Port CLConvertFullyConnectedWeights to new API

* Replace ICLKernel by IClKernel in other unrelated kernels

Resolves partially: COMPMID-4187

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I173b8f2ac645dbfd7d412f4b058c5c9655c229ee
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5402
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

9 files changed, 202 insertions, 7 deletions

diff --git a/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp b/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp
new file mode 100644
index 0000000000..49f2f68a76
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2018-2021 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 "src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.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/Utils.h"
+#include "src/core/CL/CLValidate.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+#include "support/Cast.h"
+#include "support/StringSupport.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+void ClConvertFullyConnectedWeightsKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const TensorShape &original_src_shape,
+                                                     DataLayout data_layout)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
+    // Output tensor auto initialisation if not yet initialized
+    auto_init_if_empty(*dst, *src->clone());
+
+    auto padding_info = get_padding_info({ src, dst });
+
+    ARM_COMPUTE_ERROR_THROW_ON(ClConvertFullyConnectedWeightsKernel::validate(src, dst, original_src_shape, data_layout));
+
+    const DataLayout src_data_layout = (data_layout == DataLayout::NCHW) ? DataLayout::NHWC : DataLayout::NCHW;
+
+    const int width_idx   = get_data_layout_dimension_index(src_data_layout, DataLayoutDimension::WIDTH);
+    const int height_idx  = get_data_layout_dimension_index(src_data_layout, DataLayoutDimension::HEIGHT);
+    const int channel_idx = get_data_layout_dimension_index(src_data_layout, DataLayoutDimension::CHANNEL);
+
+    const unsigned int num_elems_per_src_plane = original_src_shape[width_idx] * original_src_shape[height_idx];
+    const unsigned int num_channels            = original_src_shape[channel_idx];
+
+    const unsigned int factor_1 = (data_layout == DataLayout::NCHW) ? num_elems_per_src_plane : num_channels;
+    const unsigned int factor_2 = (data_layout == DataLayout::NCHW) ? num_channels : num_elems_per_src_plane;
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(src->element_size()));
+    build_opts.add_option("-DFACTOR_1=" + support::cpp11::to_string(factor_1));
+    build_opts.add_option("-DFACTOR_2=" + support::cpp11::to_string(factor_2));
+
+    // Create kernel
+    _kernel = create_kernel(compile_context, "convert_fc_weights", build_opts.options());
+
+    // Configure kernel window
+    Window win = calculate_max_window(*src, Steps());
+    ICLKernel::configure_internal(win);
+
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
+}
+
+Status ClConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const TensorShape &original_src_shape,
+                                                      DataLayout data_layout)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
+    ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN);
+    ARM_COMPUTE_RETURN_ERROR_ON(src->num_dimensions() != 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(1) != original_src_shape.total_size_lower(3));
+    ARM_COMPUTE_RETURN_ERROR_ON(data_layout == DataLayout::UNKNOWN);
+
+    // Checks performed when dst is configured
+    if(dst->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
+    }
+
+    return Status{};
+}
+
+void ClConvertFullyConnectedWeightsKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
+    auto       dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
+    unsigned int idx = 0;
+    add_2D_tensor_argument(idx, src, window);
+    add_2D_tensor_argument(idx, dst, window);
+    enqueue(queue, *this, window, lws_hint());
+}
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.h b/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.h
new file mode 100644
index 0000000000..11ab4d2a0d
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.h
@@ -0,0 +1,76 @@
+/*
+ * Copyright (c) 2018-2021 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.
+ */
+#ifndef ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H
+#define ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H
+
+#include "src/core/common/Macros.h"
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface to convert the 2D Fully Connected weights from NCHW to NHWC or vice versa.
+ *
+ * @note This function can be applied to the 2D weights used by a Fully Connected layer if:
+ *       - It follows a Convolution layer
+ *       - The data layout used by the network does not match the one the model has been trained in.
+ *
+ * @note This function assumes the weights are already reshaped (transposed)
+ */
+namespace opencl
+{
+namespace kernels
+{
+class ClConvertFullyConnectedWeightsKernel : public IClKernel
+{
+public:
+    ClConvertFullyConnectedWeightsKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClConvertFullyConnectedWeightsKernel);
+    /** Set the src and dst tensor.
+     *
+     * @param[in]  compile_context    The compile context to be used.
+     * @param[in]  src                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: All.
+     * @param[out] dst                The converted weights tensor info. Shape and Data Type: Same as @p src.
+     * @param[in]  original_src_shape Shape of the original src tensor (the one entering fully connected layer).
+     * @param[in]  data_layout        The data layout the weights have been trained in.
+     */
+    void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const TensorShape &original_src_shape, DataLayout data_layout);
+    /** Static function to check if given info will lead to a valid configuration of @ref ClConvertFullyConnectedWeightsKernel
+     *
+     * @param[in] src                Source weights tensor info to convert. Must be 2 dimensional. Data types supported: All.
+     * @param[in] dst                The converted weights tensor info. Shape and Data Type: Same as @p src.
+     * @param[in] original_src_shape Shape of the original src tensor (the one entering fully connected layer).
+     * @param[in] data_layout        The data layout the weights have been trained in.
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const TensorShape &original_src_shape, DataLayout data_layout);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override;
+};
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CLCONVERTFULLYCONNECTEDWEIGHTSKERNEL_H */
diff --git a/src/core/gpu/cl/kernels/ClCopyKernel.h b/src/core/gpu/cl/kernels/ClCopyKernel.h
index 9d5457d72d..b1b9672bcb 100644
--- a/src/core/gpu/cl/kernels/ClCopyKernel.h
+++ b/src/core/gpu/cl/kernels/ClCopyKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** OpenCL kernel to perform a copy between two tensors */
-class ClCopyKernel : public ICLKernel
+class ClCopyKernel : public IClKernel
 {
 public:
     ClCopyKernel() = default;
diff --git a/src/core/gpu/cl/kernels/ClCropKernel.h b/src/core/gpu/cl/kernels/ClCropKernel.h
index 92f94d471b..ec0f8e58da 100644
--- a/src/core/gpu/cl/kernels/ClCropKernel.h
+++ b/src/core/gpu/cl/kernels/ClCropKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** OpenCL kernel to perform a copy between two tensors */
-class ClCropKernel : public ICLKernel
+class ClCropKernel : public IClKernel
 {
 public:
     ClCropKernel() = default;
diff --git a/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h b/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h
index 5acfb33199..103ef00695 100644
--- a/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h
+++ b/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h
@@ -37,7 +37,7 @@ namespace kernels
 /** Interface for the depth concatenate kernel.
  *  The src tensor will be concatenated into the dst tensor.
  */
-class ClDepthConcatenateKernel : public ICLKernel
+class ClDepthConcatenateKernel : public IClKernel
 {
 public:
     /** Default constructor */
diff --git a/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h b/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h
index ff2f5619db..384b561003 100644
--- a/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h
+++ b/src/core/gpu/cl/kernels/ClDirectConvolutionKernel.h
@@ -36,7 +36,7 @@ namespace kernels
 {
 /** Interface for the  direct convolution kernel.
  */
-class ClDirectConvolutionKernel : public ICLKernel
+class ClDirectConvolutionKernel : public IClKernel
 {
 public:
     ClDirectConvolutionKernel() = default;
diff --git a/src/core/gpu/cl/kernels/ClFillKernel.h b/src/core/gpu/cl/kernels/ClFillKernel.h
index 136c1243fe..b439eac0de 100644
--- a/src/core/gpu/cl/kernels/ClFillKernel.h
+++ b/src/core/gpu/cl/kernels/ClFillKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** Interface for filling the planes of a tensor */
-class ClFillKernel : public ICLKernel
+class ClFillKernel : public IClKernel
 {
 public:
     ClFillKernel() = default;
diff --git a/src/core/gpu/cl/kernels/ClPixelWiseMultiplicationKernel.h b/src/core/gpu/cl/kernels/ClPixelWiseMultiplicationKernel.h
index 64b6aa1eda..5889b84938 100644
--- a/src/core/gpu/cl/kernels/ClPixelWiseMultiplicationKernel.h
+++ b/src/core/gpu/cl/kernels/ClPixelWiseMultiplicationKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** Interface for the pixelwise multiplication kernel. */
-class ClPixelWiseMultiplicationKernel : public ICLKernel
+class ClPixelWiseMultiplicationKernel : public IClKernel
 {
 public:
     /** Default constructor */
diff --git a/src/core/gpu/cl/kernels/ClReshapeKernel.h b/src/core/gpu/cl/kernels/ClReshapeKernel.h
index ee835c0fd3..0501b93f40 100644
--- a/src/core/gpu/cl/kernels/ClReshapeKernel.h
+++ b/src/core/gpu/cl/kernels/ClReshapeKernel.h
@@ -35,7 +35,7 @@ namespace opencl
 namespace kernels
 {
 /** Interface for the kernel to perform tensor reshaping */
-class ClReshapeKernel : public ICLKernel
+class ClReshapeKernel : public IClKernel
 {
 public:
     ClReshapeKernel() = default;