hande-qmc · ajwt · Jan 13, 2025 · Aug 1, 2023 · Aug 1, 2023 · Jan 13, 2025
diff --git a/cmake/custom/compilers/Clang.C.cmake b/cmake/custom/compilers/Clang.C.cmake
@@ -3,5 +3,6 @@ if(NOT DEFINED ENV{CFLAGS})
     set(CMAKE_C_FLAGS         "${CMAKE_C_FLAGS} -DRESTRICT=restrict -DFUNDERSCORE=1 -Qunused-arguments -fcolor-diagnostics")
     set(CMAKE_C_FLAGS_DEBUG   "-O0 -DDEBUG -g3 -Wall -Wextra -Winit-self -Wuninitialized -Wmissing-declarations -Wwrite-strings -Wno-sign-compare")
     set(CMAKE_C_FLAGS_RELEASE "-O3 -DNDEBUG")
+    set(CMAKE_C_FLAGS_RELAPPLESILICON "-O1 -DNDEBUG" CACHE STRING "For apple silicon ARM processors." FORCE)
   endif()
 endif()
diff --git a/cmake/custom/compilers/GNU.C.cmake b/cmake/custom/compilers/GNU.C.cmake
@@ -3,5 +3,6 @@ if(NOT DEFINED ENV{CFLAGS})
     set(CMAKE_C_FLAGS         "${CMAKE_C_FLAGS} -DRESTRICT=restrict -DFUNDERSCORE=1 -fPIC")
     set(CMAKE_C_FLAGS_DEBUG   "-O0 -g3 -DDEBUG -Wall -Wextra -Winit-self -Wuninitialized -Wmissing-declarations -Wwrite-strings -Wno-sign-compare")
     set(CMAKE_C_FLAGS_RELEASE "-O3 -DNDEBUG")
+    set(CMAKE_C_FLAGS_RELAPPLESILICON "-O1 -DNDEBUG" CACHE STRING "For apple silicon ARM processors." FORCE)
   endif()
 endif()
diff --git a/cmake/custom/compilers/GNU.Fortran.cmake b/cmake/custom/compilers/GNU.Fortran.cmake
@@ -4,5 +4,6 @@ if(NOT DEFINED ENV{FCFLAGS})
     set(CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS} -fimplicit-none -fautomatic -fmax-errors=5")
     set(CMAKE_Fortran_FLAGS_DEBUG   "-O0 -g -fbacktrace -Wall")
     set(CMAKE_Fortran_FLAGS_RELEASE "-O3 -funroll-all-loops -ftree-vectorize")
+    set(CMAKE_Fortran_FLAGS_RELAPPLESILICON "-O1 -funroll-all-loops -ftree-vectorize" CACHE STRING "For apple silicon ARM processors." FORCE)
   endif()
 endif()
diff --git a/cmake/custom/dsfmt.cmake b/cmake/custom/dsfmt.cmake
@@ -31,32 +31,52 @@ function(system_has_sse2 _result)
   set(_cpu_family)
   set(_cpu_model)
   set(_cpu_flags)
+  set(_cpu_brand)
+  set(_apple_silicon)
   if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
      file(READ "/proc/cpuinfo" _cpuinfo)
      string(REGEX REPLACE ".*vendor_id[ \t]*:[ \t]+([a-zA-Z0-9_-]+).*" "\\1" _vendor_id "${_cpuinfo}")
      string(REGEX REPLACE ".*cpu family[ \t]*:[ \t]+([a-zA-Z0-9_-]+).*" "\\1" _cpu_family "${_cpuinfo}")
      string(REGEX REPLACE ".*model[ \t]*:[ \t]+([a-zA-Z0-9_-]+).*" "\\1" _cpu_model "${_cpuinfo}")
      string(REGEX REPLACE ".*flags[ \t]*:[ \t]+([^\n]+).*" "\\1" _cpu_flags "${_cpuinfo}")
+     string(FIND _cpu_flags "sse2" _sse2_found)
   elseif(CMAKE_SYSTEM_NAME STREQUAL "Darwin")
      execute_process(
        COMMAND
-         "/usr/sbin/sysctl" "-n" "machdep.cpu.vendor" "machdep.cpu.model" "machdep.cpu.family" "machdep.cpu.features"
+         "/usr/sbin/sysctl" "-n" "machdep.cpu.brand_string"
        OUTPUT_VARIABLE
-         _sysctl_output_string
+         _sysctl_output
        OUTPUT_STRIP_TRAILING_WHITESPACE
-       )
-     string(REPLACE "\n" ";" _sysctl_output ${_sysctl_output_string})
-     list(GET _sysctl_output 0 _vendor_id)
-     list(GET _sysctl_output 1 _cpu_model)
-     list(GET _sysctl_output 2 _cpu_family)
-     list(GET _sysctl_output 3 _cpu_flags)
+     )
+     list(GET _sysctl_output 0 _cpu_brand)
+     string(TOLOWER "${_cpu_brand}" _cpu_brand)
+     string(FIND _cpu_brand "apple" _apple_silicon)
+     if(_apple_silicon)
+       set(_sse2_found FALSE)
+     else()
+       execute_process(
+         COMMAND
+           "/usr/sbin/sysctl" "-n" "machdep.cpu.vendor" "machdep.cpu.model" "machdep.cpu.family" "machdep.cpu.features"
+         OUTPUT_VARIABLE
+           _sysctl_output_string
+         OUTPUT_STRIP_TRAILING_WHITESPACE
+         )
+       string(REPLACE "\n" ";" _sysctl_output ${_sysctl_output_string})
+       list(GET _sysctl_output 0 _vendor_id)
+       list(GET _sysctl_output 1 _cpu_model)
+       list(GET _sysctl_output 2 _cpu_family)
+       list(GET _sysctl_output 3 _cpu_flags)
 
-     string(TOLOWER "${_cpu_flags}" _cpu_flags)
-     string(REPLACE "." "_" _cpu_flags "${_cpu_flags}")
+       string(TOLOWER "${_cpu_flags}" _cpu_flags)
+       string(REPLACE "." "_" _cpu_flags "${_cpu_flags}")
+       string(FIND _cpu_flags "sse2" _sse2_found)
+     endif()
   endif()
-  string(FIND _cpu_flags "sse2" _sse2_found)
   if(_sse2_found)
     set(${_result} TRUE PARENT_SCOPE)
     message(STATUS "CPU ${_vendor_id} with SSE2 instruction set FOUND")
+  elseif(_apple_silicon)
+    set(${_result} FALSE PARENT_SCOPE)
+    message(STATUS "Apple silicon detected, disabling SSE2 instruction set")
   endif()
-endfunction()
+endfunction()
diff --git a/cmake/downloaded/autocmake_safeguards.cmake b/cmake/downloaded/autocmake_safeguards.cmake
@@ -21,6 +21,7 @@ string(TOUPPER "${CMAKE_BUILD_TYPE}" cmake_build_type_toupper)
 if(NOT cmake_build_type_tolower STREQUAL "debug" AND
    NOT cmake_build_type_tolower STREQUAL "release" AND
    NOT cmake_build_type_tolower STREQUAL "minsizerel" AND
-   NOT cmake_build_type_tolower STREQUAL "relwithdebinfo")
+   NOT cmake_build_type_tolower STREQUAL "relwithdebinfo" AND
+   NOT cmake_build_type_tolower STREQUAL "relapplesilicon")
     message(FATAL_ERROR "Unknown build type \"${CMAKE_BUILD_TYPE}\". Allowed values are Debug, Release, RelWithDebInfo, and MinSizeRel (case-insensitive).")
 endif()
diff --git a/cmakeconfig.py b/cmakeconfig.py
@@ -43,7 +43,7 @@
   --single                               Enable usage of single precision, where appropriate [default: False].
   --backtrace                            Enable backtrace functionality [default: False].
   --popcnt                               Enable use of intrinsic popcnt [default: False].
-  --type=<TYPE>                          Set the CMake build type (debug, release, relwithdebinfo, minsizerel) [default: release].
+  --type=<TYPE>                          Set the CMake build type (debug, release, relwithdebinfo, minsizerel, relapplesilicon) [default: release].
   --generator=<STRING>                   Set the CMake build system generator [default: Unix Makefiles].
   --show                                 Show CMake command and exit.
   --cmake-executable=<CMAKE_EXECUTABLE>  Set the CMake executable [default: cmake].

diff --git a/lib/external/MurmurHash2.h b/lib/external/MurmurHash2.h
@@ -13,7 +13,8 @@
   FORCE_INLINE uint64_t rotl64 ( uint64_t x, int8_t r ){ return (x << r) | (x >> (64 - r)); }
   FORCE_INLINE uint32_t rotr32 ( uint32_t x, int8_t r ){ return (x >> r) | (x << (32 - r)); }
   FORCE_INLINE uint64_t rotr64 ( uint64_t x, int8_t r ){ return (x >> r) | (x << (64 - r)); }
-  FORCE_INLINE unsigned long long int rdtsc(){ unsigned long long int x; __asm__ volatile ("rdtsc" : "=A" (x)); return x; }
+  // This isn't referenced anywhere, and breaks for non x86 architectures.
+  //FORCE_INLINE unsigned long long int rdtsc(){ unsigned long long int x; __asm__ volatile ("rdtsc" : "=A" (x)); return x; }
 
   #define	ROTL32(x,y)	rotl32(x,y)
   #define ROTL64(x,y)	rotl64(x,y)

diff --git a/test_suite/fci/H2-RHF-cc-pVTZ-Lz_davidson/Lz.fci.in b/test_suite/fci/H2-RHF-cc-pVTZ-Lz_davidson/Lz.fci.in
@@ -16,7 +16,7 @@ fci {
         ndavidson_eigv = 4,
         ndavidson_trialvec = 8,
         davidson_maxsize = 50,
-        davidson_tol = 1e-8,
+        davidson_tol = 1e-14,
         davidson_maxiter = 200,
     },
 }
diff --git a/...dson/benchmark.out.9712b5a3.inp=Lz.fci.in → ...dson/benchmark.out.a03e0bad.inp=Lz.fci.in b/...dson/benchmark.out.9712b5a3.inp=Lz.fci.in → ...dson/benchmark.out.a03e0bad.inp=Lz.fci.in
@@ -4,32 +4,32 @@
  ================================================================
 Version information
 -------------------
-Version       | v1.5-509-gcf07666
-Commit hash   | cf076667d9f540506f71c35bc4a3f4eb8557039d
-Commit author | Brian Zhao
-Commit date   | Mon Jun 20 18:11:28 2022 +0100
-Branch        | main
+Version       | v1.6-33-ga03e0ba
+Commit hash   | a03e0bad8f01c5df93d2f9601870a9cb6ec4e5e4
+Commit author | William Van Benschoten
+Commit date   | Tue Aug 1 14:34:30 2023 -0500
+Branch        | he/apple_silicon
 
 Configuration and build information
 -----------------------------------
-Who compiled             | maf63
-Compilation hostname     | serenity
-Operating system         | Linux-5.4.0-110-generic
-CMake version            | 3.16.3
-CMake generator          | Unix Makefiles
+Who compiled             | ajwt3
+Compilation hostname     | liminal
+Operating system         | Linux-5.4.0-204-generic
+CMake version            | Not built using CMake
+CMake generator          | Not built using CMake
 Build type               | debug
-Configuration time       | 2022-06-20 17:24:01 [UTC]
-Python version           | 3.7.0
-Fortran compiler         | /usr/local/shared/ubuntu-20.04/x86_64/openmpi-4.1.2-gnu/bin/mpif90
-C compiler               | /usr/local/shared/ubuntu-20.04/x86_64/openmpi-4.1.2-gnu/bin/mpicc
-DET_SIZE set to          | 64
+Configuration time       | 2025-01-13 17:32:10 [UTC]
+Python version           | 3.7.12
+Fortran compiler         | mpif90
+C compiler               | mpicc
+DET_SIZE set to          | 32
 POP_SIZE set to          | 64
 dSFMT Mersenne exponent  | 19937
 MPI parallelization      | ON
-MPI launcher             | /usr/local/shared/ubuntu-20.04/x86_64/openmpi-4.1.2-gnu/bin/mpiexec
+MPI launcher             | unknown
 OpenMP parallelization   | OFF
-Lua version              | 5.3.5
-HDF5 library version     | 1.10.4
+Lua version              | unknown
+HDF5 library version     | unknown
 
 Further components
 ------------------
@@ -38,18 +38,18 @@ DISABLE_UUID not defined.  UUID enabled.
 PARALLEL defined.  MPI parallelization enabled.
 DISABLE_MPI3 not defined.  MPI-3 functionality enabled.
 PARALLEL defined. DISABLE_MPI3 and ENABLE_SHMEM_POSIX not defined. MPI-3 in use.
-DISABLE_SCALAPACK defined.  ScaLAPACK disabled.
+DISABLE_SCALAPACK not defined.  ScaLAPACK enabled.
 SINGLE_PRECISION not defined.  Double precision used throughout.
 USE_POPCNT not defined. Internal POPCNT procedure used.
 
 Runtime information
 -------------------
 Hostname:
-    serenity
+    liminal
 Current working dir:
-    /home/maf63/hande-release/test_suite/fci/H2-RHF-cc-pVTZ-Lz_davidson
-Started running on 20/06/2022 at 21:52:12
- Calculation UUID: ca59e489-7410-47bf-b316-0965b40325bb.
+    /home/ajwt3/code/HANDEgithub/test_suite/fci/H2-RHF-cc-pVTZ-Lz_davidson
+Started running on 13/01/2025 at 17:33:30
+ Calculation UUID: 2fb5c441-b524-4035-8920-b20d7a590ee6.
  ================================================================
 
  Input options
@@ -73,7 +73,7 @@ fci {
         ndavidson_eigv = 4,
         ndavidson_trialvec = 8,
         davidson_maxsize = 50,
-        davidson_tol = 1e-8,
+        davidson_tol = 1e-14,
         davidson_maxiter = 200,
     },
 }
@@ -150,8 +150,8 @@ fci {
  Number of electrons: 2
  Number of basis functions: 56
 
- Bit-length of integers used to store determinant bit-strings: 64
- Number of integers used to store determinant bit-strings: 1
+ Bit-length of integers used to store determinant bit-strings: 32
+ Number of integers used to store determinant bit-strings: 2
 
  Symmetry information
  --------------------
@@ -232,7 +232,7 @@ fci {
          "ndavidson_eigv": 4,   
          "ndavidson_trialvec": 8,   
          "davidson_maxsize": 50,   
-         "davidson_tol": 0.00000001,   
+         "davidson_tol": 0.00000000,   
          "hamiltonian_diagonal_only": false,   
      },
      "reference": {
@@ -274,53 +274,70 @@ fci {
  Performing Davidson diagonalisation...
 
  Davidson initialisation done!
- Time taken for Davidson initialisation 0.000064 s
-
+ Time taken for Davidson initialisation 0.000100 s
+ 
  Iteration   Basis size    delta rmsE      Time  
-         1            8   1.279287E+00    0.000698
-         2           16   4.035889E-01    0.006738
-         3           24   3.571042E-02    0.001096
-         4           32   1.505639E-02    0.000988
-         5           40   1.127360E-02    0.001363
-         6           48   4.154376E-03    0.001876
+         1            8   1.279287E+00    0.278100
+         2           16   4.035889E-01    0.000500
+         3           24   3.571042E-02    0.000400
+         4           32   1.505639E-02    0.000400
+         5           40   1.127360E-02    0.044200
+         6           48   4.154376E-03    0.002400
+ Collapsing subspace...
+         7            8   3.589387E-15    0.000200
+         8           16   1.910264E-03    0.000200
+         9           24   4.463168E-04    0.000300
+        10           32   1.155592E-04    0.000400
+        11           40   2.825781E-05    0.000700
+        12           48   6.982884E-06    0.000800
+ Collapsing subspace...
+        13            8   5.031956E-15    0.000100
+        14           16   1.253481E-06    0.000200
+        15           24   6.081517E-07    0.000300
+        16           32   2.122679E-07    0.000400
+        17           40   6.705586E-08    0.000600
+        18           48   2.496507E-08    0.000900
  Collapsing subspace...
-         7            8   4.681733E-15    0.000126
-         8           16   1.910264E-03    0.000386
-         9           24   4.463168E-04    0.000597
-        10           32   1.155592E-04    0.000961
-        11           40   2.825781E-05    0.001317
-        12           48   6.982884E-06    0.001693
+        19            8   4.497774E-15    0.000100
+        20           16   6.300068E-09    0.000200
+        21           24   2.063464E-09    0.000200
+        22           32   7.819654E-10    0.000400
+        23           40   2.682836E-10    0.000700
+        24           48   1.134005E-10    0.000800
  Collapsing subspace...
-        13            8   3.992174E-15    0.000117
-        14           16   1.253481E-06    0.000346
-        15           24   6.081517E-07    0.000523
-        16           32   2.122679E-07    0.000862
-        17           40   6.705586E-08    0.001234
-        18           48   2.496258E-08    0.001662
+        25            8   1.986803E-15    0.000100
+        26           16   2.479705E-11    0.000200
+        27           24   8.199489E-12    0.000200
+        28           32   3.075607E-12    0.000500
+        29           40   1.496427E-12    0.000600
+        30           48   5.564908E-13    0.000800
  Collapsing subspace...
-        19            8   2.592436E-15    0.000114
-        20           16   6.301052E-09    0.000333
- Eigenvalue tolerance of 1.0000E-08 reached, printing results...
+        31            8   3.523100E-15    0.000100
+        32           16   1.183845E-13    0.000200
+        33           24   3.765768E-14    0.000300
+        34           32   1.353482E-14    0.000400
+        35           40   6.577544E-15    0.000600
+ Eigenvalue tolerance of 1.0000E-14 reached, printing results...
  Davidson diagonalisation results
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
   State     Energy
       1    -1.145758846665
-      2    -0.622330308093
-      3    -0.538417004192
-      4    -0.365633761257
+      2    -0.622330308365
+      3    -0.538417004463
+      4    -0.365633764501
 
  Timing breakdown
  ----------------
 
  Wall time for each calculation section (seconds):
 
  Generic system initialisation: .06
- FCI calculation              : .27
+ FCI calculation              : .05
 
  ================================================================
- Finished running on 20/06/2022 at 21:52:12
- Wall time (seconds):                               0.09
- CPU time (per processor, seconds):                 0.33
+ Finished running on 13/01/2025 at 17:33:30
+ Wall time (seconds):                               0.42
+ CPU time (per processor, seconds):                 0.11
  ================================================================
 
diff --git a/test_suite/userconfig b/test_suite/userconfig
@@ -11,6 +11,6 @@ skip_cmd_template = tc.skip tc.args tc.test tc.error
 
 [user]
 diff = vimdiff
-benchmark = 9712b5a3
+benchmark = a03e0bad  9712b5a3
 tolerance = (1.e-10, 1.e-10, None, False)