GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-03 17:24:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

Don't create more cuda blocks than necessary

Browse Source
This commit is contained in:
Benoit Steiner 2015-11-23 11:00:10 -08:00
parent df31ca3b9e
commit 562078780a
1 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
View File

@ -219,9 +219,11 @@ inline void TensorExecutor<Expression, GpuDevice, false>::run(const Expression&
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign) if (needs_assign)
{ {
const int num_blocks = device.getNumCudaMultiProcessors() * device.maxCudaThreadsPerMultiProcessor() / device.maxCudaThreadsPerBlock();
const int block_size = device.maxCudaThreadsPerBlock(); const int block_size = device.maxCudaThreadsPerBlock();
const int max_blocks = device.getNumCudaMultiProcessors() * device.maxCudaThreadsPerMultiProcessor() / block_size;
const Index size = array_prod(evaluator.dimensions()); const Index size = array_prod(evaluator.dimensions());
// Create a least one block to ensure we won't crash if we're called with tensors of size 0.
const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, (size + block_size - 1) / block_size), 1);
LAUNCH_CUDA_KERNEL((EigenMetaKernel_NonVectorizable<TensorEvaluator<Expression, GpuDevice>, Index>), num_blocks, block_size, 0, device, evaluator, size); LAUNCH_CUDA_KERNEL((EigenMetaKernel_NonVectorizable<TensorEvaluator<Expression, GpuDevice>, Index>), num_blocks, block_size, 0, device, evaluator, size);
} }
evaluator.cleanup(); evaluator.cleanup();
@ -236,9 +238,11 @@ inline void TensorExecutor<Expression, GpuDevice, true>::run(const Expression& e
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign) if (needs_assign)
{ {
const int num_blocks = device.getNumCudaMultiProcessors() * device.maxCudaThreadsPerMultiProcessor() / device.maxCudaThreadsPerBlock();
const int block_size = device.maxCudaThreadsPerBlock(); const int block_size = device.maxCudaThreadsPerBlock();
const int max_blocks = device.getNumCudaMultiProcessors() * device.maxCudaThreadsPerMultiProcessor() / block_size;
const Index size = array_prod(evaluator.dimensions()); const Index size = array_prod(evaluator.dimensions());
// Create a least one block to ensure we won't crash if we're called with tensors of size 0.
const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, (size + block_size - 1) / block_size), 1);
LAUNCH_CUDA_KERNEL((EigenMetaKernel_Vectorizable<TensorEvaluator<Expression, GpuDevice>, Index>), num_blocks, block_size, 0, device, evaluator, size); LAUNCH_CUDA_KERNEL((EigenMetaKernel_Vectorizable<TensorEvaluator<Expression, GpuDevice>, Index>), num_blocks, block_size, 0, device, evaluator, size);
} }
evaluator.cleanup(); evaluator.cleanup();