1 files changed, 13 insertions, 62 deletions

diff --git a/src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.cpp b/src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.cpp
index 5ef2cc46ee..1c8fef2db3 100644
--- a/src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.cpp
+++ b/src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.cpp
@@ -28,7 +28,6 @@
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Utils.h"
-#include "src/core/AccessWindowStatic.h"
 #include "src/core/CL/CLValidate.h"
 #include "src/core/helpers/WindowHelpers.h"
 #include "src/core/utils/helpers/tensor_info.h"
@@ -40,41 +39,6 @@ namespace arm_compute
 {
 namespace
 {
-constexpr unsigned int num_elems_processed_per_iteration = 8;
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *input3, ITensorInfo *input4, ITensorInfo *output)
-{
-    const unsigned int input1_width = input1->dimension(0);
-    const unsigned int input2_width = input2->dimension(0);
-    const unsigned int input3_width = input3->dimension(0);
-    const unsigned int input4_width = input4->dimension(0);
-
-    // The window needs to be based on the output
-    Window             win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
-    AccessWindowStatic input1_access(input1, 0, 0, ceil_to_multiple(input1_width, num_elems_processed_per_iteration), input1->dimension(1));
-
-    const unsigned int input2_left_padding  = input1_width % num_elems_processed_per_iteration;
-    const unsigned int input2_right_padding = ((input1_width + input2_width) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration - input1_width + num_elems_processed_per_iteration -
-                                              input2_width;
-    AccessWindowStatic input2_access(input2, -input2_left_padding, 0, input2_width + input2_right_padding, input2->dimension(1));
-
-    const unsigned int input3_left_padding  = (input1_width + input2_width) % num_elems_processed_per_iteration;
-    const unsigned int input3_right_padding = ((input1_width + input2_width + input3_width) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration - input1_width - input2_width +
-                                              num_elems_processed_per_iteration - input3_width;
-    AccessWindowStatic input3_access(input3, -input3_left_padding, 0, input3_width + input3_right_padding, input3->dimension(1));
-
-    const unsigned int input4_left_padding  = (input1_width + input2_width + input3_width) % num_elems_processed_per_iteration;
-    const unsigned int input4_right_padding = (output->dimension(0) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration + num_elems_processed_per_iteration - output->dimension(0);
-    AccessWindowStatic input4_access(input4, -input4_left_padding, 0, input4_width + input4_right_padding, input4->dimension(1));
-
-    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
-    bool                   window_changed = update_window_and_padding(win, input1_access, input2_access, input3_access, input4_access, output_access);
-
-    Window win_collapsed = win.collapse(win, Window::DimZ);
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win_collapsed);
-}
 Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *input3, const ITensorInfo *input4, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, input3, input4, output);
@@ -103,7 +67,6 @@ CLWidthConcatenate4TensorsKernel::CLWidthConcatenate4TensorsKernel()
 Status CLWidthConcatenate4TensorsKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *input3, const ITensorInfo *input4, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, input3, input4, output));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), input3->clone().get(), input4->clone().get(), output->clone().get()).first);
     return Status{};
 }
 
@@ -115,15 +78,25 @@ void CLWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile
     ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, input3, input4, output);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1, input2, input3, input4, output));
 
+    auto               padding_info                      = get_padding_info({ input1, input2, input3, input4, output });
+    const unsigned int min_dimension                     = std::min(std::min(input1->dimension(0), input2->dimension(0)), std::min(input3->dimension(0), input4->dimension(0)));
+    const unsigned int num_elems_processed_per_iteration = adjust_vec_size(8, min_dimension);
+    const unsigned int vec_size_leftover                 = output->dimension(0) % num_elems_processed_per_iteration;
+
     // Add build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input1->data_type()));
     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(vec_size_leftover));
     build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input1->dimension(2)));
     build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(input1->dimension(0)));
     build_opts.add_option("-DINPUT2_WIDTH=" + support::cpp11::to_string(input2->dimension(0)));
     build_opts.add_option("-DINPUT3_WIDTH=" + support::cpp11::to_string(input3->dimension(0)));
+    build_opts.add_option("-DINPUT4_WIDTH=" + support::cpp11::to_string(input4->dimension(0)));
     build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input1->element_size()));
+    build_opts.add_option("-DINPUT1_ROTATE_N=" + support::cpp11::to_string((input1->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
+    build_opts.add_option("-DINPUT2_ROTATE_N=" + support::cpp11::to_string((input1->dimension(0) + input2->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
+    build_opts.add_option("-DINPUT3_ROTATE_N=" + support::cpp11::to_string((input1->dimension(0) + input2->dimension(0) + input3->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration));
 
     // If input have different quantization info set quantization parameters needed for the re-quantization process
     const bool have_different_qinfo = helpers::tensor_info::tensors_have_different_quantization_info(output, input1, input2, input3, input4);
@@ -151,34 +124,12 @@ void CLWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile
     _kernel = create_kernel(compile_context, "concatenate_width_x4", build_opts.options());
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input1, input2, input3, input4, output);
-    ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
-
-    ICLKernel::configure_internal(std::get<1>(win_config));
+    Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+    ICLKernel::configure_internal(win.collapse(win, Window::DimZ));
 
     // Set output valid region
     output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
-
-    // Pass paddings as arguments to the kernel
-    const unsigned int input1_width = input1->dimension(0);
-    const unsigned int input2_width = input2->dimension(0);
-    const unsigned int input3_width = input3->dimension(0);
-
-    const unsigned int input1_right_padding = ceil_to_multiple(input1_width, num_elems_processed_per_iteration) - input1_width;
-    const unsigned int input2_left_padding  = input1_width % num_elems_processed_per_iteration;
-    const unsigned int input2_right_padding = ((input1_width + input2_width) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration - input1_width + num_elems_processed_per_iteration -
-                                              input2_width;
-    const unsigned int input3_left_padding  = (input1_width + input2_width) % num_elems_processed_per_iteration;
-    const unsigned int input3_right_padding = ((input1_width + input2_width + input3_width) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration - input1_width - input2_width +
-                                              num_elems_processed_per_iteration - input3_width;
-    const unsigned int input4_left_padding  = (input1_width + input2_width + input3_width) % num_elems_processed_per_iteration;
-    unsigned int       idx0                 = 5 * num_arguments_per_4D_tensor();
-    _kernel.setArg<cl_uint>(idx0++, input1_right_padding);
-    _kernel.setArg<cl_uint>(idx0++, input2_left_padding);
-    _kernel.setArg<cl_uint>(idx0++, input2_right_padding);
-    _kernel.setArg<cl_uint>(idx0++, input3_left_padding);
-    _kernel.setArg<cl_uint>(idx0++, input3_right_padding);
-    _kernel.setArg<cl_uint>(idx0++, input4_left_padding);
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 
     // Set config_id for enabling LWS tuning
     _config_id = "concatenate_width_x4_";