blob: 5950d34242256379d54d8e6fbe18c2e1bcba5ce4 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
|
Probe C: split the remaining D3Q27 JIT failure in half.
WHAT IS BROKEN
At D3Q27 the HlbmOneParticleGather kernel fails AdaptiveCpp's SSCP JIT at
run time, every launch:
LLVMToHost: clang invocation failed with exit code 1
The kernel then never executes, so the particle never moves and its force
reads exactly 0. It builds fine -- this only shows up when running.
The same chain at D2Q9 is fine, so it is specific to 3D.
WHAT IS ALREADY KNOWN
Inserting an early return at successive points in the gather:
after the AABB setup ................................. JIT passes
+ iterator<3>::apply with an empty body .............. JIT passes
+ porosity, before the 27-direction loop ............. THIS PROBE
full gather .......................................... JIT fails
So the failure is inside the per-cell lambda, and this probe is the one
that separates its two halves.
HOW TO READ THE RESULT
JIT passes -> the 27-direction momentum-exchange loop is the culprit.
JIT fails -> the particle_porosity call is the culprit.
APPLY / REVERT (from the repo root)
patch -p1 < probe_c.patch
patch -p1 -R < probe_c.patch
RUN
nix-build -A ex.poiseulle_particle_3d_hlbm_gpu
HOME=$(mktemp -d) ./result/bin/* --name probe_c --time_steps 2 2>&1 \
| grep -c 'clang invocation failed'
0 = JIT passed, nonzero = still failing. Expect roughly one failure per
launch per step, so 2 steps gives 0 or ~2-4.
NOTE
This makes the gather compute no force at all -- it is a diagnostic, not a
fix. Revert before running anything real.
--- a/modules/core/c++/hlbm.hpp
+++ b/modules/core/c++/hlbm.hpp
@@ -188,6 +188,9 @@
return;
}
+ // --- PROBE C: stop here, before the 27-direction loop. ---
+ return;
+
kel::lbm::data<sch::Vector<T,Desc::D>> momentum;
for(uint64_t i{0u}; i < Desc::D; ++i){
momentum.at({{i}}) = 0.0;
|