devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

merging updates from upstream

Browse Source
This commit is contained in:
Deven Desai
2018-07-11 09:17:33 -04:00
parent e2b2c61533 f00d08cc0a
commit 38807a2575
55 changed files with 983 additions and 340 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
View File

@@ -161,6 +161,22 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
}
}
}
// Handle special format like NCHW, its input shape is '[1, N..., 1]' and
// broadcast shape is '[N, 1..., N]'
if (!oneByN && !nByOne) {
if (input_dims[0] == 1 && input_dims[NumDims-1] == 1 && NumDims > 2) {
nByOne = true;
oneByN = true;
for (int i = 1; i < NumDims-1; ++i) {
if (broadcast[i] != 1) {
nByOne = false;
oneByN = false;
break;
}
}
}
}
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -256,24 +272,70 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
}
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
if (oneByN) {
if (oneByN && !nByOne) {
return packetNByOne<LoadMode>(index);
} else if (nByOne) {
} else if (!oneByN && nByOne) {
return packetOneByN<LoadMode>(index);
} else if (oneByN && nByOne) {
return packetOneByNByOne<LoadMode>(index);
} else {
return packetColMajor<LoadMode>(index);
}
} else {
if (oneByN) {
if (oneByN && !nByOne) {
return packetOneByN<LoadMode>(index);
} else if (nByOne) {
} else if (!oneByN && nByOne) {
return packetNByOne<LoadMode>(index);
} else if (oneByN && nByOne) {
return packetOneByNByOne<LoadMode>(index);
} else {
return packetRowMajor<LoadMode>(index);
}
}
}
template<int LoadMode>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetOneByNByOne
(Index index) const
{
EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
Index startDim, endDim;
Index inputIndex, outputOffset, batchedIndex;
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
startDim = NumDims - 1;
endDim = 1;
} else {
startDim = 0;
endDim = NumDims - 2;
}
batchedIndex = index % m_outputStrides[startDim];
inputIndex = batchedIndex / m_outputStrides[endDim];
outputOffset = batchedIndex % m_outputStrides[endDim];
if (outputOffset + PacketSize <= m_outputStrides[endDim]) {
values[0] = m_impl.coeff(inputIndex);
return internal::pload1<PacketReturnType>(values);
} else {
for (int i = 0, cur = 0; i < PacketSize; ++i, ++cur) {
if (outputOffset + cur < m_outputStrides[endDim]) {
values[i] = m_impl.coeff(inputIndex);
} else {
++inputIndex;
inputIndex = (inputIndex == m_inputStrides[startDim] ? 0 : inputIndex);
values[i] = m_impl.coeff(inputIndex);
outputOffset = 0;
cur = 0;
}
}
return internal::pload<PacketReturnType>(values);
}
}
template<int LoadMode>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetOneByN(Index index) const
{

11
unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
View File

@@ -189,9 +189,11 @@ struct ThreadPoolDevice {
// of blocks to be evenly dividable across threads.
double block_size_f = 1.0 / CostModel::taskSize(1, cost);
Index block_size = numext::mini(n, numext::maxi<Index>(1, block_size_f));
const Index max_block_size =
numext::mini(n, numext::maxi<Index>(1, 2 * block_size_f));
const Index max_oversharding_factor = 4;
Index block_size = numext::mini(
n, numext::maxi<Index>(divup<Index>(n, max_oversharding_factor * numThreads()),
block_size_f));
const Index max_block_size = numext::mini(n, 2 * block_size);
if (block_align) {
Index new_block_size = block_align(block_size);
eigen_assert(new_block_size >= block_size);
@@ -205,7 +207,8 @@ struct ThreadPoolDevice {
(divup<int>(block_count, numThreads()) * numThreads());
// Now try to increase block size up to max_block_size as long as it
// doesn't decrease parallel efficiency.
for (Index prev_block_count = block_count; prev_block_count > 1;) {
for (Index prev_block_count = block_count;
max_efficiency < 1.0 && prev_block_count > 1;) {
// This is the next block size that divides size into a smaller number
// of blocks than the current block_size.
Index coarser_block_size = divup(n, prev_block_count - 1);

4
unsupported/Eigen/src/FFT/ei_kissfft_impl.h
View File

@@ -316,8 +316,8 @@ struct kissfft_impl
// use optimized mode for even real
fwd( dst, reinterpret_cast<const Complex*> (src), ncfft);
Complex dc = dst[0].real() + dst[0].imag();
Complex nyquist = dst[0].real() - dst[0].imag();
Complex dc(dst[0].real() + dst[0].imag());
Complex nyquist(dst[0].real() - dst[0].imag());
int k;
for ( k=1;k <= ncfft2 ; ++k ) {
Complex fpk = dst[k];