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

Made the blocking computation aware of the l3 cache

Browse Source 
Also optimized the blocking parameters to take into account the number of threads used for a computation
This commit is contained in:
Benoit Steiner
2014-10-15 15:32:59 -07:00
parent dba55041ab
commit bfdd9f3ac9
13 changed files with 117 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
unsupported/Eigen/CXX11/Tensor
View File

@@ -55,7 +55,7 @@
#include "unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h"
#include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h"
#include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h"
//#include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h"
#include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h"
#include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h"
#include "unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h"
#include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h"

2
unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
View File

@@ -766,7 +766,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
typedef typename internal::gemm_blocking_space<ColMajor, LhsScalar, RhsScalar, Dynamic, Dynamic, Dynamic> BlockingType;
// Sizes of the blocks to load in cache. See the Goto paper for details.
BlockingType blocking(m, n, k, true);
BlockingType blocking(m, n, k, 1, true);
const Index kc = blocking.kc();
const Index mc = (std::min)(m, blocking.mc());
const Index nc = (std::min)(n, blocking.nc());

13
unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
View File

@@ -152,7 +152,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
Index mc = m;
Index nc = n;
Index kc = k;
internal::computeProductBlockingSizes<LhsScalar,RhsScalar,1>(kc, mc, nc/*, num_threads*/);
internal::computeProductBlockingSizes<LhsScalar,RhsScalar,1>(kc, mc, nc, num_threads);
eigen_assert(mc <= m);
eigen_assert(nc <= n);
eigen_assert(kc <= k);
@@ -197,9 +197,10 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
// this should really be numBlockAs * n_blocks;
const Index num_kernel_promises = num_threads * n_blocks;
Promise p;
p.set_value();
std::vector<Promise> kernel_promises(num_kernel_promises, p);
std::vector<Promise> kernel_promises(num_kernel_promises);
for (int i = 0; i < kernel_promises.size(); ++i) {
kernel_promises[i].set_value();
}
for (Index k_block_idx = 0; k_block_idx < k_blocks; k_block_idx++) {
const Index k_start = k_block_idx * kc;
@@ -275,8 +276,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
need_to_pack, // need_to_pack
};
typedef decltype(Self::packRhsAndKernel<packRKArg, RhsPacker, GebpKernel>) Func;
this->m_device.enqueueNoFuture<Func, packRKArg>(&Self::packRhsAndKernel<packRKArg, RhsPacker, GebpKernel>, arg);
this->m_device.enqueueNoFuture(&Self::packRhsAndKernel<packRKArg, RhsPacker, GebpKernel>, arg);
}
}
}
@@ -338,7 +338,6 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
actual_mc, arg.kc, arg.nc, 1.0, -1, -1, 0, 0);
const Index set_idx = blockAId * arg.n_blocks + arg.n_block_idx;
eigen_assert(!(*arg.kernel_promises)[set_idx].ready());
(*arg.kernel_promises)[set_idx].set_value();
}
}