COMPMID-2063: New Winograd implementation
Refactoring of winograd code reducing the size of the binaries
about 8X.
Change-Id: If8845bda324573e1a5cf436f354ac8603e88a92e
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/959
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Anthony Barbier <Anthony.barbier@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/src/core/NEON/kernels/convolution/winograd/winograd_transforms/input.hpp b/src/core/NEON/kernels/convolution/winograd/winograd_transforms/input.hpp
new file mode 100644
index 0000000..fcbd21f
--- /dev/null
+++ b/src/core/NEON/kernels/convolution/winograd/winograd_transforms/input.hpp
@@ -0,0 +1,265 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+
+#pragma once
+
+#include "winograd.hpp"
+#include "padding.hpp"
+
+#define MEMBERFN(RTYPE) template <\
+ int InnerTileRows, int InnerTileCols,\
+ typename TIn, typename TOut, WinogradRoots Roots\
+> RTYPE InputTransform<InnerTileRows, InnerTileCols, TIn, TOut, Roots>
+
+
+#define Nx1MEMBERFN(RTYPE) template <\
+ int InnerTileRows, typename TIn, typename TOut, WinogradRoots Roots\
+> RTYPE InputTransform<InnerTileRows, 1, TIn, TOut, Roots>
+
+namespace winograd
+{
+
+MEMBERFN()::InputTransform(
+ const int kernel_rows,
+ const int kernel_cols,
+ const int n_batches,
+ const int n_rows,
+ const int n_cols,
+ const int n_channels,
+ const int padding_top,
+ const int padding_left,
+ const int padding_bottom,
+ const int padding_right
+) : _n_batches(n_batches), _n_rows(n_rows), _n_cols(n_cols), _n_channels(n_channels),
+ _inptr(nullptr), _outptr(nullptr),
+ _overlap_rows(kernel_rows - 1), _overlap_cols(kernel_cols - 1),
+ _padding_top(padding_top), _padding_left(padding_left), _padding_bottom(padding_bottom), _padding_right(padding_right),
+ _tiles_M(iceildiv(padding_top + n_rows + padding_bottom - kernel_rows + 1, InnerTileRows - kernel_rows + 1)),
+ _tiles_N(iceildiv(padding_left + n_cols + padding_right - kernel_cols + 1, InnerTileCols - kernel_cols + 1)),
+ _matrix_stride(0), _matrix_row_stride(0), _matrix_batch_stride(0),
+ _in_col_stride(0), _in_row_stride(0), _in_batch_stride(0),
+ _working_space_col_stride(n_channels),
+ _working_space_row_stride(InnerTileCols * _working_space_col_stride),
+ _working_space(nullptr)
+{
+}
+
+MEMBERFN(void)::set_input_tensor(const void* const inptr)
+{
+ set_input_tensor(inptr, _n_channels);
+}
+
+MEMBERFN(void)::set_input_tensor(const void* const inptr, const int ldcol)
+{
+ set_input_tensor(inptr, _n_cols * ldcol, ldcol);
+}
+
+MEMBERFN(void)::set_input_tensor(const void* const inptr, const int ldrow, const int ldcol)
+{
+ set_input_tensor(inptr, _n_rows * ldrow, ldrow, ldcol);
+}
+
+MEMBERFN(void)::set_input_tensor(const void* const inptr, const int ldbatch, const int ldrow, const int ldcol)
+{
+ _inptr = static_cast<const TIn *>(inptr);
+ _in_batch_stride = ldbatch;
+ _in_row_stride = ldrow;
+ _in_col_stride = ldcol;
+}
+
+MEMBERFN(void)::set_output_matrices(void * const mptr, const int ldmatrix, const int ldrow)
+{
+ _outptr = static_cast<TOut *>(mptr);
+ _matrix_stride = ldmatrix;
+ _matrix_row_stride = ldrow;
+ _matrix_batch_stride = _tiles_M * _tiles_N * ldrow;
+}
+
+Nx1MEMBERFN()::InputTransform(
+ const int kernel_rows,
+ const int kernel_cols,
+ const int n_batches,
+ const int n_rows,
+ const int n_cols,
+ const int n_channels,
+ const int padding_top,
+ const int padding_left,
+ const int padding_bottom,
+ const int padding_right
+) : InputTransform<1, InnerTileRows, TIn, TOut, Roots>::InputTransform(
+ /* Transpose rows and columns */
+ kernel_cols, kernel_rows, n_batches, n_cols, n_rows, n_channels,
+ padding_left, padding_top, padding_right, padding_bottom
+ )
+{
+}
+
+Nx1MEMBERFN(void)::set_input_tensor(const void* const inptr)
+{
+ set_input_tensor(inptr, this->_n_channels);
+}
+
+Nx1MEMBERFN(void)::set_input_tensor(const void* const inptr, const int ldcol)
+{
+ set_input_tensor(inptr, this->_n_cols * ldcol, ldcol);
+}
+
+Nx1MEMBERFN(void)::set_input_tensor(const void* const inptr, const int ldrow, const int ldcol)
+{
+ set_input_tensor(inptr, this->_n_rows * ldrow, ldrow, ldcol);
+}
+
+Nx1MEMBERFN(void)::set_input_tensor(const void* const inptr, const int ldbatch, const int ldrow, const int ldcol)
+{
+ // Transpose row and column strides
+ Base::set_input_tensor(inptr, ldbatch, ldcol, ldrow);
+}
+
+MEMBERFN(size_t)::get_working_space_size(const unsigned int nthreads) const
+{
+ return sizeof(TIn) * InnerTileRows * _working_space_row_stride * nthreads;
+}
+
+MEMBERFN(void)::set_working_space(void * const buffer)
+{
+ _working_space = static_cast<TIn *>(buffer);
+}
+
+MEMBERFN(unsigned int)::get_window(void) const
+{
+ return iceildiv(_n_channels, WINDOW_BLOCK);
+}
+
+MEMBERFN(void)::run(
+ const unsigned int start,
+ const unsigned int stop,
+ const unsigned int threadid
+)
+{
+ // Determine the channels on which to work
+ if (start >= get_window())
+ {
+ return; // No work to do beyond the end of the window
+ }
+ const unsigned int start_channel = start * WINDOW_BLOCK;
+ const unsigned int stop_channel = std::min<unsigned int>(_n_channels , stop * WINDOW_BLOCK);
+ const unsigned int n_channels = stop_channel - start_channel;
+
+ // Loop over batches
+ for (int batch = 0; batch < _n_batches; batch++)
+ {
+ const TIn* const inptr_batch = _inptr + start_channel + batch*_in_batch_stride;
+ TOut* const outptr_batch = _outptr + start_channel + batch*_matrix_batch_stride;
+
+ // Loop over rows of tiles
+ for (int tile_i = 0; tile_i < _tiles_M; tile_i++)
+ {
+ // Compute the starting and ending row of pixels within the row of tiles,
+ // hence compute the padding to apply to the top and bottom of each tile.
+ const int row_top = tile_i * (InnerTileRows - _overlap_rows) - _padding_top;
+ const int row_bottom = row_top + InnerTileRows;
+ const int row_pad_top = std::max(0, _padding_top - tile_i * (InnerTileRows - _overlap_rows));
+ const int row_pad_bottom = std::max(0, row_bottom - _n_rows);
+
+ // Get a pointer to the start of the row.
+ const int row_offset = std::min(0, row_pad_top - _padding_top);
+ const TIn* const inptr_row = inptr_batch + _in_row_stride*(row_offset + tile_i*(InnerTileRows - _overlap_rows));
+ TOut* const outptr_row = outptr_batch + tile_i*_tiles_N*_matrix_row_stride;
+
+ // Loop over tiles within the row
+ for (int tile_j = 0; tile_j < _tiles_N; tile_j++)
+ {
+ // Compute the starting and ending column of pixels within the tile,
+ // hence compute the padding to apply to the left and right of the
+ // tile.
+ const int tile_left = tile_j * (InnerTileCols - _overlap_cols) - _padding_left;
+ const int tile_right = tile_left + InnerTileCols;
+ const int tile_pad_left = std::max(0, _padding_left - tile_j * (InnerTileCols - _overlap_cols));
+ const int tile_pad_right = std::max(0, tile_right - _n_cols);
+
+ // Get a pointer to the start of the tile.
+ const int col_offset = std::min(0, tile_pad_left - _padding_left);
+ const TIn* const inptr_tile = inptr_row + _in_col_stride*(col_offset + tile_j*(InnerTileCols - _overlap_cols));
+ TOut* const outptr_tile = outptr_row + tile_j * _matrix_row_stride;
+
+ // Transform the tile, applying padding if necessary.
+ if (row_pad_top || tile_pad_left || row_pad_bottom || tile_pad_right)
+ {
+ transform_padded_tile(
+ threadid, n_channels, outptr_tile, inptr_tile,
+ row_pad_top, tile_pad_left, row_pad_bottom, tile_pad_right
+ );
+ }
+ else
+ {
+ transform_unpadded_tile(threadid, n_channels, outptr_tile, inptr_tile);
+ }
+ }
+ }
+ }
+}
+
+MEMBERFN(void)::transform_unpadded_tile(
+ const unsigned int /* threadid unused */,
+ const int n_channels,
+ TOut * const outptr,
+ const TIn * const inptr
+)
+{
+ transform_tile(
+ n_channels, inptr, _in_row_stride, _in_col_stride, outptr, _matrix_stride
+ );
+}
+
+MEMBERFN(void)::transform_padded_tile(
+ const unsigned int threadid,
+ const int n_channels,
+ TOut * const outptr,
+ const TIn * const inptr,
+ const int padding_top,
+ const int padding_left,
+ const int padding_bottom,
+ const int padding_right
+)
+{
+ padding::copy_and_pad_tile(
+ InnerTileRows, InnerTileCols, n_channels,
+ inptr, _in_row_stride, _in_col_stride,
+ static_cast<TIn *>(get_working_space(threadid)), _working_space_row_stride, _working_space_col_stride,
+ padding_top, padding_left, padding_bottom, padding_right
+ );
+
+ transform_tile(
+ n_channels, static_cast<const TIn *>(get_working_space(threadid)),
+ _working_space_row_stride, _working_space_col_stride,
+ outptr, _matrix_stride
+ );
+}
+
+MEMBERFN(void *)::get_working_space(const unsigned int threadid) const
+{
+ return _working_space + InnerTileRows * _working_space_row_stride * threadid;
+}
+
+} // namespace winograd