1 files changed, 91 insertions, 31 deletions

diff --git a/src/core/CL/kernels/CLCopyKernel.cpp b/src/core/CL/kernels/CLCopyKernel.cpp
index e14e5dafab..30a0b8fcb3 100644
--- a/src/core/CL/kernels/CLCopyKernel.cpp
+++ b/src/core/CL/kernels/CLCopyKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,38 +37,58 @@ namespace arm_compute
 {
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding = PaddingList())
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding = PaddingList(), Window *output_window = nullptr)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_ON(!padding.empty() && output_window != nullptr);
     ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > 4);
 
     // Validate output if initialized
     if(output->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding), output->tensor_shape());
+        if(output_window == nullptr)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding), output->tensor_shape());
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(input->tensor_shape(), output_window->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)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, Window *output_window)
 {
     // Output auto inizialitation if not yet initialized
     auto_init_if_empty(*output, *input);
 
     // Configure window
-    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
+    const unsigned int vec_size_x = 16 / input->element_size();
 
-    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    if(output_window == nullptr)
+    {
+        // Create and update the window (if needed)
+        Window win = calculate_max_window(*input, Steps(vec_size_x));
 
-    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+        AccessWindowHorizontal input_access(input, 0, vec_size_x);
+        AccessWindowHorizontal output_access(output, 0, vec_size_x);
 
-    bool window_changed = update_window_and_padding(win, input_access, output_access);
+        bool window_changed = update_window_and_padding(win, input_access, output_access);
 
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
+        Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+        return std::make_pair(err, win);
+    }
+    else
+    {
+        Window win = calculate_max_window(*input);
+        return std::make_pair(Status{}, win);
+    }
 }
 
 std::pair<Status, Window> validate_and_configure_window_with_padding(ITensorInfo *input, ITensorInfo *output, const PaddingList &padding)
@@ -131,14 +151,14 @@ void add_padding_as_build_options(const PaddingList &padding, CLBuildOptions &bu
 } // namespace
 
 CLCopyKernel::CLCopyKernel()
-    : _input(nullptr), _output(nullptr)
+    : _input(nullptr), _output(nullptr), _output_window(), _has_output_window(false)
 {
 }
 
-void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const PaddingList &padding)
+void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const PaddingList &padding, Window *output_window)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding, output_window));
 
     _input  = input;
     _output = output;
@@ -147,21 +167,44 @@ void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const Pa
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
 
-    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
-    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
-
     std::pair<Status, Window> win_config;
 
+    const unsigned int vec_size_x = 16 / input->info()->element_size();
+
     if(padding.empty())
     {
+        // Configure window
+        win_config = validate_and_configure_window(input->info(), output->info(), output_window);
+
+        if(output_window != nullptr)
+        {
+            _has_output_window        = true;
+            _output_window            = Window(*output_window);
+            const int  width_x        = output_window->num_iterations(0);
+            const bool multi_access_x = width_x >= static_cast<int32_t>(vec_size_x);
+            const bool remainder_x    = width_x % vec_size_x > 0;
+
+            if(multi_access_x)
+            {
+                _output_window.set(Window::DimX, Window::Dimension(output_window->x().start(), ceil_to_multiple(output_window->x().end(), vec_size_x), vec_size_x));
+                win_config.second.set(Window::DimX, Window::Dimension(win_config.second.x().start(), ceil_to_multiple(win_config.second.x().end(), vec_size_x), vec_size_x));
+            }
+
+            build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
+            build_opts.add_option_if(multi_access_x && remainder_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max<int>(width_x - vec_size_x, 0)));
+        }
+        else
+        {
+            build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
+        }
+
         // Build kernel
         _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("copy_tensor", build_opts.options()));
-
-        // Configure window
-        win_config = validate_and_configure_window(input->info(), output->info());
     }
     else
     {
+        build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x));
+
         // Add compile time options
         add_padding_as_build_options(padding, build_opts);
 
@@ -185,13 +228,13 @@ void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const Pa
     ICLKernel::configure_internal(win_config.second);
 }
 
-Status CLCopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, const PaddingList &padding)
+Status CLCopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, const PaddingList &padding, Window *output_window)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, padding));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, padding, output_window));
 
     if(padding.empty())
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
+        ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), output_window).first);
     }
     else
     {
@@ -206,16 +249,33 @@ void CLCopyKernel::run(const Window &window, cl::CommandQueue &queue)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 
-    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
-    Window slice     = collapsed.first_slice_window_3D();
+    Window slice;
 
-    do
+    if(_has_output_window)
+    {
+        slice            = window.first_slice_window_3D();
+        Window out_slice = _output_window.first_slice_window_3D();
+        do
+        {
+            unsigned int idx = 0;
+            add_3D_tensor_argument(idx, _input, slice);
+            add_3D_tensor_argument(idx, _output, out_slice);
+            enqueue(queue, *this, slice);
+        }
+        while(window.slide_window_slice_3D(slice) && _output_window.slide_window_slice_3D(out_slice));
+    }
+    else
     {
-        unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice);
-        add_3D_tensor_argument(idx, _output, slice);
-        enqueue(queue, *this, slice);
+        Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+        slice            = collapsed.first_slice_window_3D();
+        do
+        {
+            unsigned int idx = 0;
+            add_3D_tensor_argument(idx, _input, slice);
+            add_3D_tensor_argument(idx, _output, slice);
+            enqueue(queue, *this, slice);
+        }
+        while(collapsed.slide_window_slice_3D(slice));
     }
-    while(collapsed.slide_window_slice_3D(slice));
 }
 } // namespace arm_compute