COMPMID-881: RSH new arm_gemm interface.
Change-Id: I1e2a1a77097d8017c274af3f97eba6964f80f5fa
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122592
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/src/core/NEON/kernels/arm_gemm/transform.hpp b/src/core/NEON/kernels/arm_gemm/transform.hpp
new file mode 100644
index 0000000..c80bb59
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/transform.hpp
@@ -0,0 +1,122 @@
+/*
+ * Copyright (c) 2017-2018 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.
+ */
+#pragma once
+
+/*
+ * Generic transform.
+ *
+ * Assuming the untransposed case, this works by first reading <BlockBy>
+ * consecutive values from the first input row. This same number of values
+ * are then read from the next <IntBy-1> rows. Now return to the first
+ * input row and repeat.
+ *
+ * Need to cope with the work requested in either dimension not actually
+ * being a multiple of the block sizes.
+ */
+template <unsigned IntBy, unsigned int BlockBy, bool Transposed, size_t TOutSize, size_t TInSize>
+struct TransformImpl
+{
+ template <typename TOut, typename TIn>
+ static void Transform(TOut *out, const TIn *const in, const int stride,
+ const int y0, const int ymax, const int x0, const int xmax)
+ {
+ const int n_whole_y_blocks = (ymax - y0) / IntBy;
+ const int y_remainders = (ymax - y0) % IntBy;
+ const int n_y_blocks = n_whole_y_blocks + (y_remainders ? 1 : 0);
+
+ const int n_whole_x_blocks = (xmax - x0) / BlockBy;
+ const int x_remainders = (xmax - x0) % BlockBy;
+ const int n_x_blocks = n_whole_x_blocks + (x_remainders ? 1 : 0);
+
+ // "Y" loop: advance down the rows of the source IntBy rows at a time.
+ // Set up fill_rows to show the number rows to copy from, and blank_rows
+ // for the number of blank rows to add.
+ for(int y_block = 0; y_block < n_y_blocks; y_block++)
+ {
+ int fill_rows = (y_block < n_whole_y_blocks) ? IntBy : y_remainders;
+ int blank_rows = IntBy - fill_rows;
+
+ int y_base = y0 + (y_block * IntBy);
+
+ // So now advance along this block of rows, BlockBy columns at a time.
+ for(int x_block = 0; x_block < n_x_blocks; x_block++)
+ {
+ int fill_cols = (x_block < n_whole_x_blocks) ? BlockBy : x_remainders;
+ int blank_cols = BlockBy - fill_cols;
+
+ int x_base = x0 + (x_block * BlockBy);
+
+ for(int row = 0; row < fill_rows; row++)
+ {
+ for(int col = 0; col < fill_cols; col++)
+ {
+ // In-range copy. If it's transposed, we reverse the sense of rows and columns here.
+ if(Transposed)
+ {
+ *out++ = static_cast<TOut>(in[(x_base + col) * stride + y_base + row]);
+ }
+ else
+ {
+ *out++ = static_cast<TOut>(in[(y_base + row) * stride + x_base + col]);
+ }
+ }
+ // "col" tail - row is in range but column is out of range.
+ for(int col = 0; col < blank_cols; col++)
+ {
+ *out++ = static_cast<TOut>(0);
+ }
+ }
+ // "row" tail - row is out of range so fill with zeros always.
+ for(int row = 0; row < blank_rows; row++)
+ {
+ for(int col = 0; col < (fill_cols + blank_cols); col++)
+ {
+ *out++ = static_cast<TOut>(0);
+ }
+ }
+ }
+ }
+ }
+
+ template <typename T>
+ static inline void Transform(T *out, const T *const in, const int stride,
+ const int k0, const int kmax, const int x0, const int xmax)
+ {
+ Transform<T, T>(out, in, stride, k0, kmax, x0, xmax);
+ }
+};
+
+/*****************************************************************************/
+template <unsigned int IntBy, unsigned int BlockBy, bool Transposed, typename TOut, typename TIn>
+void Transform(
+ TOut *out, const TIn *const in, const int stride,
+ const int k0, const int kmax, const int x0, const int xmax)
+{
+ // Redirect to a specialised implementation predicated on argument size.
+ TransformImpl<IntBy, BlockBy, Transposed, sizeof(TOut), sizeof(TIn)>::Transform(
+ out, in, stride, k0, kmax, x0, xmax);
+}
+/*****************************************************************************/
+
+#include "transforms/list.hpp"