Remove padding from CLNormalizePlanarYUVLayerKernel

Resolve: COMPMID-3911

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: Id5615b6a8b52030fb611a1a04bcd4664b8232e90
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5451
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 32 insertions, 20 deletions

diff --git a/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp b/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp
index e78d906a9d..cf2511adec 100644
--- a/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp
+++ b/src/core/CL/kernels/CLNormalizePlanarYUVLayerKernel.cpp
@@ -64,11 +64,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, ITensorInfo *mean, ITensorInfo *std)
+std::pair<Status, Window> validate_and_configure_window_nchw(ITensorInfo *input, ITensorInfo *output)
 {
-    // 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));
@@ -78,13 +75,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 
     bool window_changed = update_window_and_padding(win, input_access, output_access);
 
-    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);
 }
@@ -106,19 +96,30 @@ void CLNormalizePlanarYUVLayerKernel::configure(const CLCompileContext &compile_
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, mean, std);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), mean->info(), std->info()));
 
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), *input->info()->clone());
+
+    auto padding_info = get_padding_info({ input, output });
+
     _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);
-    const DataType     dt                                = input->info()->data_type();
+    const DataLayout data_layout = input->info()->data_layout();
+
+    // Get number of elements to process per iterations
+    const unsigned int num_elems_processed_per_iteration = (data_layout == DataLayout::NHWC) ? adjust_vec_size(16 / input->info()->element_size(),
+                                                                                                               input->info()->dimension(0)) :
+                                                           (16 / input->info()->element_size());
+    const unsigned int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+    const DataType     dt          = input->info()->data_type();
 
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(dt)));
     build_opts.add_option(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
+    build_opts.add_option(("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->info()->dimension(0) % num_elems_processed_per_iteration)));
     build_opts.add_option(("-DNUM_CHANNELS=" + support::cpp11::to_string(input->info()->dimension(channel_idx))));
 
     std::string kernel_name = "normalize_planar_yuv_layer_";
@@ -131,13 +132,22 @@ void CLNormalizePlanarYUVLayerKernel::configure(const CLCompileContext &compile_
     }
 
     // Create kernel
-    kernel_name += lower_string(string_from_data_layout(input->info()->data_layout()));
+    kernel_name += lower_string(string_from_data_layout(data_layout));
     _kernel = create_kernel(compile_context, kernel_name, 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);
+    if(data_layout == DataLayout::NHWC)
+    {
+        Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+        ICLKernel::configure_internal(win);
+        ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
+    }
+    else
+    {
+        auto win_config = validate_and_configure_window_nchw(input->info(), 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 = "normalize_planar_yuv_layer_";
@@ -155,8 +165,10 @@ void CLNormalizePlanarYUVLayerKernel::configure(const CLCompileContext &compile_
 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);
-
+    if(input->data_layout() == DataLayout::NCHW)
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_nchw(input->clone().get(), output->clone().get()).first);
+    }
     return Status{};
 }