COMPMID-1932: Implement CLHeightConcatenateLayer.

Change-Id: I244ff2352c3c9620148dab6475aee62b218b96ef
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/871
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

3 files changed, 214 insertions, 1 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 5af8a09723..be457a74db 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -190,6 +190,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "compare_lessequal_quantized", "comparisons.cl" },
     { "concatenate_depth", "concatenate.cl" },
     { "concatenate_width", "concatenate.cl" },
+    { "concatenate_height", "concatenate.cl" },
     { "concatenate_width_x2", "concatenate.cl" },
     { "concatenate_width_x4", "concatenate.cl" },
     { "convolution_rectangle", "convolution_rectangle.cl" },
diff --git a/src/core/CL/cl_kernels/concatenate.cl b/src/core/CL/cl_kernels/concatenate.cl
index c374769423..99f4659960 100644
--- a/src/core/CL/cl_kernels/concatenate.cl
+++ b/src/core/CL/cl_kernels/concatenate.cl
@@ -330,6 +330,59 @@ __kernel void concatenate_width(
 
 #endif /* defined(WIDTH_OFFSET) && defined(DEPTH) */
 
+#if defined(HEIGHT_OFFSET) && defined(DEPTH) && defined(VEC_SIZE)
+/** This kernel concatenates the input tensor into the output tensor along the second dimension
+ *
+ * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
+ * @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16
+ * @note Vector sizes supported are 2,4,8 and 16.
+ * @note The offset for the second spatial dimension has to be passed at compile time using -DHEIGHT_OFFSET. i.e. -DHEIGHT_OFFSET=128
+ * @note Tensor depth should be given as a preprocessor argument using -DDEPTH=size. e.g. -DDEPTH=16
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/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 first source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_w                        src_stride_z * number of elements along Z 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 destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_w                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+
+__kernel void concatenate_height(
+    TENSOR4D_DECLARATION(src),
+    TENSOR4D_DECLARATION(dst))
+{
+    Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, DEPTH);
+    Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT(dst, DEPTH);
+
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
+
+#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
+    const VEC_UCHAR out = requantize(source_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT);
+    VSTORE(VEC_SIZE)
+    (out, 0, (__global DATA_TYPE *)(dst.ptr + HEIGHT_OFFSET * dst_stride_y));
+#else  /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
+    VSTORE(VEC_SIZE)
+    (source_values, 0, (__global DATA_TYPE *)(dst.ptr + HEIGHT_OFFSET * dst_stride_y));
+#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
+}
+
+#endif /* defined(HEIGHT_OFFSET) && defined(DEPTH) */
+
 /** This kernel concatenates the input tensor into the output tensor along the third dimension
  *
  * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float
@@ -370,6 +423,5 @@ __kernel void concatenate_depth(
 
     VSTORE(VEC_SIZE)
     (source_values, 0, (__global DATA_TYPE *)(dst.ptr + offsets.z));
-
 }
 #endif /* defined(DATA_TYPE) && defined(VEC_SIZE) */
diff --git a/src/core/CL/kernels/CLHeightConcatenateLayerKernel.cpp b/src/core/CL/kernels/CLHeightConcatenateLayerKernel.cpp
new file mode 100644
index 0000000000..e3f2a96281
--- /dev/null
+++ b/src/core/CL/kernels/CLHeightConcatenateLayerKernel.cpp
@@ -0,0 +1,160 @@
+/*
+ * Copyright (c) 2019 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/CLHeightConcatenateLayerKernel.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"
+
+#include <map>
+
+using namespace arm_compute;
+namespace
+{
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int height_offset, ITensorInfo *output, unsigned int &num_elems_processed_per_iteration)
+{
+    num_elems_processed_per_iteration = 4;
+    // The window needs to be based on input as we copy all the heights 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, height_offset, num_elems_processed_per_iteration);
+    bool                   window_changed = update_window_and_padding(win, input_access, output_access);
+
+    Window win_collapsed = win.collapse(win, Window::DimZ);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win_collapsed);
+}
+Status validate_arguments(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::F16, DataType::U32,
+                                                         DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) + height_offset > output->dimension(Window::DimY));
+
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != output->dimension(0));
+    for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != output->dimension(i));
+    }
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+
+    return Status{};
+}
+} // namespace
+
+CLHeightConcatenateLayerKernel::CLHeightConcatenateLayerKernel()
+    : _input(nullptr), _output(nullptr), _height_offset(0), _num_elems_processed_per_iteration()
+{
+}
+
+Status CLHeightConcatenateLayerKernel::validate(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output)
+{
+    unsigned int num_elems_processed_per_iteration;
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, height_offset, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), height_offset, output->clone().get(), num_elems_processed_per_iteration).first);
+    return Status{};
+}
+
+void CLHeightConcatenateLayerKernel::configure(const ICLTensor *input, unsigned int height_offset, ICLTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), height_offset, output->info()));
+
+    _input         = input;
+    _output        = output;
+    _height_offset = height_offset;
+
+    auto win_config = validate_and_configure_window(input->info(), height_offset, output->info(), _num_elems_processed_per_iteration);
+
+    // Add build options
+    CLBuildOptions build_opts;
+
+    switch(input->info()->element_size())
+    {
+        case 1:
+        {
+            build_opts.add_option("-DDATA_TYPE=uchar");
+            break;
+        }
+        case 2:
+        {
+            build_opts.add_option("-DDATA_TYPE=short");
+            break;
+        }
+        case 4:
+        {
+            build_opts.add_option("-DDATA_TYPE=int");
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("Unsupported input data type.");
+            break;
+        }
+    }
+
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(_num_elems_processed_per_iteration));
+    build_opts.add_option("-DHEIGHT_OFFSET=" + support::cpp11::to_string(_height_offset));
+    build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
+
+    if(is_data_type_quantized_asymmetric(input->info()->data_type()) && input->info()->quantization_info() != output->info()->quantization_info())
+    {
+        build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(input->info()->quantization_info().offset));
+        build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(output->info()->quantization_info().offset));
+        build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(input->info()->quantization_info().scale));
+        build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(output->info()->quantization_info().scale));
+    }
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("concatenate_height", build_opts.options()));
+    // Configure kernel window
+
+    ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
+
+    ICLKernel::configure_internal(std::get<1>(win_config));
+}
+
+void CLHeightConcatenateLayerKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    unsigned int idx = 0;
+    add_4D_tensor_argument(idx, _input, window);
+    add_4D_tensor_argument(idx, _output, window);
+    enqueue(queue, *this, window);
+}