COMPMID-413: Add support for QS8 and QS16 CLNormalizationLayer.

Change-Id: I1aaa9fb8d05796bbca9cfae584e084646552bb71
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/80155
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

1 files changed, 25 insertions, 12 deletions

diff --git a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
index 1afd76a375..a0607c2ba0 100644
--- a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
@@ -26,6 +26,7 @@
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/FixedPoint.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Utils.h"
@@ -46,7 +47,7 @@ BorderSize CLNormalizationLayerKernel::border_size() const
 
 void CLNormalizationLayerKernel::configure(const ICLTensor *input, const ICLTensor *squared_input, ICLTensor *output, NormalizationLayerInfo norm_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     // Output tensor auto initialization if not yet initialized
@@ -56,34 +57,46 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, const ICLTens
     ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, squared_input, output);
     ARM_COMPUTE_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd");
     ARM_COMPUTE_ERROR_ON_MSG(norm_info.type() == NormType::IN_MAP_2D, "2D In-Map Normalization not implemented");
-
-    // Set build options
-    std::set<std::string> build_opts;
-    build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
+    if(is_data_type_fixed_point(input->info()->data_type()))
+    {
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, squared_input, output);
+        ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.beta(), input);
+        ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.kappa(), input);
+        ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.scale_coeff(), input);
+    }
 
     _input         = input;
     _squared_input = squared_input;
     _output        = output;
 
-    _is_in_map                      = (norm_info.type() == NormType::IN_MAP_1D);
+    _is_in_map                      = (norm_info.type() != NormType::CROSS_MAP);
     const unsigned int border_width = _is_in_map ? std::min(norm_info.norm_size() / 2, 3U) : 0;
     _border_size                    = BorderSize(0, border_width);
 
+    const unsigned int num_elems_processed_per_iteration = (is_data_type_fixed_point(input->info()->data_type())) ? 16 : 4;
+    const unsigned int num_elems_read_per_iteration      = num_elems_processed_per_iteration + 2 * (norm_info.norm_size() / 2);
+
+    // Set build options
+    std::set<std::string> build_opts;
+    build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
+    if(is_data_type_fixed_point(input->info()->data_type()))
+    {
+        build_opts.emplace(("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())));
+    }
+    build_opts.emplace(("-DCOEFF=" + support::cpp11::to_string(norm_info.scale_coeff())));
+    build_opts.emplace(("-DBETA=" + support::cpp11::to_string(norm_info.beta())));
+    build_opts.emplace(("-DKAPPA=" + support::cpp11::to_string(norm_info.kappa())));
+    build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
+
     // Create kernel
     std::string kernel_name = (norm_info.type() == NormType::IN_MAP_1D) ? "normalization_layer_in_map_1D" : "normalization_layer_cross_map";
     _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
 
     // Set kernel static arguments
     unsigned int idx = 3 * num_arguments_per_3D_tensor(); // Skip the input and output parameters
-    _kernel.setArg<cl_float>(idx++, norm_info.scale_coeff());
-    _kernel.setArg<cl_float>(idx++, norm_info.beta());
-    _kernel.setArg<cl_float>(idx++, norm_info.kappa());
     _kernel.setArg<cl_uint>(idx++, norm_info.norm_size() / 2);
 
     // Configure kernel window
-    const unsigned int num_elems_processed_per_iteration = (_is_in_map) ? 4 : 1;
-    const unsigned int num_elems_read_per_iteration      = num_elems_processed_per_iteration + 2 * (norm_info.norm_size() / 2);
-
     Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
 
     AccessWindowHorizontal input_access(input->info(), -_border_size.left, num_elems_read_per_iteration);