COMPMID-1330: Add support for NormalizePlanarYUV operator in CL

Change-Id: Id0754b9e2bc3ef7ff2c4c21c3b89709588c41bd3
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/146637
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>

3 files changed, 313 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 0cc6e320bf..4af2b09530 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -308,6 +308,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "non_max_suppression", "nonmax.cl" },
     { "normalization_layer_cross_map", "normalization_layer.cl" },
     { "normalization_layer_in_map", "normalization_layer.cl" },
+    { "normalize_planar_yuv_layer_nchw", "normalize_planar_yuv_layer.cl" },
+    { "normalize_planar_yuv_layer_nhwc", "normalize_planar_yuv_layer.cl" },
     { "NV12_to_IYUV_bt709", "color_convert.cl" },
     { "NV12_to_RGB888_bt709", "color_convert.cl" },
     { "NV12_to_RGBA8888_bt709", "color_convert.cl" },
@@ -674,6 +676,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/normalization_layer.clembed"
     },
     {
+        "normalize_planar_yuv_layer.cl",
+#include "./cl_kernels/normalize_planar_yuv_layer.clembed"
+    },
+    {
         "batchnormalization_layer.cl",
 #include "./cl_kernels/batchnormalization_layer.clembed"
     },
diff --git a/src/core/CL/cl_kernels/normalize_planar_yuv_layer.cl b/src/core/CL/cl_kernels/normalize_planar_yuv_layer.cl
new file mode 100644
index 0000000000..dc6652449e
--- /dev/null
+++ b/src/core/CL/cl_kernels/normalize_planar_yuv_layer.cl
@@ -0,0 +1,134 @@
+/*
+ * 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(VEC_SIZE)
+
+#define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+
+/** Apply normalize_planar_yuv layer on tensors with NCHW format.
+ *
+ * @note Data type should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE e.g. -DVEC_SIZE=8
+ * @note The depth of the input tensor should be given as a preprocessor argument using -DNUM_CHANNELS e.g. -DNUM_CHANNELS=8
+ *
+ * @param[in]  src_ptr                            Pointer to the first source tensor. Supported data types: F16/F32
+ * @param[in]  src_stride_x                       Stride of the first source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                         input_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                         input_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                         input_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[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_offset_first_element_in_bytes  The offset of the first element in the destination tensor
+ * @param[in]  mean_ptr                           Pointer to the mean source tensor. Supported data types: same as @p src_ptr
+ * @param[in]  mean_stride_x                      Stride of the mean source tensor in X dimension (in bytes)
+ * @param[in]  mean_step_x                        mean_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
+ * @param[in]  std_ptr                            Pointer to the std tensor. Supported data types: same as @p src_ptr
+ * @param[in]  std_stride_x                       Stride of the std tensor in X dimension (in bytes)
+ * @param[in]  std_step_x                         std_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  std_offset_first_element_in_bytes  The offset of the first element in the var source tensor
+ */
+__kernel void normalize_planar_yuv_layer_nchw(TENSOR3D_DECLARATION(src),
+                                              TENSOR3D_DECLARATION(dst),
+                                              VECTOR_DECLARATION(mean),
+                                              VECTOR_DECLARATION(std))
+{
+    Tensor3D src  = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst  = CONVERT_TO_TENSOR3D_STRUCT(dst);
+    Vector   mean = CONVERT_TO_VECTOR_STRUCT(mean);
+    Vector   std  = CONVERT_TO_VECTOR_STRUCT(std);
+
+    const uint current_slice = get_global_id(2) % NUM_CHANNELS;
+
+    const DATA_TYPE curr_mean = *((__global DATA_TYPE *)(mean.ptr + current_slice * mean.stride_x));
+    const DATA_TYPE curr_std  = *((__global DATA_TYPE *)(std.ptr + current_slice * std.stride_x));
+
+    TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
+    TYPE res  = (data - curr_mean) / curr_std;
+
+    VSTORE(VEC_SIZE)
+    (res, 0, (__global DATA_TYPE *)dst.ptr);
+}
+
+/** Apply normalize_planar_yuv layer on tensors with NHWC format.
+ *
+ * @note Data type should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE e.g. -DVEC_SIZE=8
+ *
+ * @param[in]  src_ptr                            Pointer to the first source tensor. Supported data types: F16/F32
+ * @param[in]  src_stride_x                       Stride of the first source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                         input_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                         input_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                         input_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[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_offset_first_element_in_bytes  The offset of the first element in the destination tensor
+ * @param[in]  mean_ptr                           Pointer to the mean source tensor. Supported data types: same as @p src_ptr
+ * @param[in]  mean_stride_x                      Stride of the mean source tensor in X dimension (in bytes)
+ * @param[in]  mean_step_x                        mean_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
+ * @param[in]  std_ptr                            Pointer to the std tensor. Supported data types: same as @p src_ptr
+ * @param[in]  std_stride_x                       Stride of the std tensor in X dimension (in bytes)
+ * @param[in]  std_step_x                         std_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  std_offset_first_element_in_bytes  The offset of the first element in the var source tensor
+ */
+__kernel void normalize_planar_yuv_layer_nhwc(TENSOR3D_DECLARATION(src),
+                                              TENSOR3D_DECLARATION(dst),
+                                              VECTOR_DECLARATION(mean),
+                                              VECTOR_DECLARATION(std))
+{
+    Tensor3D src  = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst  = CONVERT_TO_TENSOR3D_STRUCT(dst);
+    Vector   mean = CONVERT_TO_VECTOR_STRUCT(mean);
+    Vector   std  = CONVERT_TO_VECTOR_STRUCT(std);
+
+    const uint current_slice = get_global_id(0);
+
+    const TYPE curr_mean = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(mean.ptr + current_slice * VEC_SIZE * mean.stride_x));
+    const TYPE curr_std  = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(std.ptr + current_slice * VEC_SIZE * std.stride_x));
+
+    TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr);
+    TYPE res  = (data - curr_mean) / curr_std;
+
+    VSTORE(VEC_SIZE)
+    (res, 0, (__global DATA_TYPE *)dst.ptr);
+}
+#endif // defined(DATA_TYPE) && defined(VEC_SIZE)
diff --git a/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp b/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp
new file mode 100644
index 0000000000..31451ef422
--- /dev/null
+++ b/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp
@@ -0,0 +1,173 @@
+/*
+ * 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/CLNormalizePlanarYUVLayerKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.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/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Window.h"
+
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *std)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, mean, std);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, std);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(mean->num_dimensions() > 1, "mean and std must be vectors");
+
+    const unsigned int channel_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(channel_idx) != mean->dimension(0));
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *mean, ITensorInfo *std)
+{
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output, *input->clone());
+
+    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
+
+    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, input->valid_region());
+
+    if(input->data_layout() == DataLayout::NHWC)
+    {
+        AccessWindowHorizontal mean_access(mean, 0, num_elems_processed_per_iteration);
+        AccessWindowHorizontal std_access(std, 0, num_elems_processed_per_iteration);
+        window_changed = window_changed || update_window_and_padding(win, mean_access, std_access);
+    }
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+CLNormalizePlanarYUVLayerKernel::CLNormalizePlanarYUVLayerKernel()
+    : _input(nullptr), _output(nullptr), _mean(nullptr), _std(nullptr)
+{
+}
+
+void CLNormalizePlanarYUVLayerKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *mean, const ICLTensor *std)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, mean, std);
+
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), *input->info()->clone());
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), mean->info(), std->info()));
+
+    _input  = input;
+    _output = output;
+    _mean   = mean;
+    _std    = std;
+
+    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+    const unsigned int channel_idx                       = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::CHANNEL);
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
+    build_opts.add_option(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
+    build_opts.add_option(("-DNUM_CHANNELS=" + support::cpp11::to_string(input->info()->dimension(channel_idx))));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("normalize_planar_yuv_layer_" + lower_string(string_from_data_layout(input->info()->data_layout())), build_opts.options()));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info(), mean->info(), std->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
+    // Set config_id for enabling LWS tuning
+    _config_id = "normalize_planar_yuv_layer_";
+    _config_id += lower_string(string_from_data_layout(input->info()->data_layout()));
+    _config_id += "_";
+    _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 CLNormalizePlanarYUVLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *std)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, mean, std));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), mean->clone().get(), std->clone().get()).first);
+
+    return Status{};
+}
+
+void CLNormalizePlanarYUVLayerKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+    Window slice     = collapsed.first_slice_window_3D();
+
+    Window slice_in = collapsed.first_slice_window_1D();
+    slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
+
+    unsigned int idx = 2 * num_arguments_per_3D_tensor();
+    add_1D_tensor_argument(idx, _mean, slice_in);
+    add_1D_tensor_argument(idx, _std, slice_in);
+
+    do
+    {
+        idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice);
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(collapsed.slide_window_slice_3D(slice));
+}