COMPMID-687: Winograd refactoring

Moved the headers into src/
Added pimpl pattern

Change-Id: I227f8b47468d8e14875d710aac8de5eb09463e2a
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111765
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>

23 files changed, 119 insertions, 14 deletions

diff --git a/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h b/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h
index 1e7ca64b8c..3ab3aa792b 100644
--- a/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h
@@ -25,17 +25,34 @@
 #define __ARM_COMPUTE_NEGEMMWINOGRADLAYERKERNEL_H__
 
 #include "arm_compute/core/NEON/INEKernel.h"
-
-#include "arm_compute/core/NEON/kernels/winograd/winograd_shim_nchw.hpp"
+#include "arm_compute/core/NEON/kernels/winograd/tensor.hpp"
 
 namespace arm_compute
 {
 class ITensor;
+class NEWinogradLayerKernel;
+class Winograd3x3F32
+{
+public:
+    friend class NEWinogradLayerKernel;
+    Winograd3x3F32(const KernelShape &kernel_shape, const Tensor4DShape input_shape, const PaddingType padding_type, void *kernel_storage);
+    ~Winograd3x3F32();
+    std::pair<void *, void *> get_nhwc_ptrs(const Tensor4DShape &input_shape, const PaddingType padding_type, void *working_space);
+    void transform_weights(const void *const kernel, void *transform_working_space);
+    void reshape_input(const Tensor4DShape &input_shape, const PaddingType padding_type, const void *const input, void *working_space);
+    void reshape_output(const Tensor4DShape &input_shape, const PaddingType padding_type, void *const output);
+    void nchw2nhwc(const Tensor4DShape &input_shape, const PaddingType padding_type, void *working_space, const void *const input);
+    void nhwc2nchw(const Tensor4DShape &input_shape, const PaddingType padding_type, void *working_space, void *const output);
+
+private:
+    class Private;
+    std::unique_ptr<Private> _pimpl;
+};
 
 class NEWinogradLayerKernel : public INEKernel
 {
 public:
-    using Winograd3x3F32 = winograd_shim_nchw::Winograd2x2_3x3GEMM<float, float>;
+    //    using Winograd3x3F32 = winograd_shim_nchw::Winograd2x2_3x3GEMM<float, float>;
 
     /** Constructor */
     NEWinogradLayerKernel();
@@ -61,9 +78,22 @@ public:
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
 
+    /* Get the memory required to instantiate a new Winograd operator.
+       */
+    static size_t get_kernel_storage_size(const KernelShape &shape);
+
+    /* Get the memory required to apply a Winograd operator to some input.
+       */
+    static size_t get_working_space_size(const Tensor4DShape &input_shape, const KernelShape &k_shape, const PaddingType padding);
+
+    /* Get the memory required to transform the kernel.
+       */
+    static size_t get_kernel_transform_working_size(const KernelShape &shape);
+
 protected:
     Winograd3x3F32 *_convolver;
-    ITensor        *_output;
+    //    std::unique_ptr<Winograd3x3F32> _conv;
+    ITensor *_output;
 };
 
 } // namespace arm_compute
diff --git a/arm_compute/runtime/NEON/functions/NEWinogradLayer.h b/arm_compute/runtime/NEON/functions/NEWinogradLayer.h
index 7dca4570e5..77707060ec 100644
--- a/arm_compute/runtime/NEON/functions/NEWinogradLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEWinogradLayer.h
@@ -67,8 +67,6 @@ public:
     NEWinogradLayer &operator=(const NEWinogradLayer &) = delete;
 
 private:
-    using Winograd3x3F32 = NEWinogradLayerKernel::Winograd3x3F32;
-
     MemoryGroup                     _memory_group;
     NEWinogradLayerKernel           _winograd_kernel;
     Tensor                          _weights_workspace;
diff --git a/scripts/clang_tidy_rules.py b/scripts/clang_tidy_rules.py
index 5b27dd5be5..7a13d045e7 100755
--- a/scripts/clang_tidy_rules.py
+++ b/scripts/clang_tidy_rules.py
@@ -91,6 +91,7 @@ def filter_clang_tidy_lines( lines ):
                ("parameter 'memory_manager' is unused" in line) or
                ("parameter 'memory_manager' is copied for each invocation but only used as a const reference" in line) or
                ("DeconvolutionLayer.cpp" in line and "casting (double + 0.5) to integer leads to incorrect rounding; consider using lround" in line) or
+               ("NEWinogradLayerKernel.cpp" in line and "use '= default' to define a trivial destructor" in line) or
                "3rdparty" in line):
                 print_context=False
                 continue
diff --git a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradLayerKernel.cpp
index b9109dcff2..fe633368c0 100644
--- a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEWinogradLayerKernel.cpp
@@ -27,9 +27,86 @@
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/TensorInfo.h"
+#include "support/ToolchainSupport.h"
+
+#include "src/core/NEON/kernels/winograd/winograd_shim_nchw.hpp"
+
+using T = winograd_shim_nchw::Winograd2x2_3x3GEMM<float, float>;
 
 namespace arm_compute
 {
+class Winograd3x3F32::Private
+{
+public:
+    Private(const KernelShape &kernel_shape, const Tensor4DShape input_shape, const PaddingType padding_type, void *kernel_storage)
+        : convolver(kernel_shape, input_shape, padding_type, kernel_storage)
+    {
+    }
+
+    T convolver;
+};
+
+Winograd3x3F32::~Winograd3x3F32()
+{
+}
+
+void Winograd3x3F32::nchw2nhwc(const Tensor4DShape &input_shape, const PaddingType padding_type, void *working_space, const void *const input)
+{
+    _pimpl->convolver.nchw2nhwc(input_shape, padding_type, working_space, reinterpret_cast<const float *>(input));
+}
+
+void Winograd3x3F32::nhwc2nchw(const Tensor4DShape &input_shape, const PaddingType padding_type, void *working_space, void *const output)
+{
+    _pimpl->convolver.nhwc2nchw(input_shape, padding_type, working_space, reinterpret_cast<float *const>(output));
+}
+
+void Winograd3x3F32::transform_weights(const void *const kernel, void *transform_working_space)
+{
+    _pimpl->convolver.transform_weights(reinterpret_cast<const float *>(kernel), transform_working_space);
+}
+
+void Winograd3x3F32::reshape_input(const Tensor4DShape &input_shape, const PaddingType padding_type, const void *const input, void *working_space)
+{
+    _pimpl->convolver.reshape_input(input_shape, padding_type, reinterpret_cast<const float *>(input), working_space);
+}
+
+void Winograd3x3F32::reshape_output(const Tensor4DShape &input_shape, const PaddingType padding_type, void *const output)
+{
+#if defined(__aarch64__)
+    _pimpl->convolver.reshape_output(input_shape, padding_type, reinterpret_cast<float *const>(output));
+#else  /* __aarch64__ */
+    ARM_COMPUTE_UNUSED(input_shape);
+    ARM_COMPUTE_UNUSED(padding_type);
+    ARM_COMPUTE_UNUSED(output);
+    ARM_COMPUTE_ERROR("Not implemented");
+#endif /* __aarch64__ */
+}
+
+std::pair<void *, void *> Winograd3x3F32::get_nhwc_ptrs(const Tensor4DShape &input_shape, const PaddingType padding_type, void *working_space)
+{
+    return _pimpl->convolver.get_nhwc_ptrs(input_shape, padding_type, working_space);
+}
+
+Winograd3x3F32::Winograd3x3F32(const KernelShape &kernel_shape, const Tensor4DShape input_shape, const PaddingType padding_type, void *kernel_storage)
+    : _pimpl(support::cpp14::make_unique<Private>(kernel_shape, input_shape, padding_type, kernel_storage))
+{
+}
+
+size_t NEWinogradLayerKernel::get_kernel_storage_size(const KernelShape &shape)
+{
+    return T::get_kernel_storage_size(shape);
+}
+
+size_t NEWinogradLayerKernel::get_working_space_size(const Tensor4DShape &input_shape, const KernelShape &k_shape, const PaddingType padding)
+{
+    return T::get_working_space_size(input_shape, k_shape, padding);
+}
+
+size_t NEWinogradLayerKernel::get_kernel_transform_working_size(const KernelShape &shape)
+{
+    return T::get_kernel_transform_working_size(shape);
+}
+
 NEWinogradLayerKernel::NEWinogradLayerKernel()
     : _convolver(nullptr), _output(nullptr)
 {
@@ -55,6 +132,6 @@ void NEWinogradLayerKernel::run(const Window &window, const ThreadInfo &info)
     const size_t num_gemms_per_thread = 16 / num_threads;
     const size_t first_gemm           = tid * num_gemms_per_thread;
     const size_t last_gemm            = (tid == (num_threads - 1)) ? 15 : first_gemm + num_gemms_per_thread - 1;
-    _convolver->execute(first_gemm, last_gemm);
+    _convolver->_pimpl->convolver.execute(first_gemm, last_gemm);
 }
 } // namespace arm_compute
diff --git a/arm_compute/core/NEON/kernels/winograd/gemm.hpp b/src/core/NEON/kernels/winograd/gemm.hpp
index 564016a646..564016a646 100644
--- a/arm_compute/core/NEON/kernels/winograd/gemm.hpp
+++ b/src/core/NEON/kernels/winograd/gemm.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/gemm/a64_sgemm.hpp b/src/core/NEON/kernels/winograd/gemm/a64_sgemm.hpp
index e1b7488c31..e1b7488c31 100644
--- a/arm_compute/core/NEON/kernels/winograd/gemm/a64_sgemm.hpp
+++ b/src/core/NEON/kernels/winograd/gemm/a64_sgemm.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/gemm/a64_sgemm_4x16.hpp b/src/core/NEON/kernels/winograd/gemm/a64_sgemm_4x16.hpp
index e74610ef27..e74610ef27 100644
--- a/arm_compute/core/NEON/kernels/winograd/gemm/a64_sgemm_4x16.hpp
+++ b/src/core/NEON/kernels/winograd/gemm/a64_sgemm_4x16.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/perf.h b/src/core/NEON/kernels/winograd/perf.h
index 11fb0c452f..11fb0c452f 100644
--- a/arm_compute/core/NEON/kernels/winograd/perf.h
+++ b/src/core/NEON/kernels/winograd/perf.h
diff --git a/arm_compute/core/NEON/kernels/winograd/profiler.hpp b/src/core/NEON/kernels/winograd/profiler.hpp
index 143192b589..143192b589 100644
--- a/arm_compute/core/NEON/kernels/winograd/profiler.hpp
+++ b/src/core/NEON/kernels/winograd/profiler.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/shims.hpp b/src/core/NEON/kernels/winograd/shims.hpp
index 249e5757f0..249e5757f0 100644
--- a/arm_compute/core/NEON/kernels/winograd/shims.hpp
+++ b/src/core/NEON/kernels/winograd/shims.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/transforms.hpp b/src/core/NEON/kernels/winograd/transforms.hpp
index 8546ee9e2e..8546ee9e2e 100644
--- a/arm_compute/core/NEON/kernels/winograd/transforms.hpp
+++ b/src/core/NEON/kernels/winograd/transforms.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/transforms/input_2x2_3x3.hpp b/src/core/NEON/kernels/winograd/transforms/input_2x2_3x3.hpp
index 7013c66ac0..ca8d012e5e 100644
--- a/arm_compute/core/NEON/kernels/winograd/transforms/input_2x2_3x3.hpp
+++ b/src/core/NEON/kernels/winograd/transforms/input_2x2_3x3.hpp
@@ -22,7 +22,8 @@
  * SOFTWARE.
  */
 #pragma once
-#include "../tensor.hpp"
+#include "arm_compute/core/NEON/kernels/winograd/tensor.hpp"
+
 
 namespace winograd {
   /* Transform an input tensor into the Winograd domain.
diff --git a/arm_compute/core/NEON/kernels/winograd/transforms/input_2x2_3x3/a64_float.hpp b/src/core/NEON/kernels/winograd/transforms/input_2x2_3x3/a64_float.hpp
index a99cbe325b..a99cbe325b 100644
--- a/arm_compute/core/NEON/kernels/winograd/transforms/input_2x2_3x3/a64_float.hpp
+++ b/src/core/NEON/kernels/winograd/transforms/input_2x2_3x3/a64_float.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/transforms/input_2x2_3x3/a64_float_channelwise.hpp b/src/core/NEON/kernels/winograd/transforms/input_2x2_3x3/a64_float_channelwise.hpp
index ad1ad55291..ad1ad55291 100644
--- a/arm_compute/core/NEON/kernels/winograd/transforms/input_2x2_3x3/a64_float_channelwise.hpp
+++ b/src/core/NEON/kernels/winograd/transforms/input_2x2_3x3/a64_float_channelwise.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/transforms/kernel_2x2_3x3.hpp b/src/core/NEON/kernels/winograd/transforms/kernel_2x2_3x3.hpp
index 033442aa14..033442aa14 100644
--- a/arm_compute/core/NEON/kernels/winograd/transforms/kernel_2x2_3x3.hpp
+++ b/src/core/NEON/kernels/winograd/transforms/kernel_2x2_3x3.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/transforms/kernel_2x2_3x3/a64_float.hpp b/src/core/NEON/kernels/winograd/transforms/kernel_2x2_3x3/a64_float.hpp
index 3dd62d1ac1..3dd62d1ac1 100644
--- a/arm_compute/core/NEON/kernels/winograd/transforms/kernel_2x2_3x3/a64_float.hpp
+++ b/src/core/NEON/kernels/winograd/transforms/kernel_2x2_3x3/a64_float.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/transforms/output_2x2_3x3.hpp b/src/core/NEON/kernels/winograd/transforms/output_2x2_3x3.hpp
index 0992c0bb44..0992c0bb44 100644
--- a/arm_compute/core/NEON/kernels/winograd/transforms/output_2x2_3x3.hpp
+++ b/src/core/NEON/kernels/winograd/transforms/output_2x2_3x3.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/transforms/output_2x2_3x3/a64_float.hpp b/src/core/NEON/kernels/winograd/transforms/output_2x2_3x3/a64_float.hpp
index 5925f9d569..5925f9d569 100644
--- a/arm_compute/core/NEON/kernels/winograd/transforms/output_2x2_3x3/a64_float.hpp
+++ b/src/core/NEON/kernels/winograd/transforms/output_2x2_3x3/a64_float.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/transforms/output_2x2_3x3/a64_float_two_stage.hpp b/src/core/NEON/kernels/winograd/transforms/output_2x2_3x3/a64_float_two_stage.hpp
index f551b12b52..f551b12b52 100644
--- a/arm_compute/core/NEON/kernels/winograd/transforms/output_2x2_3x3/a64_float_two_stage.hpp
+++ b/src/core/NEON/kernels/winograd/transforms/output_2x2_3x3/a64_float_two_stage.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/utils.hpp b/src/core/NEON/kernels/winograd/utils.hpp
index 14e709f028..14e709f028 100644
--- a/arm_compute/core/NEON/kernels/winograd/utils.hpp
+++ b/src/core/NEON/kernels/winograd/utils.hpp
diff --git a/arm_compute/core/NEON/kernels/winograd/winograd_gemm.hpp b/src/core/NEON/kernels/winograd/winograd_gemm.hpp
index c990cd0252..59afa2f5ab 100644
--- a/arm_compute/core/NEON/kernels/winograd/winograd_gemm.hpp
+++ b/src/core/NEON/kernels/winograd/winograd_gemm.hpp
@@ -26,7 +26,6 @@
 #include <cstdlib>
 #include <cassert>
 
-#include "alloc.hpp"
 #include "gemm.hpp"
 #include "profiler.hpp"
 #include "utils.hpp"
diff --git a/arm_compute/core/NEON/kernels/winograd/winograd_shim_nchw.hpp b/src/core/NEON/kernels/winograd/winograd_shim_nchw.hpp
index 4c7e291c58..c5bcffbaef 100644
--- a/arm_compute/core/NEON/kernels/winograd/winograd_shim_nchw.hpp
+++ b/src/core/NEON/kernels/winograd/winograd_shim_nchw.hpp
@@ -25,7 +25,6 @@
 #include <cstdint>
 #include <cstdlib>
 
-#include "alloc.hpp"
 #include "gemm.hpp"
 #include "profiler.hpp"
 #include "utils.hpp"
diff --git a/src/runtime/NEON/functions/NEWinogradLayer.cpp b/src/runtime/NEON/functions/NEWinogradLayer.cpp
index a9dec4ea0d..3251de4ae4 100644
--- a/src/runtime/NEON/functions/NEWinogradLayer.cpp
+++ b/src/runtime/NEON/functions/NEWinogradLayer.cpp
@@ -83,18 +83,18 @@ void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, co
 
     // Get the memory required to instantiate a new Winograd operator.
     constexpr size_t kstore_alignment          = 64;
-    const size_t     kernel_storage_per_thread = Winograd3x3F32::get_kernel_storage_size(kernel_shape);
+    const size_t     kernel_storage_per_thread = NEWinogradLayerKernel::get_kernel_storage_size(kernel_shape);
     _kernel_storage.allocator()->init(TensorInfo(TensorShape{ (kernel_storage_per_thread + kstore_alignment - 1) }, 1, DataType::U8));
     _memory_group.manage(&_kernel_storage);
 
     // Get workbench size and allocate memory
     constexpr size_t wspace_alignment = 64;
-    const size_t     ws_size          = Winograd3x3F32::get_working_space_size(in_shape, kernel_shape, padding);
+    const size_t     ws_size          = NEWinogradLayerKernel::get_working_space_size(in_shape, kernel_shape, padding);
     _workspace.allocator()->init(TensorInfo(TensorShape{ (ws_size + wspace_alignment - 1) }, 1, DataType::U8));
     _memory_group.manage(&_workspace);
 
     // Workspace for weights transform
-    const size_t weights_transform_size = Winograd3x3F32::get_kernel_transform_working_size(kernel_shape);
+    const size_t weights_transform_size = NEWinogradLayerKernel::get_kernel_transform_working_size(kernel_shape);
     _weights_workspace.allocator()->init(TensorInfo(TensorShape{ (weights_transform_size + wspace_alignment - 1) }, 1, DataType::U8));
     _memory_group.manage(&_weights_workspace);
 
@@ -125,7 +125,7 @@ void NEWinogradLayer::run()
     _conv->nchw2nhwc(in_shape, padding, _workspace.buffer(), reinterpret_cast<const float *>(_input->buffer()));
 
     //Get ptrs into the workspace
-    std::pair<float *, float *> nhwc_ptrs = _conv->get_nhwc_ptrs(in_shape, padding, _workspace.buffer());
+    std::pair<void *, void *> nhwc_ptrs = _conv->get_nhwc_ptrs(in_shape, padding, _workspace.buffer());
 
     //Setup matrices ptrs and transfor the input tensor to the appropriate form before running GEMM.
     _conv->reshape_input(in_shape, padding, nhwc_ptrs.second, _workspace.buffer());