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

Avoid calling get_future() more than once on a given promise.

Browse Source
This commit is contained in:
Benoit Steiner 2014-10-16 10:10:04 -07:00
parent bfdd9f3ac9
commit d853adffdb
2 changed files with 8 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -198,8 +198,10 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
// this should really be numBlockAs * n_blocks; // this should really be numBlockAs * n_blocks;
const Index num_kernel_promises = num_threads * n_blocks; const Index num_kernel_promises = num_threads * n_blocks;
std::vector<Promise> kernel_promises(num_kernel_promises); std::vector<Promise> kernel_promises(num_kernel_promises);
std::vector<Future> kernel_futures(num_kernel_promises);
for (int i = 0; i < kernel_promises.size(); ++i) { for (int i = 0; i < kernel_promises.size(); ++i) {
kernel_promises[i].set_value(); kernel_promises[i].set_value();
kernel_futures[i] = kernel_promises[i].get_future();
} }
for (Index k_block_idx = 0; k_block_idx < k_blocks; k_block_idx++) { for (Index k_block_idx = 0; k_block_idx < k_blocks; k_block_idx++) {
@ -218,8 +220,9 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
int blockAId = (k_block_idx * m_blocks + mt_block_idx) % num_threads; int blockAId = (k_block_idx * m_blocks + mt_block_idx) % num_threads;
for (int i = 0; i < n_blocks; ++i) { for (int i = 0; i < n_blocks; ++i) {
int future_id = (blockAId * n_blocks + i); int future_id = (blockAId * n_blocks + i);
wait_until_ready(&kernel_promises[future_id]); wait_until_ready(&kernel_futures[future_id]);
kernel_promises[future_id] = Promise(); kernel_promises[future_id] = Promise();
kernel_futures[future_id] = kernel_promises[future_id].get_future();
} }
const packLArg arg = { const packLArg arg = {
blockAs[blockAId], // blockA blockAs[blockAId], // blockA
@ -248,7 +251,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
for (int i = num_blocks; i < num_threads; ++i) { for (int i = num_blocks; i < num_threads; ++i) {
int blockAId = (k_block_idx * m_blocks + i + m_block_idx) % num_threads; int blockAId = (k_block_idx * m_blocks + i + m_block_idx) % num_threads;
int future_id = (blockAId * n_blocks + n_block_idx); int future_id = (blockAId * n_blocks + n_block_idx);
wait_until_ready(&kernel_promises[future_id]); wait_until_ready(&kernel_futures[future_id]);
} }
} }
@ -281,9 +284,9 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
} }
} }
// collect the last frame of kernel futures // Make sure all the kernels are done.
for (int i = 0; i < kernel_promises.size(); ++i) { for (int i = 0; i < kernel_futures.size(); ++i) {
wait_until_ready(&kernel_promises[i]); wait_until_ready(&kernel_futures[i]);
} }
// deallocate all of the memory for both A and B's // deallocate all of the memory for both A and B's

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

@ -46,11 +46,6 @@ static EIGEN_STRONG_INLINE void wait_until_ready(const Future* f) {
// eigen_assert(f->ready()); // eigen_assert(f->ready());
} }
static EIGEN_STRONG_INLINE void wait_until_ready(Promise* p) {
p->get_future().wait();
// eigen_assert(p->get_future().ready());
}
struct ThreadPoolDevice { struct ThreadPoolDevice {
ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : num_threads_(num_cores) { } ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : num_threads_(num_cores) { }