COMPMID-1725: Implement Pack

Change-Id: I13f6e4c600f39355f69e015409bf30dafdc5e3aa
Reviewed-on: https://review.mlplatform.org/332
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

4 files changed, 348 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index ac1d4b349e..e48ff03e05 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -423,6 +423,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "space_to_batch_nhwc", "space_to_batch.cl" },
     { "space_to_batch_static_nhwc", "space_to_batch.cl" },
     { "softmax_layer_max_shift_exp_sum_parallel", "softmax_layer.cl" },
+    { "stack_layer", "stack_layer.cl" },
     { "strided_slice", "slice_ops.cl" },
     { "suppress_non_maximum", "canny.cl" },
     { "tablelookup_U8", "tablelookup.cl" },
@@ -865,6 +866,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/space_to_batch.clembed"
     },
     {
+        "stack_layer.cl",
+#include "./cl_kernels/stack_layer.clembed"
+    },
+    {
         "tablelookup.cl",
 #include "./cl_kernels/tablelookup.clembed"
     },
diff --git a/src/core/CL/cl_kernels/stack_layer.cl b/src/core/CL/cl_kernels/stack_layer.cl
new file mode 100644
index 0000000000..bed62662ad
--- /dev/null
+++ b/src/core/CL/cl_kernels/stack_layer.cl
@@ -0,0 +1,113 @@
+/*
+ * Copyright (c) 2018 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) && defined(SRC_DIM2) && defined(DST_DIM3)
+
+#if AXIS == 0
+#define X_DST (idx_input)
+#define Y_DST (x_src)
+#define Z_DST (y_src)
+#define W_DST (z_src)
+#define K_DST (w_src)
+#elif AXIS == 1 // AXIS == 1
+#define X_DST (x_src)
+#define Y_DST (idx_input)
+#define Z_DST (y_src)
+#define W_DST (z_src)
+#define K_DST (w_src)
+#elif AXIS == 2 // AXIS == 2
+#define X_DST (x_src)
+#define Y_DST (y_src)
+#define Z_DST (idx_input)
+#define W_DST (z_src)
+#define K_DST (w_src)
+#elif AXIS == 3 // AXIS == 3
+#define X_DST (x_src)
+#define Y_DST (y_src)
+#define Z_DST (z_src)
+#define W_DST (idx_input)
+#define K_DST (w_src)
+#elif AXIS == 4 // AXIS == 4
+#define X_DST (x_src)
+#define Y_DST (y_src)
+#define Z_DST (z_src)
+#define W_DST (w_src)
+#define K_DST (idx_input)
+#else // AXIS not supported
+#error "Not supported axis"
+#endif // AXIS == 0
+
+/** OpenCL kernel to stack a rank-R tensor into one with rank-(R+1) along the axis dimension
+ *
+ * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
+ * @note The dimension to stack the tensors along has to be passed at compile time using -DAXIS. i.e. -DAXIS=1
+ * @note Dimension 2 of the input tensor must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM2=112)
+ * @note Dimension 3 of the output tensor must be passed at compile time using -DDST_DIM3 (e.g. -DDST_DIM3=112)
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  src_step_w                        src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_stride_w                      Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  dst_step_w                        dst_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in]  idx_input                         Index of the input tensor in the list of tensors to stack
+ */
+__kernel void stack_layer(
+    TENSOR4D_DECLARATION(src),
+    TENSOR4D_DECLARATION(dst),
+    unsigned int idx_input)
+{
+    uint x_src = get_global_id(0);
+    uint y_src = get_global_id(1);
+    uint z_src = (get_global_id(2) % SRC_DIM2);
+    uint w_src = (get_global_id(2) / SRC_DIM2);
+
+    __global DATA_TYPE *src = (__global DATA_TYPE *)(src_ptr + src_offset_first_element_in_bytes + x_src * sizeof(DATA_TYPE) + y_src * src_stride_y + z_src * src_stride_z + w_src * src_stride_w);
+
+    __global DATA_TYPE *dst = (__global DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + X_DST * sizeof(DATA_TYPE) + Y_DST * dst_stride_y + Z_DST * dst_stride_z + W_DST * dst_stride_w + K_DST *
+                                                     dst_stride_w * (uint)DST_DIM3);
+
+    *dst = *src;
+}
+
+#undef X_DST
+#undef Y_DST
+#undef Z_DST
+#undef W_DST
+#endif // defined(DATA_TYPE) && defined(AXIS) && defined(SRC_DIM2) && defined(DST_DIM3)
diff --git a/src/core/CL/kernels/CLStackLayerKernel.cpp b/src/core/CL/kernels/CLStackLayerKernel.cpp
new file mode 100644
index 0000000000..bac8992f7b
--- /dev/null
+++ b/src/core/CL/kernels/CLStackLayerKernel.cpp
@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 2018 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/CLStackLayerKernel.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/Error.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/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, unsigned int axis, unsigned int idx_input, unsigned int num_tensors, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::U8, DataType::S8,
+                                                         DataType::U16, DataType::S16, DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(idx_input >= num_tensors);
+    ARM_COMPUTE_RETURN_ERROR_ON(axis > input->num_dimensions());
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_stack_shape(*input, axis, num_tensors));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int axis, unsigned int num_tensors, ITensorInfo *output)
+{
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_stack_shape(*input, axis, num_tensors)));
+
+    // Configure kernel window
+    constexpr unsigned int num_elems_processed_per_iteration = 1;
+
+    // The window needs to be based on input as we copy all the depths of input
+    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+
+    bool window_changed = update_window_and_padding(win, input_access, output_access);
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+CLStackLayerKernel::CLStackLayerKernel()
+    : _input(nullptr), _output(nullptr)
+{
+}
+
+void CLStackLayerKernel::configure(const ICLTensor *input, unsigned int axis, unsigned int idx_input, unsigned int num_tensors, ICLTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), axis, idx_input, num_tensors, output->info()));
+
+    _input  = input;
+    _output = output;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), axis, num_tensors, output->info());
+
+    // Add build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_underlying_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DAXIS=" + support::cpp11::to_string(axis));
+    build_opts.add_option("-DSRC_DIM2=" + support::cpp11::to_string(input->info()->dimension(2)));
+    build_opts.add_option("-DDST_DIM3=" + support::cpp11::to_string(output->info()->dimension(3)));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("stack_layer", build_opts.options()));
+
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
+    const unsigned int idx = 2 * num_arguments_per_4D_tensor();
+    _kernel.setArg<cl_uint>(idx, idx_input);
+}
+
+Status CLStackLayerKernel::validate(const ITensorInfo *input, unsigned int axis, unsigned int idx_input, unsigned int num_tensors, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, axis, idx_input, num_tensors, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), axis, num_tensors, output->clone().get()).first);
+    return Status{};
+}
+
+void CLStackLayerKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window window_out;
+    window_out.use_tensor_dimensions(_output->info()->tensor_shape());
+
+    Window collapsed = window.collapse(ICLKernel::window(), Window::DimZ);
+
+    Window slice_in  = collapsed.first_slice_window_4D();
+    Window slice_out = window_out.first_slice_window_4D();
+
+    unsigned int idx = 0;
+    add_4D_tensor_argument(idx, _input, slice_in);
+    add_4D_tensor_argument(idx, _output, slice_out);
+    enqueue(queue, *this, slice_in);
+}
diff --git a/src/runtime/CL/functions/CLStackLayer.cpp b/src/runtime/CL/functions/CLStackLayer.cpp
new file mode 100644
index 0000000000..85adcad90c
--- /dev/null
+++ b/src/runtime/CL/functions/CLStackLayer.cpp
@@ -0,0 +1,85 @@
+/*
+ * Copyright (c) 2018 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 <complex>
+
+#include "arm_compute/runtime/CL/functions/CLStackLayer.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+CLStackLayer::CLStackLayer() // NOLINT
+    : _input(),
+      _stack_kernels(),
+      _num_inputs(0)
+{
+}
+
+void CLStackLayer::configure(const std::vector<ICLTensor *> &input, int axis, ICLTensor *output)
+{
+    _num_inputs    = input.size();
+    _stack_kernels = arm_compute::support::cpp14::make_unique<CLStackLayerKernel[]>(_num_inputs);
+
+    // Wrap around negative values
+    const unsigned int axis_u = wrap_around(axis, static_cast<int>(input[0]->info()->num_dimensions() + 1));
+
+    for(unsigned int i = 0; i < _num_inputs; i++)
+    {
+        _stack_kernels[i].configure(input[i], axis_u, i, _num_inputs, output);
+    }
+}
+
+Status CLStackLayer::validate(const std::vector<ITensorInfo *> &input, int axis, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input.size() < 2);
+
+    // Wrap around negative values
+    const unsigned int axis_u = wrap_around(axis, static_cast<int>(input[0]->num_dimensions() + 1));
+
+    const unsigned int num_inputs = input.size();
+
+    for(unsigned int i = 0; i < num_inputs; i++)
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLStackLayerKernel::validate(input[i], axis_u, i, num_inputs, output));
+    }
+
+    return Status{};
+}
+
+void CLStackLayer::run()
+{
+    for(unsigned i = 0; i < _num_inputs; i++)
+    {
+        CLScheduler::get().enqueue(_stack_kernels[i], false);
+    }
+}