GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-11 19:29:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

Re-enabled the optimized reduction CUDA code.

Browse Source
This commit is contained in:
Benoit Steiner 2016-01-11 09:07:14 -08:00
parent b557662e58
commit 780623261e
2 changed files with 3 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -527,7 +527,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
}
// Attempt to use an optimized reduction.
#if 0
else if (RunningOnGPU && data && (m_device.majorDeviceVersion() >= 3)) {
bool reducing_inner_dims = true;
for (int i = 0; i < NumReducedDims; ++i) {
@ -563,7 +562,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
return false;
}
}
#endif
return true;
}

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

@ -126,7 +126,7 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
const int block_size = 256;
const int num_per_thread = 128;
const int num_blocks = std::ceil(static_cast<float>(num_coeffs) / (block_size * num_per_thread));
LAUNCH_CUDA_KERNEL((FullReductionKernel<block_size, num_per_thread>),
LAUNCH_CUDA_KERNEL((FullReductionKernel<block_size, num_per_thread, Self, Op, Index>),
num_blocks, block_size, 0, device, reducer, self, num_coeffs, output);
}
};
@ -222,7 +222,7 @@ struct InnerReducer<Self, Op, GpuDevice> {
const int num_per_thread = 128;
const int num_blocks = 32;
LAUNCH_CUDA_KERNEL((InnerReductionKernel<num_per_thread>),
LAUNCH_CUDA_KERNEL((InnerReductionKernel<num_per_thread, Self, Op, Index>),
num_blocks, block_size, block_size*sizeof(float), device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
}
};
@ -279,7 +279,7 @@ struct OuterReducer<Self, Op, GpuDevice> {
device.maxCudaThreadsPerMultiProcessor() / block_size;
const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
LAUNCH_CUDA_KERNEL((OuterReductionKernel<num_per_thread>),
LAUNCH_CUDA_KERNEL((OuterReductionKernel<num_per_thread, Self, Op, Index>),
num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
}
};