COMPMID-1723: CL: Implement Reverse

Change-Id: Id0d4a07af24e2331161996083b0c1bab072bd405
Reviewed-on: https://review.mlplatform.org/322
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 285 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index e48ff03e05..1361d02f74 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -387,6 +387,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "reorg_layer_nhwc", "reorg_layer.cl" },
     { "reshape_layer", "reshape_layer.cl" },
     { "reshape_to_columns", "convolution_layer.cl" },
+    { "reverse", "reverse.cl" },
     { "RGB888_to_IYUV_bt709", "color_convert.cl" },
     { "RGB888_to_NV12_bt709", "color_convert.cl" },
     { "RGB888_to_RGBA8888_bt709", "color_convert.cl" },
@@ -822,6 +823,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/reshape_layer.clembed"
     },
     {
+        "reverse.cl",
+#include "./cl_kernels/reverse.clembed"
+    },
+    {
         "roi_align_layer.cl",
 #include "./cl_kernels/roi_align_layer.clembed"
     },
diff --git a/src/core/CL/cl_kernels/reverse.cl b/src/core/CL/cl_kernels/reverse.cl
new file mode 100644
index 0000000000..6afd382fec
--- /dev/null
+++ b/src/core/CL/cl_kernels/reverse.cl
@@ -0,0 +1,102 @@
+/*
+* 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(NUM_REVERSE_DIMS)
+
+#if NUM_REVERSE_DIMS > 4
+#error("Reversing more than 4 dimensions is not currently supported")
+#endif /* NUM_REVERSE_DIMS > 4 */
+
+/** Performs reverse along the specified axis.
+ *
+ * @note The data type must be given as a preprocessor argument using -DDATA_TYPE=num. e.g. -DDATA_TYPE=uint
+ * @note The number of dimensions to reverse must be given as a preprocessor argument using -DNUM_REVERSE_DIMS=num, e.g. -DNUM_REVERSE_DIMS=3
+ *
+ * @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 first 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 first 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 first 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 first source tensor
+ * @param[in]  axis_ptr                           Pointer to the source vector. Supported data types: U32
+ * @param[in]  axis_stride_x                      Stride of the first source tensor in X dimension (in bytes)
+ * @param[in]  axis_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  axis_offset_first_element_in_bytes The offset of the first element in the first 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                         output_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                         output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                         output_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 reverse(TENSOR4D_DECLARATION(src),
+                      VECTOR_DECLARATION(axis),
+                      TENSOR4D_DECLARATION(dst),
+                      const uint width,
+                      const uint height,
+                      const uint depth,
+                      const uint batches)
+{
+    Tensor4D src  = CONVERT_TO_TENSOR4D_STRUCT(src, depth);
+    Vector   axis = CONVERT_TO_VECTOR_STRUCT_NO_STEP(axis);
+    Tensor4D dst  = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(dst, depth);
+
+    const uint x_in = get_global_id(0);
+    const uint y_in = get_global_id(1);
+    const uint z_in = get_global_id(2) % depth;
+    const uint w_in = get_global_id(2) / depth;
+
+    const uint4 dims       = (uint4)(0, 1, 2, 3);
+    int4        to_reverse = (int4)(0, 0, 0, 0);
+#if NUM_REVERSE_DIMS == 1
+    const uint index = *((__global uint *)axis.ptr);
+    to_reverse       = (uint4)index == dims;
+#elif NUM_REVERSE_DIMS == 2
+    const uint2 indices = vload2(0, (__global uint *)axis.ptr);
+    to_reverse          = ((uint4)indices.s0 == dims) || ((uint4)indices.s1 == dims);
+#elif NUM_REVERSE_DIMS == 3
+    const uint2 indices01 = vload2(0, (__global uint *)axis.ptr);
+    const uint index2     = *((__global uint *)axis.ptr + 2);
+    to_reverse            = ((uint4)indices01.s0 == dims) || ((uint4)indices01.s1 == dims) || ((uint4)index2 == dims);
+#else  /* NUM_REVERSE_DIMS == 3 */
+    const uint4 indices = vload4(0, (__global uint *)axis.ptr);
+    to_reverse          = ((uint4)indices.s0 == dims) || ((uint4)indices.s1 == dims) || ((uint4)indices.s2 == dims) || ((uint4)indices.s3 == dims);
+#endif /* NUM_REVERSE_DIMS == 1 */
+    const uint x_out = to_reverse.s0 ? width - x_in - 1 : x_in;
+    const uint y_out = to_reverse.s1 ? height - y_in - 1 : y_in;
+    const uint z_out = to_reverse.s2 ? depth - z_in - 1 : z_in;
+    const uint w_out = to_reverse.s3 ? batches - w_in - 1 : w_in;
+
+    *((__global DATA_TYPE *)tensor4D_offset(&dst, x_out, y_out, z_out, w_out)) = *((__global DATA_TYPE *)src.ptr);
+}
+#endif // defined(DATA_TYPE) && defined(NUM_REVERSE_DIMS)
diff --git a/src/core/CL/kernels/CLReverseKernel.cpp b/src/core/CL/kernels/CLReverseKernel.cpp
new file mode 100644
index 0000000000..2859a51ce1
--- /dev/null
+++ b/src/core/CL/kernels/CLReverseKernel.cpp
@@ -0,0 +1,135 @@
+/*
+ * 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/CLReverseKernel.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/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Window.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, axis);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16,
+                                                         DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(axis, 1, DataType::U32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis->num_dimensions() > 1, "Axis must be a 1D tensor");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis->dimension(0) > 4, "Only up to 4 dimensions can be reversed");
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+} // namespace
+
+CLReverseKernel::CLReverseKernel()
+    : _input(nullptr), _output(nullptr), _axis(nullptr)
+{
+}
+
+void CLReverseKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *axis)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, axis);
+
+    _input  = input;
+    _output = output;
+    _axis   = axis;
+
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), *input->info()->clone());
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis->info()));
+
+    // Set kernel build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DNUM_REVERSE_DIMS=" + support::cpp11::to_string(axis->info()->dimension(0)));
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("reverse", build_opts.options()));
+
+    // Set static kernel arguments
+    unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_1D_tensor();
+    add_argument<cl_uint>(idx, input->info()->dimension(0));
+    add_argument<cl_uint>(idx, input->info()->dimension(1));
+    add_argument<cl_uint>(idx, input->info()->dimension(2));
+    add_argument<cl_uint>(idx, input->info()->dimension(3));
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output->info(), Steps());
+    ICLKernel::configure_internal(win);
+
+    // Set config_id for enabling LWS tuning
+    _config_id += "reverse_";
+    _config_id += lower_string(string_from_data_type(input->info()->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(2));
+}
+
+Status CLReverseKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis));
+    return Status{};
+}
+
+void CLReverseKernel::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(ICLKernel::window(), Window::DimZ);
+    Window slice      = collapsed.first_slice_window_4D();
+    Window axis_slice = collapsed.first_slice_window_1D();
+
+    do
+    {
+        unsigned int idx = 0;
+        add_4D_tensor_argument(idx, _input, slice);
+        add_1D_tensor_argument(idx, _axis, axis_slice);
+        add_4D_tensor_argument(idx, _output, slice);
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(collapsed.slide_window_slice_4D(slice));
+}
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLReverse.cpp b/src/runtime/CL/functions/CLReverse.cpp
new file mode 100644
index 0000000000..0f86b9f326
--- /dev/null
+++ b/src/runtime/CL/functions/CLReverse.cpp
@@ -0,0 +1,43 @@
+/*
+ * 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/CLReverse.h"
+
+#include "arm_compute/core/CL/kernels/CLReverseKernel.h"
+#include "arm_compute/core/Types.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+void CLReverse::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *axis)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLReverseKernel>();
+    k->configure(input, output, axis);
+    _kernel = std::move(k);
+}
+
+Status CLReverse::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis)
+{
+    return CLReverseKernel::validate(input, output, axis);
+}
+} // namespace arm_compute