mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-01-24 14:45:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

[SYCL clean up the code] : removing exrta #pragma unroll in SYCL which was causing issues in embeded systems

Browse Source
This commit is contained in:
mehdi-goli 2020-10-27 16:31:33 +00:00 committed by David Tellenbach
parent b9ff791fed
commit a725a3233c
1 changed files with 0 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
View File

@ -100,7 +100,6 @@ struct SecondStepFullReducer {
CoeffReturnType accumulator = *aInPtr;
scratchptr[localid] = op.finalize(accumulator);
#pragma unroll 8
for (Index offset = itemID.get_local_range(0) / 2; offset > 0; offset /= 2) {
itemID.barrier(cl::sycl::access::fence_space::local_space);
if (localid < offset) {
@ -154,7 +153,6 @@ class FullReductionKernelFunctor {
Index start = Evaluator::PacketSize * globalid;
// vectorizable parts
PacketReturnType packetAccumulator = op.template initializePacket<PacketReturnType>();
#pragma unroll(8 / Evaluator::PacketSize)
for (Index i = start; i < VectorizedRange; i += step) {
op.template reducePacket<PacketReturnType>(evaluator.impl().template packet<Unaligned>(i), &packetAccumulator);
}
@ -293,7 +291,6 @@ struct PartialReductionKernel {
const Index per_thread_local_stride = PannelParameters::LocalThreadSizeR * reduce_elements_num_groups;
const Index per_thread_global_stride =
rt == reduction_dim::outer_most ? num_coeffs_to_preserve * per_thread_local_stride : per_thread_local_stride;
#pragma unroll 8
for (Index i = globalRId; i < num_coeffs_to_reduce; i += per_thread_local_stride) {
op.reduce(evaluator.impl().coeff(global_offset), &accumulator);
localOffset += per_thread_local_stride;
@ -391,7 +388,6 @@ struct SecondStepPartialReduction {
OutScalar accumulator = op.initialize();
// num_coeffs_to_reduce is not bigger that 256
#pragma unroll 8
for (Index i = 0; i < num_coeffs_to_reduce; i++) {
op.reduce(*in_ptr, &accumulator);
in_ptr += num_coeffs_to_preserve;