LWFA on cpu shows different (wrong) physics

[PrometheusPi]

When running the default LWFA example with 32.cfgthere is a major difference between using gpus or cpus.

In the following you will see png pictures of the simulation after 600 iterations/time steps.

The current master brach:

The current devbranch on k80GPUs:

The current devbranch on laserCPUs:

The last is definitely wrong, an a_0=8.0 would cause a blow out, not these streams.

[PrometheusPi]

@kossag14 This is probably not an issue that effects you. I think it is more likely the difference you see arise from #2687.

[PrometheusPi]

as recommended by @psychocoderHPC offline I checked how PIConGPU performs on CPUs (lasernodes) at 60 commits before tags/0.4.0-rc1.

The results looks great! We have "correct" physics! ✨ 🎉

600 iterations of 57bcf5cc592b3bf9ad14902d1cadbd4ec8d76193 on laser nodes:

Probably the laser refactoring #2587 (or #2624) caused this bug.

[PrometheusPi]

@psychocoderHPC as you expected, calling the laser functor is causing the error.

With the fix:

diff --git a/include/picongpu/fields/LaserPhysics.hpp b/include/picongpu/fields/LaserPhysics.hpp
index 66f8c84..09bab6f 100644
--- a/include/picongpu/fields/LaserPhysics.hpp
+++ b/include/picongpu/fields/LaserPhysics.hpp
@@ -79,12 +79,6 @@ namespace fields
 
             uint32_t cellOffsetInSuperCellFromInitPlaneY = LaserFunctor::Unitless::initPlaneY % SuperCellSize::y::value;
 
-            auto accLaserFunctor = laserFunctor(
-                acc,
-                localSuperCellOffset,
-                mappings::threads::WorkerCfg< numWorkers >{ workerIdx }
-            );
-
             mappings::threads::ForEachIdx<
                 mappings::threads::IdxConfig<
                     planeSize,
@@ -96,6 +90,12 @@ namespace fields
                     uint32_t const
                 )
                 {
+                   auto accLaserFunctor = laserFunctor(
+                       acc,
+                       localSuperCellOffset,
+                       mappings::threads::WorkerCfg< numWorkers >{ workerIdx }
+                    );
+
                     /* cell index within the superCell */
                     DataSpace< simDim > cellIdxInSuperCell = DataSpaceOperations< simDim >::template map< LaserPlaneSizeInSuperCell >( linearIdx );
                     cellIdxInSuperCell.y() += cellOffsetInSuperCellFromInitPlaneY;

the issue is solved:

png picture after 600 iterations with above fix on laser cpu nodes:

[ax3l]

thx for the fix in #2689! will add it to 0.4.0-rc2