pecos/tps

virtual method of a class cannot be used in other class' method on gpu hip

Closed this issue · 2 comments

dreamer2368 commented

This is demonstrated in DGNonLinearForm::evalFaceFlux_gpu() as a minimal example on the branch gpu_dryair_transport.

  // get state and gradient
  for (int eq = 0; eq < num_equation; eq++) {
    u1[eq] = d_uk_el1[eq + k * num_equation + iface * maxIntPoints * num_equation];
    u2[eq] = d_uk_el2[eq + k * num_equation + iface * maxIntPoints * num_equation];
  }

  // evaluate flux
  // d_rsolver->Eval_LF(u1, u2, nor, Rflux);

  // For some unknown reason, the hip build is very sensitive to
  // how Eval_LF is called.  If we copy the state to u1, u2 and
  // normal to nor and then call as above (which behaves correctly
  // for CUDA), the resulting flux is nonsense.  But, if we just
  // point into d_uk_el1 at the right spots, all is well.  This
  // problem seems to have something to do with virtual functions
  // and pointer arguments, but I don't understand it.  However,
  // this approach is working.
  d_rsolver->Eval_LF(d_uk_el1 + k * num_equation + iface * maxIntPoints * num_equation,
                     d_uk_el2 + k * num_equation + iface * maxIntPoints * num_equation,
                     d_shapeWnor1 + offsetShape1 + maxDofs + 1 + k * (maxDofs + 1 + dim),
                     Rflux);

A known problem in this approach is that it is not applicable to other places where data array indexing order is different.

While this is working at least for d_rsolver->Eval_LF, following the same approach for d_flux->ComputeViscousFluxes does not work (see line 371 in src/dgNonlinearForm.cpp). Through a small test on the branch issue145, it turns out that the problem is the calling of other objects' virtual methods, such as mixture->computeSpeciesEnthalpies or transport->ComputeFluxTransportProperties within d_flux->ComputeViscousFluxes.

  • When all these methods are hard-coded within the function, d_flux->ComputeViscousFluxes works.

  • If these virtual methods are changed to non-virtual, they work properly on gpu hip. On the branch issue145 (see line 95 of src/equation_of_state.cpp), GasMixture::computeSpeciesEnthalpies is defined as non-virtual, which can pass cyl3d.gpu.test. (For demonstration purpose, I hard-coded transport->ComputeFluxTransportProperties within d_flux->ComputeViscousFluxes, on line 283 of src/fluxs.cpp. So other tests will fail except cyl3d.gpu.test.)

  • Defining these methods to be const did not make a difference.

  • Both computeSpeciesEnthalpies and ComputeFluxTransportProperties are defined to be MFEM_HOST_DEVICE.

  • For an unknown reason, mixture->ComputePressure works even within other objects' method. Since this is defined inline, I also tried inline for other methods, but it didn't work.

  • Just as const RiemannSolver *d_rsolver = rsolver_;, I attempted to define a pointer GasMixture *d_mix = mixture; within ComputeViscousFluxes routine, although this did not work. Perhaps the pointer must be defined outside MFEM_FORALL loop, but this is not attempted yet.

It seems that the last point is the fundamental reason for the failure. @trevilo , why should these objects be defined outside MFEM_FORALL loop? Do you think there would be any workaround for this?

dreamer2368 commented

On the commit 4ff9390cfeaa1a876c0bda712ee5337f72cb0afe, GasMixture object pointer is created outside the MFEM_FORALL loop, and taken as an input argument of ComputeViscousFluxes. In this branch, TransportProperties::ComputeFluxTransportProperties is hard-coded within ComputeViscousFluxes.

This returns a memory access fault on gpu hip path:

------------------------------------
  _______ _____   _____
 |__   __|  __ \ / ____|
    | |  | |__) | (___  
    | |  |  ___/ \___ \ 
    | |  | |     ____) | 
    |_|  |_|    |_____/ 

TPS Version:  1.1 (dev)
Git Version:  e798d61
MFEM Version: MFEM v4.3 (release)
------------------------------------

Options used:
   --no-version
   --runFile inputs/input.dtconst.cyl.ini
   --no-debug

Caching input file -> inputs/input.dtconst.cyl.ini

---------------------------------
MFEM Device configuration:
Device configuration: hip,cpu
Memory configuration: host-std,hip
---------------------------------


TPS is using non-GPU-aware MPI.
Fluid is set to the preset 'dry_air'. Input options in [plasma_models] will not be used.
number of elements on rank 0 = 6153
min elements/partition       = 6153
max elements/partition       = 6153
restart: 0
Initial time-step: 2.53434e-06s

[INLET]: Patch number                      = 1 
[INLET]: Total Surface Area                = 9.78349e-03
[INLET]: # of boundary faces               = 133 
[INLET]: # of participating MPI partitions = 1 

[OUTLET]: Patch number                      = 2 
[OUTLET]: Total Surface Area                = 0.00978
[OUTLET]: # of boundary faces               = 131 
[OUTLET]: # of participating MPI partitions = 1 
Iteration = 0: wall clock time/iter = 0.000 (secs)
Memory access fault by GPU node-1 (Agent handle: 0x1c6afc0) on address 0x1d9d000. Reason: Page not present or supervisor privilege.

It seems that there are similar issues (ROCm/tensorflow-upstream#302) with gfx803 architecture.

trevilo commented

This issue appears to be hardware specific. We were seeing the problems on our gfx803 card, but recent testing on a gfx90a (after unifying the _CUDA_ and _HIP_ paths in #192) doesn't show any problems. Since gfx803 is an old card that isn't fully supported anymore, I don't think we should worry about this unless we start seeing issues on newer hardware.