summaryrefslogtreecommitdiff
path: root/probe_c.patch
diff options
context:
space:
mode:
Diffstat (limited to 'probe_c.patch')
-rw-r--r--probe_c.patch55
1 files changed, 55 insertions, 0 deletions
diff --git a/probe_c.patch b/probe_c.patch
new file mode 100644
index 0000000..5950d34
--- /dev/null
+++ b/probe_c.patch
@@ -0,0 +1,55 @@
+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;