Fix computeProductBlockingSizes with m==0, and add respective unit test.

This commit is contained in:
Gael Guennebaud 2015-03-31 15:19:57 +02:00
parent bd76d837e6
commit ae01c05e18
2 changed files with 31 additions and 5 deletions

View File

@ -249,10 +249,9 @@ void evaluateProductBlockingSizesHeuristic(Index& k, Index& m, Index& n, Index n
actual_lm = l2;
max_mc = 576;
}
Index mc = (std::min<Index>)(actual_lm/(3*k*sizeof(LhsScalar)), max_mc);
if (mc > Traits::mr) mc -= mc % Traits::mr;
else if (mc==0) return;
m = (m%mc)==0 ? mc
: (mc - Traits::mr * ((mc/*-1*/-(m%mc))/(Traits::mr*(m/mc+1))));
}

View File

@ -134,7 +134,7 @@ void zero_sized_objects(const MatrixType& m)
}
}
template<int>
void bug_127()
{
// Bug 127
@ -159,6 +159,7 @@ void bug_127()
a*b;
}
template<int>
void unaligned_objects()
{
// Regression test for the bug reported here:
@ -188,6 +189,29 @@ void unaligned_objects()
}
}
template<typename T>
EIGEN_DONT_INLINE
Index test_compute_block_size(Index m, Index n, Index k)
{
Index mc(m), nc(n), kc(k);
internal::computeProductBlockingSizes<T,T>(kc, mc, nc);
return kc+mc+nc;
}
template<typename T>
Index compute_block_size()
{
Index ret = 0;
ret += test_compute_block_size<T>(0,1,1);
ret += test_compute_block_size<T>(1,0,1);
ret += test_compute_block_size<T>(1,1,0);
ret += test_compute_block_size<T>(0,0,1);
ret += test_compute_block_size<T>(0,1,0);
ret += test_compute_block_size<T>(1,0,0);
ret += test_compute_block_size<T>(0,0,0);
return ret;
}
void test_product_extra()
{
for(int i = 0; i < g_repeat; i++) {
@ -198,6 +222,9 @@ void test_product_extra()
CALL_SUBTEST_4( product_extra(MatrixXcd(internal::random<int>(1,EIGEN_TEST_MAX_SIZE/2), internal::random<int>(1,EIGEN_TEST_MAX_SIZE/2))) );
CALL_SUBTEST_1( zero_sized_objects(MatrixXf(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );
}
CALL_SUBTEST_5( bug_127() );
CALL_SUBTEST_6( unaligned_objects() );
CALL_SUBTEST_5( bug_127<0>() );
CALL_SUBTEST_6( unaligned_objects<0>() );
CALL_SUBTEST_7( compute_block_size<float>() );
CALL_SUBTEST_7( compute_block_size<double>() );
CALL_SUBTEST_7( compute_block_size<std::complex<double> >() );
}