COMPMID-2386: Add support for CLMeanStdNormalizationLayer

Change-Id: I0323b2410b222fd08933da22de455e798a60a0b1
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1297
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 280 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index b734fd291c..51acd9f339 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -379,6 +379,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "lktracker_stage1", "optical_flow_pyramid_lk.cl" },
     { "magnitude_phase", "magnitude_phase.cl" },
     { "mean_stddev_accumulate", "mean_stddev.cl" },
+    { "mean_stddev_normalization", "mean_stddev_normalization.cl" },
     { "memset", "memset.cl" },
     { "minmax", "minmaxloc.cl" },
     { "minmax_border", "minmaxloc.cl" },
@@ -818,6 +819,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/mean_stddev.clembed"
     },
     {
+        "mean_stddev_normalization.cl",
+#include "./cl_kernels/mean_stddev_normalization.clembed"
+    },
+    {
         "memset.cl",
 #include "./cl_kernels/memset.clembed"
     },
diff --git a/src/core/CL/cl_kernels/mean_stddev_normalization.cl b/src/core/CL/cl_kernels/mean_stddev_normalization.cl
new file mode 100644
index 0000000000..9667737c65
--- /dev/null
+++ b/src/core/CL/cl_kernels/mean_stddev_normalization.cl
@@ -0,0 +1,124 @@
+/*
+ * 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 "helpers.h"
+
+#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(EPSILON) && defined(WIDTH)
+/** This function normalizes the input 2D tensor across the first dimension with respect to mean and standard deviation of the same dimension.
+ *
+ * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @attention Data type should be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
+ * @attention Width of the input tensor should be passed using the -DWIDTH compile flag, e.g. -DWIDTH=16
+ * @attention Normalization epsilon parameter should be given as a preprocessor argument with -DEPSILON=value. e.g. -DEPSILON=0.001f
+ *
+ * @param[in]  input_ptr                            Pointer to the first source tensor. Supported data types: F16/F32
+ * @param[in]  input_stride_x                       Stride of the first 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 first 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_offset_first_element_in_bytes  The offset of the first element in the first source tensor
+ * @param[out] output_ptr                           (Optional) Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      (Optional) Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      (Optional) Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
+ */
+__kernel void mean_stddev_normalization(
+    IMAGE_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    IMAGE_DECLARATION(output)
+#endif /* IN_PLACE */
+)
+{
+    // Get pixels pointer
+    Image in = CONVERT_TO_IMAGE_STRUCT(input);
+#ifdef IN_PLACE
+    Image out = in;
+#else  /* IN_PLACE */
+    Image out = CONVERT_TO_IMAGE_STRUCT(output);
+#endif /* IN_PLACE */
+
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    sum = 0.f;
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    sum_sq = 0.f;
+    // Calculate partial sum
+    int i = 0;
+    for(; i <= (WIDTH - VEC_SIZE); i += VEC_SIZE)
+    {
+        // Load data
+        VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+        data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)offset(&in, i, 0));
+
+        sum += data;
+        sum_sq += data * data;
+    }
+    // Perform reduction
+#if VEC_SIZE > 8
+    sum.s01234567 += sum.s89abcdef;
+    sum_sq.s01234567 += sum_sq.s89abcdef;
+#endif // VEC_SIZE > 8
+#if VEC_SIZE > 4
+    sum.s0123 += sum.s4567;
+    sum_sq.s0123 += sum_sq.s4567;
+#endif // VEC_SIZE > 4
+#if VEC_SIZE > 2
+    sum.s01 += sum.s23;
+    sum_sq.s01 += sum_sq.s23;
+#endif // VEC_SIZE > 2
+    sum.s0 += sum.s1;
+    sum_sq.s0 += sum_sq.s1;
+    // Left-overs loop
+    for(; i < WIDTH; ++i)
+    {
+        DATA_TYPE data = *((__global DATA_TYPE *)offset(&in, i, 0));
+
+        sum.s0 += data;
+        sum_sq.s0 += data * data;
+    }
+
+    DATA_TYPE mean       = sum.s0 / WIDTH;
+    DATA_TYPE var        = (sum_sq.s0 / WIDTH) - (mean * mean);
+    DATA_TYPE stddev_inv = 1.f / sqrt(var + EPSILON);
+
+    i = 0;
+    for(; i <= (WIDTH - VEC_SIZE); i += VEC_SIZE)
+    {
+        VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+        data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)offset(&in, i, 0));
+
+        VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+        res = (data - mean) * stddev_inv;
+        VSTORE(VEC_SIZE)
+        (res, 0, (__global DATA_TYPE *)offset(&out, i, 0));
+    }
+    for(; i < WIDTH; ++i)
+    {
+        DATA_TYPE data = *((__global DATA_TYPE *)offset(&in, i, 0));
+
+        *((__global DATA_TYPE *)offset(&out, i, 0)) = (data - mean) * stddev_inv;
+    }
+}
+#endif /* defined(VEC_SIZE) && defined(DATA_TYPE) && defined(EPSILON) && defined(WIDTH) */
diff --git a/src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp b/src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp
new file mode 100644
index 0000000000..a9baf24fa6
--- /dev/null
+++ b/src/core/CL/kernels/CLMeanStdDevNormalizationKernel.cpp
@@ -0,0 +1,151 @@
+/*
+ * 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/CLMeanStdDevNormalizationKernel.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/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Window.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, float epsilon)
+{
+    ARM_COMPUTE_UNUSED(epsilon);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 2, "Input tensor cannot have more than 2 dimensions");
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+
+    // Checks performed when output is configured
+    if((output != nullptr) && (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{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    if(output != nullptr)
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+        // Output auto inizialitation if not yet initialized
+        auto_init_if_empty(*output, *input);
+    }
+
+    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
+
+    // This kernel doesn't need padding
+    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    if(output != nullptr)
+    {
+        output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+    }
+
+    return std::make_pair(Status{}, win);
+}
+} // namespace
+
+CLMeanStdDevNormalizationKernel::CLMeanStdDevNormalizationKernel()
+    : _input(nullptr), _output(nullptr), _run_in_place(false)
+{
+}
+
+void CLMeanStdDevNormalizationKernel::configure(ICLTensor *input, ICLTensor *output, float epsilon)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+
+    _run_in_place = (output == nullptr) || (output == input);
+
+    ARM_COMPUTE_ERROR_THROW_ON(CLMeanStdDevNormalizationKernel::validate(input->info(), (output != nullptr) ? output->info() : nullptr, epsilon));
+
+    _input  = input;
+    _output = output;
+
+    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+
+    // 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("-DEPSILON=" + float_to_string_with_full_precision(epsilon));
+    build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
+    build_opts.add_option_if(_run_in_place, "-DIN_PLACE");
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("mean_stddev_normalization", build_opts.options()));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), (_run_in_place) ? nullptr : output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
+    // Set config_id for enabling LWS tuning
+    _config_id = "mean_stddev_normalization_layer_";
+    _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));
+}
+
+Status CLMeanStdDevNormalizationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, float epsilon)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, epsilon));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (output != nullptr) ? output->clone().get() : nullptr).first);
+    return Status{};
+}
+
+void CLMeanStdDevNormalizationKernel::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 = window.first_slice_window_2D();
+    // Set slice step equal to width to force gws[0] to 1, as each thread normalizes across all rows
+    slice.set_dimension_step(Window::DimX, _input->info()->dimension(0));
+
+    do
+    {
+        unsigned int idx = 0;
+        add_2D_tensor_argument(idx, _input, slice);
+        if(!_run_in_place)
+        {
+            add_2D_tensor_argument(idx, _output, slice);
+        }
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(window.slide_window_slice_2D(slice));
+}
+} // namespace arm_compute