// This file is part of Eigen, a lightweight C++ template library // for linear algebra. // // Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com> // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
: m_impl(op.expression(), device), m_padding(op.padding()), m_paddingValue(op.padding_value()), m_device(device)
{ // The padding op doesn't change the rank of the tensor. Directly padding a scalar would lead // to a vector, which doesn't make sense. Instead one should reshape the scalar into a vector // of 1 element first and then pad.
EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
// Compute offsets in the output tensor corresponding to the desc.offset().
DSizes<Index, NumDims> output_offsets; for (int i = NumDims - 1; i > 0; --i) { constint dim = IsColMajor ? i : NumDims - i - 1; constint stride_dim = IsColMajor ? dim : dim + 1;
output_offsets[dim] = offset / m_outputStrides[stride_dim];
offset -= output_offsets[dim] * m_outputStrides[stride_dim];
}
output_offsets[inner_dim_idx] = offset;
// Offsets in the input corresponding to output offsets.
DSizes<Index, NumDims> input_offsets = output_offsets; for (int i = 0; i < NumDims; ++i) { constint dim = IsColMajor ? i : NumDims - i - 1;
input_offsets[dim] = input_offsets[dim] - m_padding[dim].first;
}
// Compute offset in the input buffer (at this point it might be illegal and // point outside of the input buffer, because we don't check for negative // offsets, it will be autocorrected in the block iteration loop below).
Index input_offset = 0; for (int i = 0; i < NumDims; ++i) { constint dim = IsColMajor ? i : NumDims - i - 1;
input_offset += input_offsets[dim] * m_inputStrides[dim];
}
// Destination buffer and scratch buffer both indexed from 0 and have the // same dimensions as the requested block (for destination buffer this // property is guaranteed by `desc.destination()`).
Index output_offset = 0; const DSizes<Index, NumDims> output_strides =
internal::strides<Layout>(desc.dimensions());
// NOTE(ezhulenev): We initialize bock iteration state for `NumDims - 1` // dimensions, skipping innermost dimension. In theory it should be possible // to squeeze matching innermost dimensions, however in practice that did // not show any improvements in benchmarks. Also in practice first outer // dimension usually has padding, and will prevent squeezing.
// Initialize output block iterator state. Dimension in this array are // always in inner_most -> outer_most order (col major layout).
array<BlockIteratorState, NumDims - 1> it; for (int i = 0; i < NumDims - 1; ++i) { constint dim = IsColMajor ? i + 1 : NumDims - i - 2;
it[i].count = 0;
it[i].size = desc.dimension(dim);
const Index input_inner_dim_size = static_cast<Index>(m_impl.dimensions()[inner_dim_idx]);
// Total output size. const Index output_size = desc.size();
// We will fill inner dimension of this size in the output. It might be // larger than the inner dimension in the input, so we might have to pad // before/after we copy values from the input inner dimension. const Index output_inner_dim_size = desc.dimension(inner_dim_idx);
// How many values to fill with padding BEFORE reading from the input inner // dimension. const Index output_inner_pad_before_size =
input_offsets[inner_dim_idx] < 0
? numext::mini(numext::abs(input_offsets[inner_dim_idx]),
output_inner_dim_size)
: 0;
// How many values we can actually copy from the input inner dimension. const Index output_inner_copy_size = numext::mini( // Want to copy from input.
(output_inner_dim_size - output_inner_pad_before_size), // Can copy from input.
numext::maxi(input_inner_dim_size - (input_offsets[inner_dim_idx] +
output_inner_pad_before_size),
Index(0)));
eigen_assert(output_inner_copy_size >= 0);
// How many values to fill with padding AFTER reading from the input inner // dimension. const Index output_inner_pad_after_size =
(output_inner_dim_size - output_inner_copy_size -
output_inner_pad_before_size);
// Sanity check, sum of all sizes must be equal to the output size.
eigen_assert(output_inner_dim_size ==
(output_inner_pad_before_size + output_inner_copy_size +
output_inner_pad_after_size));
// Keep track of current coordinates and padding in the output.
DSizes<Index, NumDims> output_coord = output_offsets;
DSizes<Index, NumDims> output_padded; for (int i = 0; i < NumDims; ++i) { constint dim = IsColMajor ? i : NumDims - i - 1;
output_padded[dim] = isPaddingAtIndexForDim(output_coord[dim], dim);
}
// Prepare storage for the materialized padding result. consttypename TensorBlock::Storage block_storage =
TensorBlock::prepareStorage(desc, scratch);
// TODO(ezhulenev): Squeeze multiple non-padded inner dimensions into a // single logical inner dimension.
// When possible we squeeze writes for the innermost (only if non-padded) // dimension with the first padded dimension. This allows to reduce the // number of calls to LinCopy and better utilize vector instructions. constbool squeeze_writes =
NumDims > 1 && // inner dimension is not padded
(input_inner_dim_size == m_dimensions[inner_dim_idx]) && // and equal to the block inner dimension
(input_inner_dim_size == output_inner_dim_size);
// Maximum coordinate on a squeeze dimension that we can write to. const Index squeeze_max_coord =
squeeze_writes ? numext::mini( // max non-padded element in the input static_cast<Index>(m_dimensions[squeeze_dim] -
m_padding[squeeze_dim].second), // max element in the output buffer static_cast<Index>(output_offsets[squeeze_dim] +
desc.dimension(squeeze_dim)))
: static_cast<Index>(0);
// Iterate copying data from `m_impl.data()` to the output buffer. for (Index size = 0; size < output_size;) { // Detect if we are in the padded region (exclude innermost dimension). bool is_padded = false; for (int j = 1; j < NumDims; ++j) { constint dim = IsColMajor ? j : NumDims - j - 1;
is_padded = output_padded[dim]; if (is_padded) break;
}
if (is_padded) { // Fill single innermost dimension with padding value.
size += output_inner_dim_size;
// Update iteration state for only `squeeze_num - 1` processed inner // dimensions, because we have another iteration state update at the end // of the loop that will update iteration state for the last inner // processed dimension.
it[0].count += (squeeze_num - 1);
input_offset += it[0].input_stride * (squeeze_num - 1);
output_offset += it[0].output_stride * (squeeze_num - 1);
output_coord[squeeze_dim] += (squeeze_num - 1);
} else { // Single read from innermost dimension.
size += output_inner_dim_size;
{ // Fill with padding before copying from input inner dimension. const Index out = output_offset;
{ // Copy data from input inner dimension. const Index out = output_offset + output_inner_pad_before_size; const Index in = input_offset + output_inner_pad_before_size;
{ // Fill with padding after copying from input inner dimension. const Index out = output_offset + output_inner_pad_before_size +
output_inner_copy_size;
const Index initialIndex = index;
Index inputIndex = 0;
EIGEN_UNROLL_LOOP for (int i = NumDims - 1; i > 0; --i) { const Index firstIdx = index; const Index lastIdx = index + PacketSize - 1; const Index lastPaddedLeft = m_padding[i].first * m_outputStrides[i]; const Index firstPaddedRight = (m_dimensions[i] - m_padding[i].second) * m_outputStrides[i]; const Index lastPaddedRight = m_outputStrides[i+1];
if (!isLeftPaddingCompileTimeZero(i) && lastIdx < lastPaddedLeft) { // all the coefficient are in the padding zone. return internal::pset1<PacketReturnType>(m_paddingValue);
} elseif (!isRightPaddingCompileTimeZero(i) && firstIdx >= firstPaddedRight && lastIdx < lastPaddedRight) { // all the coefficient are in the padding zone. return internal::pset1<PacketReturnType>(m_paddingValue);
} elseif ((isLeftPaddingCompileTimeZero(i) && isRightPaddingCompileTimeZero(i)) || (firstIdx >= lastPaddedLeft && lastIdx < firstPaddedRight)) { // all the coefficient are between the 2 padding zones. const Index idx = index / m_outputStrides[i];
inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
index -= idx * m_outputStrides[i];
} else { // Every other case return packetWithPossibleZero(initialIndex);
}
}
const Index lastIdx = index + PacketSize - 1; const Index firstIdx = index; const Index lastPaddedLeft = m_padding[0].first; const Index firstPaddedRight = (m_dimensions[0] - m_padding[0].second); const Index lastPaddedRight = m_outputStrides[1];
if (!isLeftPaddingCompileTimeZero(0) && lastIdx < lastPaddedLeft) { // all the coefficient are in the padding zone. return internal::pset1<PacketReturnType>(m_paddingValue);
} elseif (!isRightPaddingCompileTimeZero(0) && firstIdx >= firstPaddedRight && lastIdx < lastPaddedRight) { // all the coefficient are in the padding zone. return internal::pset1<PacketReturnType>(m_paddingValue);
} elseif ((isLeftPaddingCompileTimeZero(0) && isRightPaddingCompileTimeZero(0)) || (firstIdx >= lastPaddedLeft && lastIdx < firstPaddedRight)) { // all the coefficient are between the 2 padding zones.
inputIndex += (index - m_padding[0].first); return m_impl.template packet<Unaligned>(inputIndex);
} // Every other case return packetWithPossibleZero(initialIndex);
}
const Index initialIndex = index;
Index inputIndex = 0;
EIGEN_UNROLL_LOOP for (int i = 0; i < NumDims - 1; ++i) { const Index firstIdx = index; const Index lastIdx = index + PacketSize - 1; const Index lastPaddedLeft = m_padding[i].first * m_outputStrides[i+1]; const Index firstPaddedRight = (m_dimensions[i] - m_padding[i].second) * m_outputStrides[i+1]; const Index lastPaddedRight = m_outputStrides[i];
if (!isLeftPaddingCompileTimeZero(i) && lastIdx < lastPaddedLeft) { // all the coefficient are in the padding zone. return internal::pset1<PacketReturnType>(m_paddingValue);
} elseif (!isRightPaddingCompileTimeZero(i) && firstIdx >= firstPaddedRight && lastIdx < lastPaddedRight) { // all the coefficient are in the padding zone. return internal::pset1<PacketReturnType>(m_paddingValue);
} elseif ((isLeftPaddingCompileTimeZero(i) && isRightPaddingCompileTimeZero(i)) || (firstIdx >= lastPaddedLeft && lastIdx < firstPaddedRight)) { // all the coefficient are between the 2 padding zones. const Index idx = index / m_outputStrides[i+1];
inputIndex += (idx - m_padding[i].first) * m_inputStrides[i];
index -= idx * m_outputStrides[i+1];
} else { // Every other case return packetWithPossibleZero(initialIndex);
}
}
const Index lastIdx = index + PacketSize - 1; const Index firstIdx = index; const Index lastPaddedLeft = m_padding[NumDims-1].first; const Index firstPaddedRight = (m_dimensions[NumDims-1] - m_padding[NumDims-1].second); const Index lastPaddedRight = m_outputStrides[NumDims-1];
if (!isLeftPaddingCompileTimeZero(NumDims-1) && lastIdx < lastPaddedLeft) { // all the coefficient are in the padding zone. return internal::pset1<PacketReturnType>(m_paddingValue);
} elseif (!isRightPaddingCompileTimeZero(NumDims-1) && firstIdx >= firstPaddedRight && lastIdx < lastPaddedRight) { // all the coefficient are in the padding zone. return internal::pset1<PacketReturnType>(m_paddingValue);
} elseif ((isLeftPaddingCompileTimeZero(NumDims-1) && isRightPaddingCompileTimeZero(NumDims-1)) || (firstIdx >= lastPaddedLeft && lastIdx < firstPaddedRight)) { // all the coefficient are between the 2 padding zones.
inputIndex += (index - m_padding[NumDims-1].first); return m_impl.template packet<Unaligned>(inputIndex);
} // Every other case return packetWithPossibleZero(initialIndex);
}
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.