GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-06-04 18:54:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

Tweak cost model for tensor contraction when parallelizing over the inner dimension.

Browse Source 
https://bitbucket.org/snippets/rmlarsen/MexxLo
This commit is contained in:
Rasmus Munk Larsen 2019-04-12 13:35:10 -07:00
parent 9a3f06d836
commit 039ee52125
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -1169,7 +1169,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
TensorOpCost contractionCostPerInnerDim(Index m, Index n, Index k) const { TensorOpCost contractionCostPerInnerDim(Index m, Index n, Index k) const {
// Compute cost. // Compute cost.
const int output_packet_size = internal::unpacket_traits<PacketReturnType>::size; const int output_packet_size = internal::unpacket_traits<PacketReturnType>::size;
TensorOpCost cost(0, 0, (computeBandwidth(true, m, n, k) * m) * n); TensorOpCost cost(0, 0, (computeBandwidth(true, m, n, k) * m) * n, true, output_packet_size);
// Output stores. // Output stores.
cost += TensorOpCost(0, sizeof(CoeffReturnType), 0, true, output_packet_size); cost += TensorOpCost(0, sizeof(CoeffReturnType), 0, true, output_packet_size);
TensorOpCost lhsCost = this->m_leftImpl.costPerCoeff(true) * m; TensorOpCost lhsCost = this->m_leftImpl.costPerCoeff(true) * m;
@ -1192,8 +1192,8 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
int num_threads = 1; int num_threads = 1;
double min_cost = total_parallel_cost; double min_cost = total_parallel_cost;
double kPerThreadOverHead = 4000; double kPerThreadOverHead = 4000;
double kFixedOverHead = 100000; double kFixedOverHead = 50000;
for (int nt = 2; nt <= this->m_device.numThreads(); nt++) { for (int nt = 2; nt <= this->m_device.numThreads(); nt += 2) {
double sequential_cost = double sequential_cost =
kFixedOverHead + nt * (reduction_cost + kPerThreadOverHead); kFixedOverHead + nt * (reduction_cost + kPerThreadOverHead);
double parallel_cost = total_parallel_cost / nt + sequential_cost; double parallel_cost = total_parallel_cost / nt + sequential_cost;