2 files changed, 220 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/convolution/common/padding.cpp b/src/core/NEON/kernels/convolution/common/padding.cpp
new file mode 100644
index 0000000000..b50067b4e0
--- /dev/null
+++ b/src/core/NEON/kernels/convolution/common/padding.cpp
@@ -0,0 +1,130 @@
+/*
+ * 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.
+ */
+#include <cstring>
+#include <cstdint>
+
+#include "arm_compute/core/NEON/kernels/convolution/common/arm.hpp"
+#include "arm_compute/core/NEON/kernels/convolution/common/padding.hpp"
+
+namespace padding
+{
+
+template <typename T>
+void copy_and_pad_tile(
+  const unsigned int tile_rows,
+  const unsigned int tile_cols,
+  const unsigned int n_channels,
+  const T* const inptr,
+  const unsigned int in_row_stride,
+  const unsigned int in_col_stride,
+  T* const outptr,
+  const unsigned int out_row_stride,
+  const unsigned int out_col_stride,
+  const unsigned int pad_top,
+  const unsigned int pad_left,
+  const unsigned int pad_bottom,
+  const unsigned int pad_right,
+  const T pad_value
+)
+{
+  for (unsigned int out_i = 0; out_i < tile_rows; out_i++)
+  {
+    for (unsigned int out_j = 0; out_j < tile_cols; out_j++)
+    {
+      T* const output = outptr + out_i*out_row_stride + out_j*out_col_stride;
+
+      if (out_i < pad_top || tile_rows - pad_bottom <= out_i ||
+          out_j < pad_left || tile_cols - pad_right <= out_j)
+      {
+        for (unsigned int n = 0; n < n_channels; n++)
+        {
+          output[n] = pad_value;
+        }
+      }
+      else
+      {
+        const auto in_i = out_i - pad_top, in_j = out_j - pad_left;
+        const T* const input = inptr + in_i*in_row_stride + in_j*in_col_stride;
+        std::memcpy(output, input, n_channels * sizeof(T));
+      }
+    }
+  }
+}
+
+template void copy_and_pad_tile(
+  unsigned int, unsigned int, unsigned int,
+  const uint8_t *, unsigned int, unsigned int,
+  uint8_t *, unsigned int, unsigned int,
+  unsigned int, unsigned int, unsigned int, unsigned int, uint8_t
+);
+
+template void copy_and_pad_tile(
+  unsigned int, unsigned int, unsigned int,
+  const float *, unsigned int, unsigned int,
+  float *, unsigned int, unsigned int,
+  unsigned int, unsigned int, unsigned int, unsigned int, float
+);
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+template void copy_and_pad_tile(
+  unsigned int, unsigned int, unsigned int,
+  const float16_t *, unsigned int, unsigned int,
+  float16_t *, unsigned int, unsigned int,
+  unsigned int, unsigned int, unsigned int, unsigned int, float16_t
+);
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+template <unsigned int TileRows, unsigned int TileCols>
+void CopyCropped<TileRows, TileCols>::execute(
+  const size_t size,
+  const void * const inptr,
+  const size_t in_row_stride,
+  const size_t in_col_stride,
+  void * const outptr,
+  const size_t out_row_stride,
+  const size_t out_col_stride,
+  const unsigned int pad_top,
+  const unsigned int pad_left,
+  const unsigned int pad_bottom,
+  const unsigned int pad_right
+)
+{
+  for (unsigned int out_i = 0, in_i = pad_top; in_i < TileRows - pad_bottom; out_i++, in_i++)
+  {
+    for (unsigned int out_j = 0, in_j = pad_left; in_j < TileCols - pad_right; out_j++, in_j++)
+    {
+      std::memcpy(
+        static_cast<uint8_t *>(outptr) + out_i*out_row_stride + out_j*out_col_stride,
+        static_cast<const uint8_t *>(inptr) + in_i*in_row_stride + in_j*in_col_stride,
+        size
+      );
+    }
+  }
+}
+
+template class CopyCropped<2, 2>;
+template class CopyCropped<3, 3>;
+template class CopyCropped<4, 4>;
+
+}  // namespace padding
diff --git a/src/core/NEON/kernels/convolution/common/qasymm8.cpp b/src/core/NEON/kernels/convolution/common/qasymm8.cpp
new file mode 100644
index 0000000000..1de9ebf28a
--- /dev/null
+++ b/src/core/NEON/kernels/convolution/common/qasymm8.cpp
@@ -0,0 +1,90 @@
+/*
+ * 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.
+ */
+
+#include <algorithm>
+#include <cassert>
+#include <cstdint>
+#include <cmath>
+#include <limits>
+
+#include "arm_compute/core/NEON/kernels/convolution/common/qasymm8.hpp"
+
+namespace qasymm8
+{
+#if(__ANDROID__ || BARE_METAL)
+template <typename T> T round(T val) {  return ::round(val); }
+template <typename T> T exp2(T val) { return ::exp2(val); }
+template <typename T> T log2(T val) { return ::log2(val); }
+#else  /* (__ANDROID__ || BARE_METAL) */
+template <typename T> T round(T val) { return std::round(val); }
+template <typename T> T exp2(T val) { return std::exp2(val); }
+template <typename T> T log2(T val) { return std::log2(val); }
+#endif  /* (__ANDROID__ || BARE_METAL) */
+
+uint8_t QAsymm8Params::quantize(const float value) const
+{
+  const float transformed = value / scale + offset;
+  return static_cast<uint8_t>(round(std::max(0.0f, std::min(255.0f, transformed))));
+}
+
+float QAsymm8Params::dequantize(const uint8_t value) const
+{
+  return scale * (static_cast<float>(value) - offset);
+}
+
+QAsymm8RescaleParams QAsymm8RescaleParams::make_rescale_params(
+  const QAsymm8Params& weight_quant,
+  const QAsymm8Params& input_quant,
+  const QAsymm8Params& output_quant
+)
+{
+  // Based on the gemmlowp approach: https://github.com/google/gemmlowp/blob/master/doc/quantization_example.cc
+  const float rescale = weight_quant.scale * input_quant.scale / output_quant.scale;
+  const float shiftf = round(log2(0.5f / rescale));
+  const float multf = exp2(31.0f + shiftf)*rescale;
+
+  int64_t shift = static_cast<int64_t>(shiftf);
+  int64_t mult = static_cast<int64_t>(multf);
+
+  if (mult == (1ll << 31))
+  {
+    mult /= 2;
+    shift--;
+  }
+
+  assert(shift >= 0);
+  assert(mult <= std::numeric_limits<int32_t>::max());
+
+  return QAsymm8RescaleParams(
+    static_cast<int32_t>(shift),
+    static_cast<int32_t>(mult),
+    rescale
+  );
+}
+
+QAsymm8RescaleParams::QAsymm8RescaleParams(int32_t shift, int32_t multi, float rescale)
+  : shift(shift), multiplier(multi), rescale(rescale)
+{
+}
+}