COMPMID-935 - Implementing Convolution with Winograd on OpenCL (part 4)

Implemented Winograd Output Transform (2x2,3x3) on OpenCL
Implemented CLWinogradConvolutionLayer on OpenCL

Change-Id: I6a113fc5f052ca07f878d2b800d2ab003f84af65
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/125148
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

5 files changed, 211 insertions, 14 deletions

diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index 9c69800928..7b785bb8da 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -55,6 +55,7 @@ inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *i
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the matrix B must be <= 3");
 
     if(!is_interleaved_transposed)
     {
@@ -174,7 +175,7 @@ inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *inpu
 } // namespace
 
 CLGEMMMatrixMultiplyKernel::CLGEMMMatrixMultiplyKernel()
-    : _input0(nullptr), _input1(nullptr), _output(nullptr)
+    : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true)
 {
 }
 
@@ -192,9 +193,10 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
     // Perform validate step
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info));
 
-    _input0 = input0;
-    _input1 = input1;
-    _output = output;
+    _input0         = input0;
+    _input1         = input1;
+    _output         = output;
+    _slide_matrix_b = _input1->info()->num_dimensions() >= _input0->info()->num_dimensions();
 
     const DataType data_type = input0->info()->data_type();
     const int      fp_pos    = input0->info()->fixed_point_position();
@@ -257,6 +259,9 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
                                       "-DALPHA=" + float_to_string_with_full_precision(alpha));
     }
 
+    // Do not slide matrix B if _slide_matrix_b = false
+    build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
+
     std::string kernel_name;
     if(is_interleaved_transposed)
     {
@@ -365,7 +370,7 @@ void CLGEMMMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &que
         Window slice_b = slice;
         // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
         // This scenario can happen when the matrix multiplication is used to perform a convolution operation
-        if(_input1->info()->num_dimensions() < 3)
+        if(!_slide_matrix_b)
         {
             slice_b = slice_matrix_b;
         }
@@ -374,9 +379,9 @@ void CLGEMMMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &que
         add_2D_tensor_argument(idx, _input0, slice);
         add_2D_tensor_argument(idx, _input1, slice_b);
         add_2D_tensor_argument(idx, _output, slice);
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[3]));
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[3]));
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[3]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
         enqueue(queue, *this, slice, _lws_hint);
     }
     while(window.slide_window_slice_3D(slice));
diff --git a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
index 5489fde818..f69a39e4ad 100644
--- a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
@@ -76,15 +76,18 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     }
 
     AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    window_changed = window_changed || update_window_and_padding(win, input_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);
+        window_changed = window_changed || update_window_and_padding(win, input_access, output_access);
         output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), input->tensor_shape()));
     }
+    else
+    {
+        window_changed = window_changed || update_window_and_padding(win, input_access);
+    }
 
     // Collapse along the Z direction
     Window collapsed = win.collapse(win, Window::DimZ);
diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
index 3dbbe157b2..655b82bf66 100644
--- a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
@@ -76,7 +76,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
     AccessWindowStatic    output_access(output, 0, 0, output->dimension(0), output->dimension(1));
     window_changed = update_window_and_padding(win, input_access, output_access);
-    output_access.set_valid_region(win, input->valid_region());
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
 
     Window win_collapsed = win.collapse(win, Window::DimZ);
 
diff --git a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
index 72adb5f358..3b9350f9ba 100644
--- a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
@@ -44,11 +44,11 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_dims.width != 3 || kernel_dims.height != 3, "Winograd input transform only supports 3x3 kernels");
     ARM_COMPUTE_UNUSED(kernel_dims);
 
-    const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, conv_info, Size2D(3U, 3U));
-
     // Validate configured output
     if(output->total_size() != 0)
     {
+        const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, conv_info, kernel_dims);
+
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }
@@ -151,7 +151,8 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
 Status CLWinogradInputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(validate_arguments(input, output, conv_info, kernel_dims));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, conv_info, kernel_dims));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), conv_info, kernel_dims).first);
 
     return Status{};
 }
diff --git a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
new file mode 100644
index 0000000000..c9823275eb
--- /dev/null
+++ b/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp
@@ -0,0 +1,188 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "support/ToolchainSupport.h"
+
+#include <cmath>
+
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims, const Size2D &num_tiles)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area());
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_dims.width != 3 || kernel_dims.height != 3, "Only 3x3 kernels are supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(static_cast<unsigned int>(std::sqrt(input->dimension(2))) != 4, "Only 2x2 output tile is supported");
+    ARM_COMPUTE_UNUSED(kernel_dims);
+
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
+    }
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input, output_convolved_dims, DataLayout::NCHW));
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    constexpr unsigned int num_elems_processed_per_iteration = 1;
+
+    Window win            = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    bool   window_changed = false;
+
+    AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration);
+    AccessWindowStatic    output_access(output, 0, 0, ceil_to_multiple(output->dimension(0), 2), ceil_to_multiple(output->dimension(1), 2));
+
+    if(bias != nullptr)
+    {
+        AccessWindowStatic bias_access(bias, 0, 0, bias->dimension(0), bias->dimension(1));
+        window_changed = update_window_and_padding(win, input_access, bias_access, output_access);
+    }
+    else
+    {
+        window_changed = update_window_and_padding(win, input_access, output_access);
+    }
+    output->set_valid_region(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
+
+CLWinogradOutputTransformKernel::CLWinogradOutputTransformKernel()
+    : _input(nullptr), _bias(nullptr), _output(nullptr)
+{
+}
+
+void CLWinogradOutputTransformKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims,
+                                                const Size2D &num_tiles)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_UNUSED(kernel_dims);
+
+    // Output tensor auto initialization if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input->info(), output_convolved_dims, DataLayout::NCHW)));
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info(), kernel_dims, output_convolved_dims, num_tiles));
+
+    _input  = input;
+    _bias   = bias;
+    _output = output;
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option_if(_bias != nullptr, std::string("-DHAS_BIAS"));
+    build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles.width));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("winograd_output_transform_2x2_3x3_nchw", build_opts.options()));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure(win_config.second);
+
+    // Set config_id for enabling LWS tuning
+    _config_id = "winograd_output_transform_2x2_3x3";
+    _config_id += lower_string(string_from_data_type(input->info()->data_type()));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output->info()->dimension(1));
+}
+
+Status CLWinogradOutputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims,
+                                                 const Size2D &num_tiles)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, (bias != nullptr ? bias->clone().get() : nullptr), output, kernel_dims, output_convolved_dims, num_tiles));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), output->clone().get()).first);
+
+    return Status{};
+}
+
+void CLWinogradOutputTransformKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    // Get initial windows
+    Window slice = window.first_slice_window_3D();
+    slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
+
+    // Setup output slice
+    Window slice_out(slice);
+    slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    if(_bias != nullptr)
+    {
+        unsigned int idx1 = 2 * num_arguments_per_3D_tensor();
+        Window       slice_biases;
+        slice_biases.use_tensor_dimensions(_bias->info()->tensor_shape());
+        add_1D_tensor_argument(idx1, _bias, slice_biases);
+    }
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice_out);
+        enqueue(queue, *this, slice, _lws_hint);
+    }
+    while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_out));
+}
\ No newline at end of file