COMPMID-1218 Implementing Batch to Space on OpenCL

Change-Id: I12ba4c0c35f086ea3f395970b85af5bf8f94850b
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/145052
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

4 files changed, 372 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 6f45756c12..29fd672a96 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -153,6 +153,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "arithmetic_add", "arithmetic_op.cl" },
     { "arithmetic_sub", "arithmetic_op.cl" },
     { "arithmetic_div", "arithmetic_op.cl" },
+    { "batch_to_space", "batch_to_space.cl" },
+    { "batch_to_space_static", "batch_to_space.cl" },
     { "batchnormalization_layer_nchw", "batchnormalization_layer.cl" },
     { "batchnormalization_layer_nhwc", "batchnormalization_layer.cl" },
     { "bitwise_or", "bitwise_op.cl" },
@@ -456,6 +458,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/arithmetic_op_quantized.clembed"
     },
     {
+        "batch_to_space.cl",
+#include "./cl_kernels/batch_to_space.clembed"
+    },
+    {
         "bitwise_op.cl",
 #include "./cl_kernels/bitwise_op.clembed"
     },
diff --git a/src/core/CL/cl_kernels/batch_to_space.cl b/src/core/CL/cl_kernels/batch_to_space.cl
new file mode 100644
index 0000000000..3043c2cf17
--- /dev/null
+++ b/src/core/CL/cl_kernels/batch_to_space.cl
@@ -0,0 +1,131 @@
+/*
+ * 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 withoutput restriction, including withoutput 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(BATCH_SIZE)
+/** Batch to space transformation.
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2
+ *
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  input_stride_x                       Stride of the 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 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 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[in]  batch_id                             The input tensor batch id
+ * @param[in]  block_shape_ptr                      Pointer to the source tensor. Supported data types: S32
+ * @param[in]  block_shape_stride_x                 Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  block_shape_step_x                   block_shape_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  block_shape_stride_y                 Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  block_shape_step_y                   block_shape_stride_y * number of elements along Y 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 source 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
+ */
+__kernel void batch_to_space(
+    TENSOR3D_DECLARATION(input),
+    const int batch_id,
+    VECTOR_DECLARATION(block_shape),
+    TENSOR4D_DECLARATION(output))
+{
+    Tensor3D in    = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor4D out   = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+    Vector   block = CONVERT_TO_VECTOR_STRUCT_NO_STEP(block_shape);
+
+    const int block_x = *((__global int *)vector_offset(&block, 0));
+    const int block_y = *((__global int *)vector_offset(&block, 1));
+
+    const int r = (BATCH_SIZE / (block_x * block_y));
+    const int x = get_global_id(0);
+    const int y = get_global_id(1);
+    const int z = get_global_id(2);
+    const int w = batch_id % r;
+
+    const int out_x = x * block_x + (batch_id / r) % block_x;
+    const int out_y = y * block_y + (batch_id / r) / block_x;
+
+    *((__global DATA_TYPE *)tensor4D_offset(&out, out_x, out_y, z, w)) = *((__global DATA_TYPE *)in.ptr);
+}
+#endif // defined(DATA_TYPE) && defined(BATCH_SIZE)
+
+#if defined(DATA_TYPE) && defined(BATCH_SIZE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y)
+/** Batch to space transformation.
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2
+ * @note The block shape x must be passed at compile time using -DBLOCK_SHAPE_X. e.g. -DBLOCK_SHAPE_X=2
+ * @note The block shape y must be passed at compile time using -DBLOCK_SHAPE_Y. e.g. -DBLOCK_SHAPE_Y=2
+ *
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  input_stride_x                       Stride of the 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 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 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[in]  batch_id                             The input tensor batch id
+ * @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 source 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
+ */
+__kernel void batch_to_space_static(
+    TENSOR3D_DECLARATION(input),
+    const int batch_id,
+    TENSOR4D_DECLARATION(output))
+{
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+
+    const int block_x = BLOCK_SHAPE_X;
+    const int block_y = BLOCK_SHAPE_Y;
+
+    const int r = (BATCH_SIZE / (block_x * block_y));
+    const int x = get_global_id(0);
+    const int y = get_global_id(1);
+    const int z = get_global_id(2);
+    const int w = batch_id % r;
+
+    const int out_x = x * block_x + (batch_id / r) % block_x;
+    const int out_y = y * block_y + (batch_id / r) / block_x;
+
+    *((__global DATA_TYPE *)tensor4D_offset(&out, out_x, out_y, z, w)) = *((__global DATA_TYPE *)in.ptr);
+}
+#endif // defined(DATA_TYPE) && defined(BATCH_SIZE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp b/src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp
new file mode 100644
index 0000000000..e08d6f6ec5
--- /dev/null
+++ b/src/core/CL/kernels/CLBatchToSpaceLayerKernel.cpp
@@ -0,0 +1,172 @@
+/*
+ * 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/CLBatchToSpaceLayerKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+using namespace arm_compute::misc::shape_calculator;
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *block_info, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, block_info, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(block_info, 1, DataType::S32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+
+    // Validate output if initialized
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 4);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+Status validate_arguments_static(const ITensorInfo *input, const int block_shape_x, const int block_shape_y, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+    ARM_COMPUTE_RETURN_ERROR_ON(block_shape_x <= 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(block_shape_y <= 0);
+
+    // Validate output if initialized
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[0] != (block_shape_x * output->tensor_shape()[0]));
+        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[1] != (block_shape_x * output->tensor_shape()[1]));
+        ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[2] != output->tensor_shape()[2]);
+        ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[3] % (block_shape_x * block_shape_y) != 0);
+        ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 4);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+} // namespace
+
+CLBatchToSpaceLayerKernel::CLBatchToSpaceLayerKernel()
+    : _input(nullptr), _block_shape(nullptr), _output(nullptr)
+{
+}
+
+void CLBatchToSpaceLayerKernel::configure(const ICLTensor *input, const ICLTensor *block_shape, ICLTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), block_shape->info(), output->info()));
+
+    _input       = input;
+    _block_shape = block_shape;
+    _output      = output;
+
+    // Create kernel
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DBATCH_SIZE=" + support::cpp11::to_string(input->info()->dimension(3)));
+    build_opts.add_option("-DWIDTH_IN=" + support::cpp11::to_string(input->info()->dimension(0)));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("batch_to_space", build_opts.options()));
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+    ICLKernel::configure_internal(win);
+}
+
+void CLBatchToSpaceLayerKernel::configure(const ICLTensor *input, const int32_t block_shape_x, const int32_t block_shape_y, ICLTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    TensorShape output_shape = compute_batch_to_space_shape(input->info(), block_shape_x, block_shape_y);
+    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type());
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_static(input->info(), block_shape_x, block_shape_y, output->info()));
+
+    _input  = input;
+    _output = output;
+
+    // Create kernel
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DBATCH_SIZE=" + support::cpp11::to_string(input->info()->dimension(3)));
+    build_opts.add_option("-DBLOCK_SHAPE_X=" + support::cpp11::to_string(block_shape_x));
+    build_opts.add_option("-DBLOCK_SHAPE_Y=" + support::cpp11::to_string(block_shape_y));
+    build_opts.add_option("-DWIDTH_IN=" + support::cpp11::to_string(input->info()->dimension(0)));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("batch_to_space_static", build_opts.options()));
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+    ICLKernel::configure_internal(win);
+}
+
+Status CLBatchToSpaceLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *block_shape, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, block_shape, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, block_shape, output));
+    return Status{};
+}
+
+Status CLBatchToSpaceLayerKernel::validate(const ITensorInfo *input, const int32_t block_shape_x, const int32_t block_shape_y, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_static(input, block_shape_x, block_shape_y, output));
+    return Status{};
+}
+
+void CLBatchToSpaceLayerKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window slice_in  = window.first_slice_window_3D();
+    Window slice_out = window.first_slice_window_4D();
+
+    Window vector_slice = window.first_slice_window_1D();
+    vector_slice.set(Window::DimX, Window::Dimension(0, 0, 0));
+
+    slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+    slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    slice_out.set(3, Window::Dimension(0, 0, 0));
+
+    int batch_id = 0;
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice_in);
+        add_argument(idx, batch_id);
+        if(_block_shape != nullptr)
+        {
+            add_1D_tensor_argument(idx, _block_shape, vector_slice);
+        }
+        add_4D_tensor_argument(idx, _output, slice_out);
+        enqueue(queue, *this, slice_in);
+
+        ++batch_id;
+    }
+    while(window.slide_window_slice_3D(slice_in));
+}
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLBatchToSpaceLayer.cpp b/src/runtime/CL/functions/CLBatchToSpaceLayer.cpp
new file mode 100644
index 0000000000..7919b131c8
--- /dev/null
+++ b/src/runtime/CL/functions/CLBatchToSpaceLayer.cpp
@@ -0,0 +1,63 @@
+/*
+ * 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/runtime/CL/functions/CLBatchToSpaceLayer.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+
+using namespace arm_compute;
+
+CLBatchToSpaceLayer::CLBatchToSpaceLayer()
+    : _batch_to_space_kernel()
+{
+}
+
+void CLBatchToSpaceLayer::configure(const ICLTensor *input, const ICLTensor *block_shape, ICLTensor *output)
+{
+    _batch_to_space_kernel.configure(input, block_shape, output);
+}
+
+void CLBatchToSpaceLayer::configure(const ICLTensor *input, int32_t block_shape_x, int32_t block_shape_y, ICLTensor *output)
+{
+    _batch_to_space_kernel.configure(input, block_shape_x, block_shape_y, output);
+}
+
+Status CLBatchToSpaceLayer::validate(const ITensorInfo *input, const ITensorInfo *block_shape, const ITensorInfo *output)
+{
+    return CLBatchToSpaceLayerKernel::validate(input, block_shape, output);
+}
+
+Status CLBatchToSpaceLayer::validate(const ITensorInfo *input, int32_t block_shape_x, int32_t block_shape_y, const ITensorInfo *output)
+{
+    return CLBatchToSpaceLayerKernel::validate(input, block_shape_x, block_shape_y, output);
+}
+
+void CLBatchToSpaceLayer::run()
+{
+    CLScheduler::get().enqueue(_batch_to_space_kernel, true);
+}