GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-22 04:27:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

Reworked the TensorExecutor code to support in place evaluation.

Browse Source
This commit is contained in:
Benoit Steiner 2014-08-13 08:22:05 -07:00
parent 647622281e
commit 439feca139
1 changed files with 58 additions and 48 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
View File

@ -11,7 +11,7 @@
#define EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H
#ifdef EIGEN_USE_THREADS
#include <future>
#include <future>"
#endif
namespace Eigen {
@ -28,34 +28,38 @@ namespace internal {
// Default strategy: the expression is evaluated with a single cpu thread.
template<typename Expression, typename Device = DefaultDevice, bool Vectorizable = TensorEvaluator<Expression, Device>::PacketAccess>
struct TensorExecutor
class TensorExecutor
{
public:
typedef typename Expression::Index Index;
EIGEN_DEVICE_FUNC
static inline void run(const Expression& expr, const Device& device = Device())
{
TensorEvaluator<Expression, Device> evaluator(expr, device);
evaluator.evalSubExprsIfNeeded();
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign)
{
const Index size = evaluator.dimensions().TotalSize();
for (Index i = 0; i < size; ++i) {
evaluator.evalScalar(i);
}
}
evaluator.cleanup();
}
};
template<typename Expression>
struct TensorExecutor<Expression, DefaultDevice, true>
class TensorExecutor<Expression, DefaultDevice, true>
{
public:
typedef typename Expression::Index Index;
static inline void run(const Expression& expr, const DefaultDevice& device = DefaultDevice())
{
TensorEvaluator<Expression, DefaultDevice> evaluator(expr, device);
evaluator.evalSubExprsIfNeeded();
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign)
{
const Index size = evaluator.dimensions().TotalSize();
static const int PacketSize = unpacket_traits<typename TensorEvaluator<Expression, DefaultDevice>::PacketReturnType>::size;
const int VectorizedSize = (size / PacketSize) * PacketSize;
@ -66,7 +70,7 @@ struct TensorExecutor<Expression, DefaultDevice, true>
for (Index i = VectorizedSize; i < size; ++i) {
evaluator.evalScalar(i);
}
}
evaluator.cleanup();
}
};
@ -107,15 +111,17 @@ struct EvalRange<Evaluator, Index, true> {
};
template<typename Expression, bool Vectorizable>
struct TensorExecutor<Expression, ThreadPoolDevice, Vectorizable>
class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable>
{
public:
typedef typename Expression::Index Index;
static inline void run(const Expression& expr, const ThreadPoolDevice& device)
{
typedef TensorEvaluator<Expression, ThreadPoolDevice> Evaluator;
Evaluator evaluator(expr, device);
evaluator.evalSubExprsIfNeeded();
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign)
{
const Index size = evaluator.dimensions().TotalSize();
static const int PacketSize = Vectorizable ? unpacket_traits<typename Evaluator::PacketReturnType>::size : 1;
@ -138,7 +144,7 @@ struct TensorExecutor<Expression, ThreadPoolDevice, Vectorizable>
if (numblocks * blocksize < size) {
EvalRange<Evaluator, Index>::run(&evaluator, numblocks * blocksize, size);
}
}
evaluator.cleanup();
}
};
@ -157,19 +163,23 @@ __global__ void EigenMetaKernel(Evaluator eval, unsigned int size) {
}
template<typename Expression, bool Vectorizable>
struct TensorExecutor<Expression, GpuDevice, Vectorizable>
class TensorExecutor<Expression, GpuDevice, Vectorizable>
{
public:
typedef typename Expression::Index Index;
static inline void run(const Expression& expr, const GpuDevice& device)
{
TensorEvaluator<Expression, GpuDevice> evaluator(expr, device);
evaluator.evalSubExprsIfNeeded();
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign)
{
const int num_blocks = getNumCudaMultiProcessors() * maxCudaThreadsPerMultiProcessor() / maxCudaThreadsPerBlock();
const int block_size = maxCudaThreadsPerBlock();
const Index size = evaluator.dimensions().TotalSize();
EigenMetaKernel<TensorEvaluator<Expression, GpuDevice> > <<<num_blocks, block_size, 0, device.stream()>>>(evaluator, size);
eigen_assert(cudaGetLastError() == cudaSuccess);
assert(cudaGetLastError() == cudaSuccess);
}
evaluator.cleanup();
}
};