fix(qudarap): exclude fake_propagate CUDA path from PRODUCTION builds (#255)

Guard the test-only `_QSim_fake_propagate` launcher and its host entry
point with `!defined(PRODUCTION)`.

In PRODUCTION builds, make `QSim::fake_propagate` fail fast on the host
side instead of compiling the CUDA path.

This avoids the CUDA 12.5.x `cicc` compile-time blow-up in `QSim.cu`
Release builds while leaving the normal runtime propagation code
unchanged and preserving fake-propagate behavior in non-PRODUCTION
builds.

Verified with:
- direct Release `nvcc` compile of `QSim.cu`
- `cmake --build /opt/eic-opticks/build --parallel --target install`

An actual fix for #57

Author: plexoos

qudarap/QSim.cc

@@ -1519,9 +1519,9 @@ quad2* QSim::UploadFakePRD(const NP* ip, const NP* prd) // static
     return d_prd ;
 }
 
-
-
+#if !defined(PRODUCTION)
 extern void QSim_fake_propagate_launch(dim3 numBlocks, dim3 threadsPerBlock, qsim* sim, quad2* prd );
+#endif
 
 /**
 
@@ -1545,6 +1545,12 @@ using common QEvt functionality
 
 void QSim::fake_propagate( const NP* prd, unsigned type )
 {
+#if defined(PRODUCTION)
+    (void)prd;
+    (void)type;
+    LOG(fatal) << "QSim::fake_propagate is disabled in PRODUCTION builds";
+    std::raise(SIGINT);
+#else
     const NP* ip = sev->getInputPhoton();
     int num_ip = ip ? ip->shape[0] : 0 ;
     assert( num_ip > 0 );
@@ -1584,6 +1590,7 @@ void QSim::fake_propagate( const NP* prd, unsigned type )
 
 
     LOG(LEVEL) << "]" ;
+#endif
 }
 
 

qudarap/QSim.cu

@@ -648,6 +648,7 @@ extern void QSim_photon_launch(dim3 numBlocks, dim3 threadsPerBlock, qsim* sim,
 }
 
 
+#if !defined(PRODUCTION)
 /**
 _QSim_fake_propagate
 -----------------------
@@ -686,8 +687,7 @@ extern void QSim_fake_propagate_launch(dim3 numBlocks, dim3 threadsPerBlock, qsi
     _QSim_fake_propagate<<<numBlocks,threadsPerBlock>>>( sim, prd );
 }
 
-
-
+#endif
 
 __global__ void _QSim_boundary_lookup_all(qsim* sim, quad* lookup, unsigned width, unsigned height )
 {

qudarap/qsim.h

@@ -24,10 +24,8 @@ Canonical use is from CSGOptiX/CSGOptiX7.cu:simulate
 
 #if defined(__CUDACC__) || defined(__CUDABE__)
    #define QSIM_METHOD __device__
-   #define QSIM_NOINLINE __noinline__
 #else
    #define QSIM_METHOD
-   #define QSIM_NOINLINE
 #endif
 
 #include "OpticksGenstep.h"
@@ -70,11 +68,6 @@ Canonical use is from CSGOptiX/CSGOptiX7.cu:simulate
 
 struct qcerenkov ;
 
-static QSIM_METHOD QSIM_NOINLINE float qsim_rng_uniform(RNG *rng)
-{
-    return curand_uniform(rng);
-}
-
 struct qsim
 {
     qbase*              base ;
@@ -1093,13 +1086,9 @@ inline QSIM_METHOD int qsim::propagate_at_boundary(unsigned& flag, RNG& rng, sct
 
 
 #if !defined(PRODUCTION) && defined(DEBUG_TAG)
-    const float u_boundary_burn = qsim_rng_uniform(&rng) ;  // needed for random consumption alignment with Geant4 G4OpBoundaryProcess::PostStepDoIt
+    const float u_boundary_burn = curand_uniform(&rng) ;  // needed for random consumption alignment with Geant4 G4OpBoundaryProcess::PostStepDoIt
 #endif
-    // Keep the scalar Philox draw behind a noinline helper. CUDA 12.5.1 can
-    // spend pathological compile time when this curand state-machine is fully
-    // inlined into this large function at high optimization. CUDA 13.0.2
-    // compiles cleanly either way, but this preserves the original RNG stream.
-    const float u_reflect = qsim_rng_uniform(&rng);
+    const float u_reflect = curand_uniform(&rng) ;
     bool reflect = u_reflect > TransCoeff  ;
 
 #if !defined(PRODUCTION) && defined(DEBUG_TAG)