GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-13 04:09:10 +08:00
Code Issues Packages Projects Releases Wiki Activity

Optimized the performance of narrow reductions on CUDA devices

Browse Source
This commit is contained in:
Benoit Steiner 2016-02-29 10:48:16 -08:00
parent e9bea614ec
commit 3284842045
2 changed files with 18 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -544,8 +544,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
Op reducer(m_reducer);
internal::InnerReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve);
return false;
return internal::InnerReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve);
}
bool preserving_inner_dims = true;
@ -561,8 +560,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
Op reducer(m_reducer);
internal::OuterReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve);
return false;
return internal::OuterReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve);
}
}
return true;

18
unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
View File

@ -230,9 +230,14 @@ struct InnerReducer<Self, Op, GpuDevice> {
assert(false && "Should only be called to reduce floats on a gpu device");
}
static EIGEN_DEVICE_FUNC void run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
static EIGEN_DEVICE_FUNC bool run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
typedef typename Self::Index Index;
// It's faster to use the usual code.
if (num_coeffs_to_reduce <= 32) {
return true;
}
const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
const int block_size = 256;
const int num_per_thread = 128;
@ -255,6 +260,8 @@ struct InnerReducer<Self, Op, GpuDevice> {
LAUNCH_CUDA_KERNEL((InnerReductionKernel<num_per_thread, Self, Op, Index>),
num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
return false;
}
};
@ -301,9 +308,14 @@ struct OuterReducer<Self, Op, GpuDevice> {
assert(false && "Should only be called to reduce floats on a gpu device");
}
static EIGEN_DEVICE_FUNC void run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
static EIGEN_DEVICE_FUNC bool run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
typedef typename Self::Index Index;
// It's faster to use the usual code.
if (num_coeffs_to_reduce <= 32) {
return true;
}
const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
const int block_size = 256;
const int num_per_thread = 16;
@ -326,6 +338,8 @@ struct OuterReducer<Self, Op, GpuDevice> {
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);
return false;
}
};