COMPMID-1648: CLNormalizationLayer IN_MAP_2D support for NHWC for FP32/FP16

Change-Id: I49f1d865f5e7562f1d80db849353a89ef77e6a9e

3 files changed, 110 insertions, 13 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 33e66705e3..3a002e808d 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -326,7 +326,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "non_linear_filter_disk5x5", "non_linear_filter5x5.cl" },
     { "non_max_suppression", "nonmax.cl" },
     { "normalization_layer_cross_map", "normalization_layer.cl" },
-    { "normalization_layer_in_map", "normalization_layer.cl" },
+    { "normalization_layer_in_map_nchw", "normalization_layer.cl" },
+    { "normalization_layer_in_map_nhwc", "normalization_layer.cl" },
     { "normalize_planar_yuv_layer_nchw", "normalize_planar_yuv_layer.cl" },
     { "normalize_planar_yuv_layer_nhwc", "normalize_planar_yuv_layer.cl" },
     { "normalize_planar_yuv_layer_q8_nchw", "normalize_planar_yuv_layer_quantized.cl" },
diff --git a/src/core/CL/cl_kernels/normalization_layer.cl b/src/core/CL/cl_kernels/normalization_layer.cl
index 0b6df39c9a..390f8fcbeb 100644
--- a/src/core/CL/cl_kernels/normalization_layer.cl
+++ b/src/core/CL/cl_kernels/normalization_layer.cl
@@ -32,6 +32,7 @@
 #define LOAD_OP(offset, ptr) vload4(offset, ptr)
 #define STORE_OP(data, offset, ptr) vstore4(data, offset, ptr)
 
+#if defined(NUM_SLICES)
 /** Apply cross-map normalization.
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
@@ -91,9 +92,10 @@ __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input),
 
     STORE_OP(normalized_pixel, 0, (__global DATA_TYPE *)out.ptr);
 }
+#endif /* defined(NUM_SLICES) */
 
 #if defined(WIDTH_SIZE)
-/** Apply in-map normalization.
+/** Apply in-map normalization when tensors are in the NCHW data layout format.
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16
@@ -117,8 +119,8 @@ __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input),
  * @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 normalization_layer_in_map(TENSOR3D_DECLARATION(input),
-                                         TENSOR3D_DECLARATION(output))
+__kernel void normalization_layer_in_map_nchw(TENSOR3D_DECLARATION(input),
+                                              TENSOR3D_DECLARATION(output))
 {
     Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(input);
     Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
@@ -170,3 +172,83 @@ __kernel void normalization_layer_in_map(TENSOR3D_DECLARATION(input),
     STORE_OP(normalized_pixel, 0, (__global DATA_TYPE *)out.ptr);
 }
 #endif // defined(WIDTH_SIZE)
+
+#if defined(NUM_SLICES)
+/** Apply in-map normalization when tensors are in the NHWC data layout format.
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16
+ * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5
+ * @note The number of slices should be given as a preprocessor argument using -DNUM_SLICES=size. e.g. -DNUM_SLICES=192
+ * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA
+ *
+ * @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_stride_z                       Stride of the first 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[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 first 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 first 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 normalization_layer_in_map_nhwc(TENSOR3D_DECLARATION(input),
+                                              TENSOR3D_DECLARATION(output))
+{
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
+
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    acc = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))0;
+    const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    coeff_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(COEFF);
+    const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    beta_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(BETA);
+    const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    kappa_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(KAPPA);
+
+    const int current_cols = get_global_id(1);
+    const int first_col    = max(-(int)RADIUS, -current_cols);
+    const int last_col     = min((int)RADIUS, (int)get_global_size(1) - 1 - current_cols);
+
+#if defined(IN_MAP_2D)
+    const int current_rows = get_global_id(2);
+    const int first_row    = max(-(int)RADIUS, -current_rows);
+    const int last_row     = min((int)RADIUS, (int)NUM_SLICES - 1 - current_rows);
+#endif /* defined(IN_MAP_2D) */
+
+#if defined(IN_MAP_2D)
+    for(int j = first_row; j <= last_row; ++j)
+    {
+#endif /* defined(IN_MAP_2D) */
+        for(int i = first_col; i <= last_col; ++i)
+        {
+#if defined(IN_MAP_2D)
+            VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+            values = LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&in, 0, i, j));
+#else  /* defined(IN_MAP_2D) */
+            VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+            values = LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&in, 0, i, 0));
+#endif /* defined(IN_MAP_2D) */
+            acc = ADD_OP(acc, MUL_OP(values, values));
+        }
+#if defined(IN_MAP_2D)
+    }
+#endif /* defined(IN_MAP_2D) */
+
+    acc = ADD_OP(MUL_OP(acc, coeff_v), kappa_v);
+    const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    normalized = POW_OP(acc, beta_v);
+    const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    normalized_pixel = DIV_OP(LOAD_OP(0, (__global DATA_TYPE *)in.ptr), normalized);
+
+    STORE_OP(normalized_pixel, 0, (__global DATA_TYPE *)out.ptr);
+}
+#endif /* defined(NUM_SLICES) */
diff --git a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
index 67357da7d1..9623ec6a89 100644
--- a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
@@ -37,20 +37,21 @@ using namespace arm_compute;
 
 namespace
 {
+constexpr unsigned int num_elems_processed_per_iteration = 4;
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, NormalizationLayerInfo norm_info)
 {
     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_DATA_LAYOUT_NOT_IN(input, DataLayout::NCHW, DataLayout::NHWC);
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
 
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_layout() == DataLayout::NHWC && norm_info.type() == NormType::IN_MAP_2D,
-                                    "Only Cross-map and 1D In-map normalization is supported for NHWC layout");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd");
 
     // 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_DATA_LAYOUT(input, output);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
     }
 
@@ -62,8 +63,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     // Output tensor auto initialization if not yet initialized
     auto_init_if_empty(*output, *input->clone());
 
-    const unsigned int num_elems_processed_per_iteration = 4;
-
     const unsigned int norm_idx              = get_normalization_dimension_index(input->data_layout(), norm_info);
     const bool         is_norm_accross_width = norm_idx == 0;
 
@@ -118,15 +117,14 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, ICLTensor *ou
     _input  = input;
     _output = output;
 
-    const unsigned int num_elems_processed_per_iteration = 4;
-    const bool         is_in_map_2D                      = (norm_info.type() == NormType::IN_MAP_2D);
-
     const DataLayout   data_layout  = input->info()->data_layout();
     const unsigned int norm_idx     = get_normalization_dimension_index(data_layout, norm_info);
     _is_norm_across_width           = norm_idx == 0;
     const unsigned int border_width = _is_norm_across_width ? num_elems_processed_per_iteration - 1 : 0;
     _border_size                    = BorderSize(0, border_width);
 
+    const bool is_in_map_2D = (norm_info.type() == NormType::IN_MAP_2D);
+
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
@@ -140,8 +138,24 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, ICLTensor *ou
     build_opts.add_option_if(norm_info.is_in_map() || (data_layout == DataLayout::NHWC && norm_info.is_cross_map()), "-DWIDTH_SIZE=" + support::cpp11::to_string(input->info()->dimension(0)));
 
     // Create kernel
-    std::string kernel_name = _is_norm_across_width ? "normalization_layer_in_map" : "normalization_layer_cross_map";
-    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+    std::string kernel_name;
+    if(norm_info.is_in_map())
+    {
+        kernel_name = "normalization_layer_in_map_" + lower_string(string_from_data_layout(data_layout));
+    }
+    else
+    {
+        if(data_layout == DataLayout::NCHW)
+        {
+            kernel_name = "normalization_layer_cross_map";
+        }
+        else
+        {
+            // 1D Cross-Map normalization in NHWC is the same as 1D In-Map normalization in NCHW
+            kernel_name = "normalization_layer_in_map_nchw";
+        }
+    }
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Configure kernel window
     auto win_config = validate_and_configure_window(input->info(), output->info(), norm_info);