Multigrid clean up phase, and fix shell variables for libfabric bug work around on Frontier

Grid/algorithms/multigrid/MultiGrid.h

@@ -35,3 +35,8 @@ Author: Peter Boyle <pboyle@bnl.gov>
 #include <Grid/algorithms/multigrid/GeneralCoarsenedMatrixMultiRHSV2.h>
 #include <Grid/algorithms/multigrid/MrhsPromotedOperator.h>
 #include <Grid/algorithms/multigrid/Smoothers.h>
+#include <Grid/algorithms/multigrid/PVdagMMultiGridParams.h>
+// PVdagMOperators.h / MrhsMultiGrid.h / PVdagMMultiGrid.h /
+// DenseCoarseMatrix.h are NOT in this umbrella: consumers of the PVdagM
+// chain include PVdagMMultiGrid.h explicitly (it pulls the dense stack
+// and BLAS).

Grid/stencil/Stencil.h

@@ -453,9 +453,23 @@ public:
 	double *dbuf =(double *) packet.recv_buf;
 	float  *fbuf =(float  *) packet.compressed_recv_buf;
 
+	// BUG FIX 2026-09-07: the lane structure is a carefully designed
+	// GPU optimisation -- lane = threadIdx.y puts adjacent threads on
+	// adjacent words, fully coalesced.  On CPU builds there are no
+	// SIMT lanes and acceleratorSIMTlane() is identically zero, so
+	// this loop converted only 1/nsimd of the words and the rest of
+	// the halo silently kept STALE buffer content (deterministic
+	// wrong answers whenever the previous occupant differed; caught
+	// on the 2-rank laptop build via Benchmark_dwf's sloppy Cshift
+	// check).  GRID_SIMT keeps the optimisation; CPU loops the lanes.
 	accelerator_forNB(ss,outer,nsimd,{
+#ifdef GRID_SIMT
 	  int lane = acceleratorSIMTlane(nsimd);
 	  dbuf[ss*nsimd+lane] = fbuf[ss*nsimd+lane]; //conversion
+#else
+	  for(int lane=0;lane<nsimd;lane++)
+	    dbuf[ss*nsimd+lane] = fbuf[ss*nsimd+lane]; //conversion
+#endif
 	});
 
       } else if ( sizeof(word)==4){
@@ -467,8 +481,13 @@ public:
 	uint16_t *hbuf =(uint16_t *) packet.compressed_recv_buf;
 
 	accelerator_forNB(ss,outer,nsimd,{
+#ifdef GRID_SIMT
 	  int lane = acceleratorSIMTlane(nsimd);
 	  fbuf[ss*nsimd+lane] = ((uint32_t)hbuf[ss*nsimd+lane])<<16; //copy back and pad each word with zeroes
+#else
+	  for(int lane=0;lane<nsimd;lane++)
+	    fbuf[ss*nsimd+lane] = ((uint32_t)hbuf[ss*nsimd+lane])<<16; //copy back and pad each word with zeroes
+#endif
 	});
 
       } else {
@@ -515,9 +534,15 @@ public:
 	double *dbuf =(double *) packet.send_buf;
 	float  *fbuf =(float  *) packet.compressed_send_buf;
 
+	// BUG FIX 2026-09-07: CPU lane coverage -- see DecompressPacket.
 	accelerator_forNB(ss,outer,nsimd,{
+#ifdef GRID_SIMT
 	  int lane = acceleratorSIMTlane(nsimd);
 	  fbuf[ss*nsimd+lane] = dbuf[ss*nsimd+lane]; // convert fp64 to fp32
+#else
+	  for(int lane=0;lane<nsimd;lane++)
+	    fbuf[ss*nsimd+lane] = dbuf[ss*nsimd+lane]; // convert fp64 to fp32
+#endif
 	});
 
       } else if ( sizeof(word)==4){
@@ -526,8 +551,13 @@ public:
 	uint16_t *hbuf =(uint16_t *) packet.compressed_send_buf;
 	
 	accelerator_forNB(ss,outer,nsimd,{
+#ifdef GRID_SIMT
 	  int lane = acceleratorSIMTlane(nsimd);
 	  hbuf[ss*nsimd+lane] = fbuf[ss*nsimd+lane]>>16; // convert as in Bagel/BFM ; bfloat16 ; s7e8 Intel patent
+#else
+	  for(int lane=0;lane<nsimd;lane++)
+	    hbuf[ss*nsimd+lane] = fbuf[ss*nsimd+lane]>>16; // convert as in Bagel/BFM ; bfloat16 ; s7e8 Intel patent
+#endif
 	});
 
       } else {

systems/Frontier/sourceme-rocm6.4.sh

@@ -14,3 +14,8 @@ export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
 export LD_LIBRARY_PATH=/opt/rocm-6.4.0/lib/llvm/lib/:$LD_LIBRARY_PATH
 module load emacs
 export FI_MR_CACHE_MONITOR=kdreg2       # REQUIRED for device-buffer MPI on Slingshot: libfabric memhooks monitor (default) is defective, see systems/WorkArounds.txt (libfabric #11451)
+
+# Standing avoidance of the CXI NO_TRANSLATION fault on persistent device
+# buffers under hipMalloc/hipFree churn (libfabric issue #12775; ruling
+# 2026-09-07). 
+export FI_HMEM_ROCR_USE_DMABUF=0

systems/Frontier/sourceme-rocm7.0.sh

@@ -12,3 +12,8 @@ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CRAY_LD_LIBRARY_PATH
 
 module load emacs
 export FI_MR_CACHE_MONITOR=kdreg2       # REQUIRED for device-buffer MPI on Slingshot: libfabric memhooks monitor (default) is defective, see systems/WorkArounds.txt (libfabric #11451)
+
+# Standing avoidance of the CXI NO_TRANSLATION fault on persistent device
+# buffers under hipMalloc/hipFree churn (libfabric issue #12775; ruling
+# 2026-09-07). 
+export FI_HMEM_ROCR_USE_DMABUF=0

systems/Frontier/sourceme-rocm7.2.sh

@@ -3,17 +3,17 @@ module unload slate 2>/dev/null
 echo spack
 . /autofs/nccs-svm1_home1/paboyle/spack-frontier/share/spack/setup-env.sh
 
-export SLATE=`spack find --paths slate   | grep ^slate  | awk '{print $2}' `
+#export SLATE=`spack find --paths slate   | grep ^slate  | awk '{print $2}' `
-export BLASPP=`spack find --paths blaspp | grep ^blaspp | awk '{print $2}' `
+#export BLASPP=`spack find --paths blaspp | grep ^blaspp | awk '{print $2}' `
-export LAPACKPP=`spack find --paths lapackpp | grep ^lapackpp | awk '{print $2}' `
+#export LAPACKPP=`spack find --paths lapackpp | grep ^lapackpp | awk '{print $2}' `
 
-echo SLATE    $SLATE
+#echo SLATE    $SLATE
-echo LAPACKPP $LAPACKPP
+#echo LAPACKPP $LAPACKPP
-echo BLASPP   $BLASPP
+#echo BLASPP   $BLASPP
 
-ls $SLATE/lib64/libslate.so $LAPACKPP/lib64/liblapackpp.so 2>/dev/null || ls $SLATE/lib $LAPACKPP/lib
+#ls $SLATE/lib64/libslate.so $LAPACKPP/lib64/liblapackpp.so 2>/dev/null || ls $SLATE/lib $LAPACKPP/lib
-ldd $SLATE/lib64/libslate.so | grep -E "blaspp|lapackpp|rocblas|amdhip|mpi_cray|libsci"
+#ldd $SLATE/lib64/libslate.so | grep -E "blaspp|lapackpp|rocblas|amdhip|mpi_cray|libsci"
-ldd $LAPACKPP/lib64/liblapackpp.so | grep -E "blaspp|libsci"
+#ldd $LAPACKPP/lib64/liblapackpp.so | grep -E "blaspp|libsci"
 
 export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
 export MPFR=`spack find --paths mpfr    | grep ^mpfr  | awk '{print $2}' `
@@ -24,9 +24,14 @@ module load cce/21.0.0
 module load cpe/26.03
 module load rocm/7.2.0
 export LD_LIBRARY_PATH=/opt/rocm-7.2.0/lib/llvm/lib/:$LD_LIBRARY_PATH
-export LD_LIBRARY_PATH=$LAPACKPP/lib64:$LD_LIBRARY_PATH
+#export LD_LIBRARY_PATH=$LAPACKPP/lib64:$LD_LIBRARY_PATH
-export LD_LIBRARY_PATH=$BLASPP/lib64:$LD_LIBRARY_PATH
+#export LD_LIBRARY_PATH=$BLASPP/lib64:$LD_LIBRARY_PATH
-export LD_LIBRARY_PATH=$SLATE/lib64:$LD_LIBRARY_PATH
+#export LD_LIBRARY_PATH=$SLATE/lib64:$LD_LIBRARY_PATH
 export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CRAY_LD_LIBRARY_PATH
 module load emacs
+
+# Standing avoidance of the CXI NO_TRANSLATION fault on persistent device
+# buffers under hipMalloc/hipFree churn (libfabric issue #12775; ruling
+# 2026-09-07).  
 export FI_MR_CACHE_MONITOR=kdreg2       # REQUIRED for device-buffer MPI on Slingshot: libfabric memhooks monitor (default) is defective, see systems/WorkArounds.txt (libfabric #11451)
+export FI_HMEM_ROCR_USE_DMABUF=0