = input;
+    _output = output;
+
+    constexpr unsigned int num_elems_processed_per_iteration = 1;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
+    AccessWindowStatic     output_access(output->info(), 0, 0, output->info()->tensor_shape().x(), output->info()->tensor_shape().y());
+    update_window_and_padding(win, input_access, output_access);
+
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEReshapeLayerKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INESimpleKernel::window(), window);
+
+    switch(_input->info()->data_type())
+    {
+        case DataType::U8:
+        case DataType::S8:
+        case DataType::QS8:
+            reshape_tensor<uint8_t>(window, _input, _output);
+            break;
+        case DataType::U16:
+        case DataType::S16:
+        case DataType::QS16:
+        case DataType::F16:
+            reshape_tensor<uint16_t>(window, _input, _output);
+            break;
+        case DataType::U32:
+        case DataType::S32:
+        case DataType::F32:
+            reshape_tensor<uint32_t>(window, _input, _output);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported data type!");
+    }
+}
diff --git a/src/runtime/CL/functions/CLReshapeLayer.cpp b/src/runtime/CL/functions/CLReshapeLayer.cpp
new file mode 100644
index 0000000000..2ce83dceb4
--- /dev/null
+++ b/src/runtime/CL/functions/CLReshapeLayer.cpp
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2017 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/runtime/CL/functions/CLReshapeLayer.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/kernels/CLReshapeLayerKernel.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+void CLReshapeLayer::configure(const ICLTensor *input, ICLTensor *output)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLReshapeLayerKernel>();
+    k->configure(input, output);
+    _kernel = std::move(k);
+}
diff --git a/src/runtime/NEON/functions/NEReshapeLayer.cpp b/src/runtime/NEON/functions/NEReshapeLayer.cpp
new file mode 100644
index 0000000000..fef4e0cc19
--- /dev/null
+++ b/src/runtime/NEON/functions/NEReshapeLayer.cpp
@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2017 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/runtime/NEON/functions/NEReshapeLayer.h"
+
+#include "arm_compute/core/NEON/kernels/NEReshapeLayerKernel.h"
+#include "support/ToolchainSupport.h"
+
+#include <utility>
+
+using namespace arm_compute;
+
+void NEReshapeLayer::configure(const ITensor *input, ITensor *output)
+{
+    auto k = arm_compute::support::cpp14::make_unique<NEReshapeLayerKernel>();
+    k->configure(input, output);
+    _kernel = std::move(k);
+}