mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2024-12-15 07:10:37 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merged in rmlarsen/eigen (pull request PR-178)

Browse Source 
Eigen Tensor cost model part 2: Thread scheduling for standard evaluators and reductions.
This commit is contained in:
Benoit Steiner 2016-04-15 14:53:15 -07:00
commit eb669f989f
4 changed files with 99 additions and 97 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h
View File

@ -10,9 +10,9 @@
#ifndef EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H #ifndef EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H
#define EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H #define EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H
#if !defined(EIGEN_USE_GPU) //#if !defined(EIGEN_USE_GPU)
#define EIGEN_USE_COST_MODEL //#define EIGEN_USE_COST_MODEL
#endif //#endif
namespace Eigen { namespace Eigen {

5
unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
View File

@ -189,6 +189,11 @@ struct TensorEvaluator<const Derived, Device>
return loadConstant(m_data+index); return loadConstant(m_data+index);
} }
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized,
internal::unpacket_traits<PacketReturnType>::size);
}
EIGEN_DEVICE_FUNC const Scalar* data() const { return m_data; } EIGEN_DEVICE_FUNC const Scalar* data() const { return m_data; }
protected: protected:

73
unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
View File

@ -59,9 +59,16 @@ class TensorExecutor<Expression, DefaultDevice, true>
{ {
const Index size = array_prod(evaluator.dimensions()); const Index size = array_prod(evaluator.dimensions());
const int PacketSize = unpacket_traits<typename TensorEvaluator<Expression, DefaultDevice>::PacketReturnType>::size; const int PacketSize = unpacket_traits<typename TensorEvaluator<Expression, DefaultDevice>::PacketReturnType>::size;
// Manually unroll this loop since compilers don't do it.
const Index UnrolledSize = (size / (4 * PacketSize)) * 4 * PacketSize;
for (Index i = 0; i < UnrolledSize; i += 4*PacketSize) {
evaluator.evalPacket(i);
evaluator.evalPacket(i+PacketSize);
evaluator.evalPacket(i+2*PacketSize);
evaluator.evalPacket(i+3*PacketSize);
}
const Index VectorizedSize = (size / PacketSize) * PacketSize; const Index VectorizedSize = (size / PacketSize) * PacketSize;
for (Index i = UnrolledSize; i < VectorizedSize; i += PacketSize) {
for (Index i = 0; i < VectorizedSize; i += PacketSize) {
evaluator.evalPacket(i); evaluator.evalPacket(i);
} }
for (Index i = VectorizedSize; i < size; ++i) { for (Index i = VectorizedSize; i < size; ++i) {
@ -78,8 +85,9 @@ class TensorExecutor<Expression, DefaultDevice, true>
#ifdef EIGEN_USE_THREADS #ifdef EIGEN_USE_THREADS
template <typename Evaluator, typename Index, bool Vectorizable> template <typename Evaluator, typename Index, bool Vectorizable>
struct EvalRange { struct EvalRange {
static void run(Evaluator evaluator, const Index first, const Index last) { static void run(Evaluator* evaluator_in, const Index first, const Index last) {
eigen_assert(last > first); Evaluator evaluator = *evaluator_in;
eigen_assert(last >= first);
for (Index i = first; i < last; ++i) { for (Index i = first; i < last; ++i) {
evaluator.evalScalar(i); evaluator.evalScalar(i);
} }
@ -88,19 +96,26 @@ struct EvalRange {
template <typename Evaluator, typename Index> template <typename Evaluator, typename Index>
struct EvalRange<Evaluator, Index, true> { struct EvalRange<Evaluator, Index, true> {
static void run(Evaluator evaluator, const Index first, const Index last) { static void run(Evaluator* evaluator_in, const Index first, const Index last) {
eigen_assert(last > first); Evaluator evaluator = *evaluator_in;
eigen_assert(last >= first);
Index i = first; Index i = first;
static const int PacketSize = unpacket_traits<typename Evaluator::PacketReturnType>::size; const int PacketSize = unpacket_traits<typename Evaluator::PacketReturnType>::size;
if (last - first >= PacketSize) { if (last - first >= PacketSize) {
eigen_assert(first % PacketSize == 0); eigen_assert(first % PacketSize == 0);
Index lastPacket = last - (last % PacketSize); Index last_chunk_offset = last - 4 * PacketSize;
for (; i < lastPacket; i += PacketSize) { // Manually unroll this loop since compilers don't do it.
for (; i <= last_chunk_offset; i += 4*PacketSize) {
evaluator.evalPacket(i);
evaluator.evalPacket(i+PacketSize);
evaluator.evalPacket(i+2*PacketSize);
evaluator.evalPacket(i+3*PacketSize);
}
last_chunk_offset = last - PacketSize;
for (; i <= last_chunk_offset; i += PacketSize) {
evaluator.evalPacket(i); evaluator.evalPacket(i);
} }
} }
for (; i < last; ++i) { for (; i < last; ++i) {
evaluator.evalScalar(i); evaluator.evalScalar(i);
} }
@ -108,8 +123,7 @@ struct EvalRange<Evaluator, Index, true> {
}; };
template <typename Expression, bool Vectorizable> template <typename Expression, bool Vectorizable>
class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable> class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable> {
{
public: public:
typedef typename Expression::Index Index; typedef typename Expression::Index Index;
static inline void run(const Expression& expr, const ThreadPoolDevice& device) static inline void run(const Expression& expr, const ThreadPoolDevice& device)
@ -119,25 +133,35 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable>
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign) if (needs_assign)
{ {
const Index PacketSize = Vectorizable ? unpacket_traits<typename Evaluator::PacketReturnType>::size : 1;
const Index size = array_prod(evaluator.dimensions()); const Index size = array_prod(evaluator.dimensions());
int num_threads = device.numThreads();
static const int PacketSize = Vectorizable ? unpacket_traits<typename Evaluator::PacketReturnType>::size : 1; #ifdef EIGEN_USE_COST_MODEL
if (num_threads > 1) {
int blocksz = std::ceil<int>(static_cast<float>(size)/device.numThreads()) + PacketSize - 1; num_threads = TensorCostModel<ThreadPoolDevice>::numThreads(
size, evaluator.costPerCoeff(Vectorizable), num_threads);
}
#endif
if (num_threads == 1) {
EvalRange<Evaluator, Index, Vectorizable>::run(&evaluator, 0, size);
} else {
Index blocksz = std::ceil<Index>(static_cast<float>(size)/num_threads) + PacketSize - 1;
const Index blocksize = numext::maxi<Index>(PacketSize, (blocksz - (blocksz % PacketSize))); const Index blocksize = numext::maxi<Index>(PacketSize, (blocksz - (blocksz % PacketSize)));
const unsigned int numblocks = static_cast<unsigned int>(size / blocksize); const Index numblocks = size / blocksize;
Barrier barrier(numblocks); Barrier barrier(numblocks);
for (unsigned int i = 0; i < numblocks; ++i) { for (int i = 0; i < numblocks; ++i) {
device.enqueue_with_barrier(&barrier, &EvalRange<Evaluator, Index, Vectorizable>::run, evaluator, i*blocksize, (i+1)*blocksize); device.enqueue_with_barrier(
&barrier, &EvalRange<Evaluator, Index, Vectorizable>::run,
&evaluator, i * blocksize, (i + 1) * blocksize);
} }
if (numblocks * blocksize < size) {
if (static_cast<Index>(numblocks) * blocksize < size) { EvalRange<Evaluator, Index, Vectorizable>::run(
EvalRange<Evaluator, Index, Vectorizable>::run(evaluator, numblocks * blocksize, size); &evaluator, numblocks * blocksize, size);
} }
barrier.Wait(); barrier.Wait();
} }
}
evaluator.cleanup(); evaluator.cleanup();
} }
}; };
@ -226,7 +250,6 @@ inline void TensorExecutor<Expression, GpuDevice, Vectorizable>::run(
#endif // __CUDACC__ #endif // __CUDACC__
#endif // EIGEN_USE_GPU #endif // EIGEN_USE_GPU
} // end namespace internal } // end namespace internal
} // end namespace Eigen } // end namespace Eigen

106
unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
View File

@ -214,7 +214,7 @@ struct FullReducer {
static EIGEN_DEVICE_FUNC void run(const Self& self, Op& reducer, const Device&, typename Self::CoeffReturnType* output) { static EIGEN_DEVICE_FUNC void run(const Self& self, Op& reducer, const Device&, typename Self::CoeffReturnType* output) {
const typename Self::Index num_coeffs = array_prod(self.m_impl.dimensions()); const typename Self::Index num_coeffs = array_prod(self.m_impl.dimensions());
*output = InnerMostDimReducer<Self, Op>::reduce(self, 0, num_coeffs, reducer); *output = InnerMostDimReducer<Self, Op, Vectorizable>::reduce(self, 0, num_coeffs, reducer);
} }
}; };
@ -222,18 +222,19 @@ struct FullReducer {
#ifdef EIGEN_USE_THREADS #ifdef EIGEN_USE_THREADS
// Multithreaded full reducers // Multithreaded full reducers
template <typename Self, typename Op, template <typename Self, typename Op,
bool vectorizable = (Self::InputPacketAccess & Op::PacketAccess)> bool Vectorizable = (Self::InputPacketAccess & Op::PacketAccess)>
struct FullReducerShard { struct FullReducerShard {
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(const Self& self, typename Self::Index firstIndex, static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(const Self& self, typename Self::Index firstIndex,
typename Self::Index numValuesToReduce, Op& reducer, typename Self::Index numValuesToReduce, Op& reducer,
typename Self::CoeffReturnType* output) { typename Self::CoeffReturnType* output) {
*output = InnerMostDimReducer<Self, Op, vectorizable>::reduce( *output = InnerMostDimReducer<Self, Op, Vectorizable>::reduce(
self, firstIndex, numValuesToReduce, reducer); self, firstIndex, numValuesToReduce, reducer);
} }
}; };
template <typename Self, typename Op> // Multithreaded full reducer
struct FullReducer<Self, Op, ThreadPoolDevice, false> { template <typename Self, typename Op, bool Vectorizable>
struct FullReducer<Self, Op, ThreadPoolDevice, Vectorizable> {
static const bool HasOptimizedImplementation = !Op::IsStateful; static const bool HasOptimizedImplementation = !Op::IsStateful;
static const int PacketSize = static const int PacketSize =
unpacket_traits<typename Self::PacketReturnType>::size; unpacket_traits<typename Self::PacketReturnType>::size;
@ -247,79 +248,44 @@ struct FullReducer<Self, Op, ThreadPoolDevice, false> {
*output = reducer.finalize(reducer.initialize()); *output = reducer.finalize(reducer.initialize());
return; return;
} }
const std::size_t num_threads = device.numThreads(); #ifdef EIGEN_USE_COST_MODEL
const TensorOpCost cost =
self.m_impl.costPerCoeff(Vectorizable) +
TensorOpCost(0, 0, internal::functor_traits<Op>::Cost, Vectorizable,
PacketSize);
const int num_threads = TensorCostModel<ThreadPoolDevice>::numThreads(
num_coeffs, cost, device.numThreads());
#else
const int num_threads = device.numThreads();
#endif
if (num_threads == 1) { if (num_threads == 1) {
*output = InnerMostDimReducer<Self, Op, false>::reduce(self, 0, num_coeffs, reducer); *output =
InnerMostDimReducer<Self, Op, Vectorizable>::reduce(self, 0, num_coeffs, reducer);
return; return;
} else { }
const Index blocksize = std::floor<Index>(static_cast<float>(num_coeffs) / num_threads); const Index blocksize =
const unsigned int numblocks = blocksize > 0 ? static_cast<unsigned int>(num_coeffs / blocksize) : 0; std::floor<Index>(static_cast<float>(num_coeffs) / num_threads);
eigen_assert(num_coeffs >= static_cast<Index>(numblocks) * blocksize); const Index numblocks = blocksize > 0 ? num_coeffs / blocksize : 0;
eigen_assert(num_coeffs >= numblocks * blocksize);
Barrier barrier(numblocks); Barrier barrier(numblocks);
MaxSizeVector<typename Self::CoeffReturnType> shards(numblocks, reducer.initialize()); MaxSizeVector<typename Self::CoeffReturnType> shards(numblocks, reducer.initialize());
for (unsigned int i = 0; i < numblocks; ++i) { for (Index i = 0; i < numblocks; ++i) {
device.enqueue_with_barrier(&barrier, &FullReducerShard<Self, Op, false>::run, self, device.enqueue_with_barrier(&barrier, &FullReducerShard<Self, Op, Vectorizable>::run,
i * blocksize, blocksize, reducer, &shards[i]);
}
typename Self::CoeffReturnType finalShard;
if (static_cast<Index>(numblocks) * blocksize < num_coeffs) {
finalShard = InnerMostDimReducer<Self, Op, false>::reduce(
self, numblocks * blocksize, num_coeffs - numblocks * blocksize, reducer);
} else {
finalShard = reducer.initialize();
}
barrier.Wait();
for (unsigned int i = 0; i < numblocks; ++i) {
reducer.reduce(shards[i], &finalShard);
}
*output = reducer.finalize(finalShard);
}
}
};
template <typename Self, typename Op>
struct FullReducer<Self, Op, ThreadPoolDevice, true> {
static const bool HasOptimizedImplementation = !Op::IsStateful;
static const int PacketSize =
unpacket_traits<typename Self::PacketReturnType>::size;
// launch one reducer per thread and accumulate the result.
static void run(const Self& self, Op& reducer, const ThreadPoolDevice& device,
typename Self::CoeffReturnType* output) {
typedef typename Self::Index Index;
const Index num_coeffs = array_prod(self.m_impl.dimensions());
if (num_coeffs == 0) {
*output = reducer.finalize(reducer.initialize());
return;
}
const std::size_t num_threads = device.numThreads();
if (num_threads == 1) {
*output = InnerMostDimReducer<Self, Op, true>::reduce(self, 0, num_coeffs, reducer);
return;
}
const Index blocksize = std::floor<Index>(static_cast<float>(num_coeffs) / num_threads);
const unsigned int numblocks = blocksize > 0 ? static_cast<unsigned int>(num_coeffs / blocksize) : 0;
eigen_assert(num_coeffs >= static_cast<Index>(numblocks) * blocksize);
Barrier barrier(numblocks);
MaxSizeVector<typename Self::CoeffReturnType> shards(numblocks, reducer.initialize());
for (unsigned int i = 0; i < numblocks; ++i) {
device.enqueue_with_barrier(&barrier, &FullReducerShard<Self, Op, true>::run,
self, i * blocksize, blocksize, reducer, self, i * blocksize, blocksize, reducer,
&shards[i]); &shards[i]);
} }
typename Self::CoeffReturnType finalShard; typename Self::CoeffReturnType finalShard;
if (static_cast<Index>(numblocks) * blocksize < num_coeffs) { if (numblocks * blocksize < num_coeffs) {
finalShard = InnerMostDimReducer<Self, Op, true>::reduce( finalShard = InnerMostDimReducer<Self, Op, Vectorizable>::reduce(
self, numblocks * blocksize, num_coeffs - numblocks * blocksize, reducer); self, numblocks * blocksize, num_coeffs - numblocks * blocksize,
reducer);
} else { } else {
finalShard = reducer.initialize(); finalShard = reducer.initialize();
} }
barrier.Wait(); barrier.Wait();
for (unsigned int i = 0; i < numblocks; ++i) {
for (Index i = 0; i < numblocks; ++i) {
reducer.reduce(shards[i], &finalShard); reducer.reduce(shards[i], &finalShard);
} }
*output = reducer.finalize(finalShard); *output = reducer.finalize(finalShard);
@ -498,13 +464,21 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
static bool size_large_enough(Index total_size) {
#ifndef EIGEN_USE_COST_MODEL
return total_size > 1024 * 1024;
#else
return true || total_size;
#endif
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool evalSubExprsIfNeeded(CoeffReturnType* data) { EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool evalSubExprsIfNeeded(CoeffReturnType* data) {
m_impl.evalSubExprsIfNeeded(NULL); m_impl.evalSubExprsIfNeeded(NULL);
// Use the FullReducer if possible. // Use the FullReducer if possible.
if (RunningFullReduction && internal::FullReducer<Self, Op, Device>::HasOptimizedImplementation && if (RunningFullReduction && internal::FullReducer<Self, Op, Device>::HasOptimizedImplementation &&
((RunningOnGPU && (m_device.majorDeviceVersion() >= 3)) || ((RunningOnGPU && (m_device.majorDeviceVersion() >= 3)) ||
(!RunningOnGPU && (internal::array_prod(m_impl.dimensions()) > 1024 * 1024)))) { (!RunningOnGPU && size_large_enough(internal::array_prod(m_impl.dimensions()))))) {
bool need_assign = false; bool need_assign = false;
if (!data) { if (!data) {