COMPMID-667: Add validation static method to NEON GEMMlowp

Change-Id: I8a470cc1351593ad8eeaf4ec92e04865e83d4f3c
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/96147
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 67 insertions, 25 deletions

diff --git a/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp b/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp
index 7f83144e12..7d79c664d1 100644
--- a/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMTranspose1xWKernel.cpp
@@ -41,45 +41,87 @@
 
 using namespace arm_compute;
 
-void NEGEMMTranspose1xWKernel::configure(const ITensor *input, ITensor *output)
+namespace
+{
+TensorShape get_output_shape(const ITensorInfo *input)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QS8, DataType::QS16, DataType::U8, DataType::S8, DataType::U16, DataType::S16, DataType::U32, DataType::S32,
-                                                  DataType::F16,
-                                                  DataType::F32);
-    ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+    TensorShape  output_shape{ input->tensor_shape() };
+    const size_t transpose_w = 16 / input->element_size();
+    output_shape.set(0, input->dimension(1) * transpose_w);
+    output_shape.set(1, static_cast<size_t>(std::ceil((input->dimension(0) / static_cast<float>(transpose_w)))));
+    return output_shape;
+}
 
-    TensorShape  output_shape{ input->info()->tensor_shape() };
-    const size_t transpose_w = 16 / input->info()->element_size();
-    output_shape.set(0, input->info()->dimension(1) * transpose_w);
-    output_shape.set(1, static_cast<size_t>(std::ceil((input->info()->dimension(0) / static_cast<float>(transpose_w)))));
+Error validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::U8, DataType::S8,
+                                                         DataType::QS16, DataType::U16, DataType::S16, DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
 
-    // Output tensor auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), get_output_shape(input));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    }
 
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    return Error{};
+}
 
-    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
     const int          scale_x                           = num_elems_processed_per_iteration;
-
-    _input  = input;
-    _output = output;
+    bool               window_changed                    = false;
 
     // Configure kernel window
-    Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
 
     ARM_COMPUTE_ERROR_ON_MSG((win.x().end() / scale_x) == 0, "Transposed shape would be 0 in the second dimension");
 
-    AccessWindowTranspose output_access(output->info(), 0, 0, num_elems_processed_per_iteration, 1, scale_x, 1.f / scale_x);
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+    window_changed = window_changed || update_window_and_padding(win, input_access);
 
-    update_window_and_padding(win,
-                              AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration),
-                              output_access);
+    // Configure window in case of configured output
+    if(output->total_size() != 0)
+    {
+        AccessWindowTranspose output_access(output, 0, 0, num_elems_processed_per_iteration, 1, scale_x, 1.f / scale_x);
+        window_changed = window_changed || update_window_and_padding(win, output_access);
+        output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), input->tensor_shape()));
+    }
 
-    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), input->info()->tensor_shape()));
+    Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+void NEGEMMTranspose1xWKernel::configure(const ITensor *input, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Output tensor auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), get_output_shape(input->info()), 1, input->info()->data_type(), input->info()->fixed_point_position());
+
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
+
+    _input  = input;
+    _output = output;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    INEKernel::configure(win_config.second);
+}
+
+Error NEGEMMTranspose1xWKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
 
-    INEKernel::configure(win);
+    return Error{};
 }
 
 void NEGEMMTranspose1xWKernel::run(const Window &window, const ThreadInfo &info)