Add FEEvaluationImplHangingNodes::RunTypes::scalar #12999
Conversation
| template <typename T> | ||
| struct UnivariateShapeDataTrait; | ||
|
|
||
| template <> | ||
| struct UnivariateShapeDataTrait<double> | ||
| { | ||
| using value_type = double; | ||
| }; | ||
|
|
||
| template <> | ||
| struct UnivariateShapeDataTrait<float> | ||
| { | ||
| using value_type = float; | ||
| }; | ||
|
|
||
| template <typename T, std::size_t width> | ||
| struct UnivariateShapeDataTrait<VectorizedArray<T, width>> | ||
| { | ||
| using value_type = T; | ||
| }; |
There was a problem hiding this comment.
@kronbichler Do we have somewhere a functionality like this?
There was a problem hiding this comment.
No, I don't think so, but it would be useful to have in a central place (read: vectorization.h). And in general, I think you can simplify this by saying
template <typename T>
struct UnivariateShapeDataTrait
{
using value_type = T;
};and then only a specialization for VectorizedArray<T, width>.
peterrum
force-pushed
the
FEEvaluationImplHangingNodes_scalar_
branch
from
f9dd536
to
38b0b56
Compare
| unsigned int d, | ||
| bool transpose, | ||
| bool skip_borders> | ||
| static inline DEAL_II_ALWAYS_INLINE void |
There was a problem hiding this comment.
How big is the effect of inlining this function? I am hesitant to inline this in debug mode, but it might make sense in release mode. Can you wrap this into an ifndef DEBUG?
| {{0, points - 1, 0, points - 1}}}}}}}}; | ||
|
|
||
| const std:: | ||
| array<std::array<std::array<std::array<unsigned int, 3>, 2>, 2>, 3> |
There was a problem hiding this comment.
Can you use ndarray here and in a few other places?
| return (fe_degree == -1 || fe_degree > 2) ? | ||
| FEEvaluationImplHangingNodesRunnerTypes::vectorized : | ||
| FEEvaluationImplHangingNodesRunnerTypes::scalar; |
There was a problem hiding this comment.
I suggest we also run the vectorized case only if we have Number::size() > 2, because the vectorized code path executes two sides of an interpolation, so with 2 SIMD lanes the scalar code does strictly less operations.
peterrum
force-pushed
the
FEEvaluationImplHangingNodes_scalar_
branch
from
2421935
to
c3565f8
Compare
|
Can you rebase this to master to see if the code now runs? |
peterrum
force-pushed
the
FEEvaluationImplHangingNodes_scalar_
branch
from
553ff34
to
3e188b0
Compare
|
/rebuild |
| enum class VectorizationTypes | ||
| { | ||
| /** | ||
| * TODO |
| } | ||
|
|
||
| template <unsigned int direction> | ||
| static inline DEAL_II_ALWAYS_INLINE void |
There was a problem hiding this comment.
Also here, I would not inline in debug mode.
| } | ||
|
|
||
| template <bool do_x, bool do_y, bool do_z> | ||
| inline DEAL_II_ALWAYS_INLINE void |
| enum class HelperType | ||
| { | ||
| /** | ||
| * TODO |
peterrum
force-pushed
the
FEEvaluationImplHangingNodes_scalar_
branch
from
6d7847a
to
5d3a0f4
Compare
peterrum
force-pushed
the
FEEvaluationImplHangingNodes_scalar_
branch
from
5d3a0f4
to
6b70043
Compare
| #ifndef DEBUG | ||
| inline DEAL_II_ALWAYS_INLINE | ||
| #endif |
There was a problem hiding this comment.
What do you think about simplifying the call to these #ifdefs by defining (for this file only)
#ifdef DEBUG
#define DEAL_II_ALWAYS_INLINE_RELEASE
#else
#define DEAL_II_ALWAYS_INLINE_RELEASE DEAL_II_ALWAYS_INLINE
#endifand then use this flag, and at the end of the relevant section call
#undef DEAL_II_ALWAYS_INLINE_RELEASEI think this gives us a good intermediate state to merge for now. We can then later analyze and see if we want to make additional changes before the release.
alternative to #12781