COMPMID-3712 Remove OpenCL padding: CLDepthwiseConvolutionLayer3x3NHWCKernel FP16/32

Removed unused N from partial block loading macro
Created utility to assert change in padding

Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Change-Id: Ifdd30c66dbf5f2842c6b2d939000613d5011708e
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4192
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 109 insertions, 115 deletions

diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index 5a0d2d0a62..876ef1ec5d 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -124,37 +124,33 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
                                                         const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
                                                         ITensorInfo *output_multipliers, ITensorInfo *output_shifts)
 {
-    const size_t weights_width  = 3;
-    const size_t weights_height = 3;
-
-    // Get convolved dimensions
-    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(
-                                         *input, TensorInfo(TensorShape(weights_width, weights_height), 1, weights->data_type()).set_data_layout(DataLayout::NCHW), conv_info, depth_multiplier, dilation);
+    ARM_COMPUTE_UNUSED(weights);
+    ARM_COMPUTE_UNUSED(depth_multiplier);
 
-    auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape).set_quantization_info(output->quantization_info()));
+    const bool   is_stride_1_dilation_1           = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1) && dilation.x() == 1 && dilation.y() == 1);
+    unsigned int num_rows_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
 
-    const bool is_qasymm              = is_data_type_quantized_asymmetric(input->data_type());
-    const bool is_stride_1_dilation_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1) && dilation.x() == 1 && dilation.y() == 1);
+    Window win{};
+    Status err{};
 
-    const unsigned int num_rows_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
-    const unsigned int num_elems_accessed_per_iteration = is_qasymm ? 4 : (8 / input->element_size());
-    const unsigned int num_rows_read_per_iteration      = num_rows_processed_per_iteration + 2;
-    const unsigned int num_rows_written_per_iteration   = std::ceil(num_rows_processed_per_iteration / static_cast<float>(conv_info.stride().first));
+    if(is_data_type_quantized_asymmetric(input->data_type()))
+    {
+        const unsigned int num_elems_accessed_per_iteration = 4;
+        const unsigned int num_rows_read_per_iteration      = num_rows_processed_per_iteration + 2;
+        const unsigned int num_rows_written_per_iteration   = std::ceil(num_rows_processed_per_iteration / static_cast<float>(conv_info.stride().first));
 
-    BorderSize border_size;
-    border_size = BorderSize(conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
+        BorderSize border_size;
+        border_size = BorderSize(conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
 
-    // Configure kernel window
-    Window win = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_written_per_iteration));
+        // Configure kernel window
+        win = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_written_per_iteration));
 
-    AccessWindowStatic input_access(input, 0, -border_size.top, ceil_to_multiple(input->dimension(0), num_elems_accessed_per_iteration),
-                                    ceil_to_multiple(input->dimension(1) + border_size.bottom, num_rows_read_per_iteration));
-    AccessWindowRectangle output_access(output, 0, 0, num_elems_accessed_per_iteration, num_rows_written_per_iteration);
+        AccessWindowStatic input_access(input, 0, -border_size.top, ceil_to_multiple(input->dimension(0), num_elems_accessed_per_iteration),
+                                        ceil_to_multiple(input->dimension(1) + border_size.bottom, num_rows_read_per_iteration));
+        AccessWindowRectangle output_access(output, 0, 0, num_elems_accessed_per_iteration, num_rows_written_per_iteration);
 
-    bool window_changed = false;
+        bool window_changed = false;
 
-    if(is_qasymm)
-    {
         if((output_multipliers != nullptr) && (output_shifts != nullptr))
         {
             AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_accessed_per_iteration);
@@ -166,27 +162,28 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
             Status err = ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "output_multipliers and output_shifts must be non-nullptr for quantized input");
             return std::make_pair(err, win);
         }
+
+        if(bias != nullptr)
+        {
+            AccessWindowHorizontal bias_access(bias, 0, num_elems_accessed_per_iteration);
+            window_changed = window_changed || update_window_and_padding(win, bias_access);
+        }
+        output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+
+        err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
     }
     else
     {
-        AccessWindowStatic weights_access(weights, 0, 0, ceil_to_multiple(weights->dimension(0), num_elems_accessed_per_iteration), weights->dimension(1));
-        window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
-    }
-
-    if(bias != nullptr)
-    {
-        AccessWindowHorizontal bias_access(bias, 0, num_elems_accessed_per_iteration);
-        window_changed = window_changed || update_window_and_padding(win, bias_access);
+        unsigned int num_elems_accessed_per_iteration = adjust_vec_size(4 / input->element_size(), input->dimension(0));
+        win                                           = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_processed_per_iteration));
     }
-    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
 
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
     return std::make_pair(err, win);
 }
 } // namespace
 
 CLDepthwiseConvolutionLayer3x3NHWCKernel::CLDepthwiseConvolutionLayer3x3NHWCKernel()
-    : _num_rows_processed_per_iteration(1), _num_planes_processed_per_iteration(1)
+    : _num_planes_processed_per_iteration(1)
 {
 }
 
@@ -211,15 +208,16 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext
                                                   conv_info, depth_multiplier, act_info, dilation,
                                                   (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
                                                   (output_shifts != nullptr) ? output_shifts->info() : nullptr));
+
+    auto padding_info = get_padding_info({ input, weights, biases, output });
+
     auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
                                                     conv_info, depth_multiplier, dilation,
                                                     (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
                                                     (output_shifts != nullptr) ? output_shifts->info() : nullptr);
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-
-    const bool is_stride_1            = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
-    const bool is_stride_1_dilation_1 = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
 
+    const bool is_stride_1              = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
+    const bool is_stride_1_dilation_1   = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
     const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->info()->data_type());
     const bool is_dot8_supported        = dot8_supported(CLKernelLibrary::get().get_device()) && !is_quantized_per_channel;
 
@@ -228,31 +226,37 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext
     _weights                            = weights;
     _biases                             = biases;
     _conv_stride_y                      = conv_info.stride().second;
-    _num_rows_processed_per_iteration   = is_stride_1_dilation_1 ? 2 : 1;
     _num_planes_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
     _output_multipliers                 = output_multipliers;
     _output_shifts                      = output_shifts;
     _is_quantized                       = is_data_type_quantized_asymmetric(input->info()->data_type());
 
-    // If QASYMM8 and the 8 bit dot product is available, force _num_planes_processed_per_iteration to 1
-    if(is_dot8_supported && _is_quantized)
+    if(_is_quantized)
     {
-        _num_planes_processed_per_iteration = 1;
-    }
+        _border_size = BorderSize(is_stride_1 ? 0 : conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
 
-    _border_size = BorderSize(_is_quantized && is_stride_1 ? 0 : conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
+        // If QASYMM8 and the 8 bit dot product is available, force _num_planes_processed_per_iteration to 1
+        if(is_dot8_supported)
+        {
+            _num_planes_processed_per_iteration = 1;
+        }
+    }
 
-    const unsigned int num_elems_accessed_per_iteration = _is_quantized ? 4 : (8 / input->info()->element_size());
+    unsigned int num_elems_accessed_per_iteration = _is_quantized ? 4 : adjust_vec_size(4 / input->info()->element_size(), input->info()->dimension(0));
+    unsigned int num_rows_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
 
     CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
     build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation())));
-    build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_accessed_per_iteration));
+    build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
     build_opts.add_option("-DSRC_DIM_2=" + support::cpp11::to_string(_input->info()->dimension(2)));
     build_opts.add_option("-DCONV_PAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
     build_opts.add_option("-DCONV_PAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
-    build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
-    build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
+    build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(input->info()->dimension(0) % num_elems_accessed_per_iteration));
+    build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
+    build_opts.add_option_if(_input->info()->tensor_shape().total_size_upper(3) > 1,
+                             "-DDST_DEPTH=" + support::cpp11::to_string(static_cast<int>(std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)))));
 
     if(_is_quantized)
     {
@@ -291,7 +295,6 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext
             build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
         }
 
-        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
         build_opts.add_option("-DWEIGHTS_TYPE=" + get_cl_type_from_data_type(weights->info()->data_type()));
         build_opts.add_option("-DWEIGHTS_PROMOTED_TYPE=" + get_cl_promoted_type_from_data_type(weights->info()->data_type()));
     }
@@ -299,22 +302,23 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext
     {
         build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
         build_opts.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
-        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
     }
 
     if(is_stride_1_dilation_1)
     {
-        build_opts.add_option("-DNUM_ROWS_PROCESSED=" + support::cpp11::to_string(_num_rows_processed_per_iteration));
+        build_opts.add_option("-DNUM_ROWS_PROCESSED=" + support::cpp11::to_string(num_rows_processed_per_iteration));
         build_opts.add_option("-DNUM_PLANES_PROCESSED=" + support::cpp11::to_string(_num_planes_processed_per_iteration));
+        build_opts.add_option("-DDST_DIM_1=" + support::cpp11::to_string(_output->info()->dimension(1)));
         build_opts.add_option("-DDST_DIM_2=" + support::cpp11::to_string(_output->info()->dimension(2)));
+        build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string((input->info()->dimension(1) + conv_info.pad_left() + conv_info.pad_right()) % num_rows_processed_per_iteration));
     }
     else
     {
         build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
         build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(_conv_stride_y));
+        build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
+        build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
     }
-    build_opts.add_option_if(_input->info()->tensor_shape().total_size_upper(3) > 1,
-                             "-DDST_DEPTH=" + support::cpp11::to_string(static_cast<int>(std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)))));
 
     std::string kernel_name;
     // Create kernel
@@ -331,12 +335,11 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext
         kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
     }
 
-    build_opts.add_option_if(input->info()->data_type() == DataType::F16, "-DIS_F16");
-    build_opts.add_option_if(input->info()->data_type() == DataType::F32, "-DIS_F32");
-
     ICLKernel::configure_internal(win_config.second);
     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
 
+    ARM_COMPUTE_ERROR_ON(!_is_quantized && has_padding_changed(padding_info));
+
     // Set config_id for enabling LWS tuning
     _config_id = kernel_name;
     _config_id += "_";
@@ -364,7 +367,6 @@ Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *inp
                                                               (output_multipliers != nullptr) ? output_multipliers->clone().get() : nullptr,
                                                               (output_shifts != nullptr) ? output_shifts->clone().get() : nullptr)
                                 .first);
-
     return Status{};
 }
 
@@ -373,23 +375,11 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::Com
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
 
-    // Collapse window
-    Window       window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
-    const size_t total_batches    = _input->info()->tensor_shape().total_size_upper(3);
+    const size_t total_batches = _input->info()->tensor_shape().total_size_upper(3);
 
-    Window win = window_collapsed;
+    Window win = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
     win.set(Window::DimZ, Window::Dimension(0, std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)) * total_batches, 1));
 
-    // Create input window and adjust
-    Window win_in = win;
-    win_in.set_dimension_step(Window::DimY, _num_rows_processed_per_iteration);
-    win_in.set_dimension_step(Window::DimZ, _conv_stride_y);
-
-    ARM_COMPUTE_ERROR_ON((win_in.y().step() < window.y().step()) || (win_in.z().step() < window.z().step()));
-
-    Window slice_in  = win_in.first_slice_window_4D();
-    Window slice_out = win.first_slice_window_4D();
-
     unsigned int idx = 2 * num_arguments_per_4D_tensor() + (_is_quantized ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
 
     if(_is_quantized)
@@ -409,60 +399,64 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::Com
         add_1D_tensor_argument(idx, _biases, win_biases);
     }
 
-    // Calculate the max_offset.
-    // max_offset is the offset for the last NOT valid value in the Z dimension (spatial dimension Y for NHWC)
-    //  |******************|
-    //  |     pad_top      |
-    //  |******************|
-    //  |                  |
-    //  |      plane0      |
-    //  |      batch0      |
-    //  |__________________|
-    //  |******************|       Batch 0
-    //  |    pad_bottom    |
-    //  |     pad_top      |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |      batch0      |
-    //  |__________________|-----> max_offset
-    //  |******************|
-    //  |    pad_bottom    |
-    //  |     pad_top      |
-    //  |******************|
-    //  |                  |
-    //  |      plane0      |
-    //  |      batch1      |
-    //  |__________________|
-    //  |******************|       Batch 1
-    //  |    pad_bottom    |
-    //  |     pad_top      |
-    //  |******************|
-    //  |                  |
-    //  |      plane1      |
-    //  |      batch1      |
-    //  |__________________|
-    //  |     pad_bottom   |
-    //  |******************|
-    const int max_offset = _input->info()->strides_in_bytes().z() * _input->info()->dimension(2) - (_input->info()->padding().bottom + _input->info()->padding().top) *
-                           _input->info()->strides_in_bytes().y();
-    _kernel.setArg(idx, max_offset);
+    if(_is_quantized)
+    {
+        // Calculate the max_offset.
+        // max_offset is the offset for the last NOT valid value in the Z dimension (spatial dimension Y for NHWC)
+        //  |******************|
+        //  |     pad_top      |
+        //  |******************|
+        //  |                  |
+        //  |      plane0      |
+        //  |      batch0      |
+        //  |__________________|
+        //  |******************|       Batch 0
+        //  |    pad_bottom    |
+        //  |     pad_top      |
+        //  |******************|
+        //  |                  |
+        //  |      plane1      |
+        //  |      batch0      |
+        //  |__________________|-----> max_offset
+        //  |******************|
+        //  |    pad_bottom    |
+        //  |     pad_top      |
+        //  |******************|
+        //  |                  |
+        //  |      plane0      |
+        //  |      batch1      |
+        //  |__________________|
+        //  |******************|       Batch 1
+        //  |    pad_bottom    |
+        //  |     pad_top      |
+        //  |******************|
+        //  |                  |
+        //  |      plane1      |
+        //  |      batch1      |
+        //  |__________________|
+        //  |     pad_bottom   |
+        //  |******************|
+        const int max_offset = _input->info()->strides_in_bytes().z() * _input->info()->dimension(2) - (_input->info()->padding().bottom + _input->info()->padding().top) *
+                               _input->info()->strides_in_bytes().y();
+        _kernel.setArg(idx, max_offset);
+    }
 
+    Window slice = win.first_slice_window_4D();
     do
     {
         unsigned int idx = 0;
-        add_4D_tensor_argument(idx, _input, slice_in);
-        add_4D_tensor_argument(idx, _output, slice_out);
+        add_4D_tensor_argument(idx, _input, slice);
+        add_4D_tensor_argument(idx, _output, slice);
         if(_is_quantized)
         {
-            add_2D_tensor_argument(idx, _weights, slice_out);
+            add_2D_tensor_argument(idx, _weights, slice);
         }
         else
         {
-            add_3D_tensor_argument(idx, _weights, slice_out);
+            add_3D_tensor_argument(idx, _weights, slice);
         }
-        enqueue(queue, *this, slice_out, lws_hint());
+        enqueue(queue, *this, slice, lws_hint());
     }
-    while(win.slide_window_slice_4D(slice_out) && win_in.slide_window_slice_4D(slice_in));
+    while(win.slide_window_slice_4D(slice));
 }
 } // namespace arm_compute