Add support for linear systems stored in more partitions than the num…

…ber of MPI ranks (#66)

* Add umpire memory pool options to general
* Add IJVectorReadMultipartBinary
* Add IJMatrixReadMultipartBinary
* Add HYPRE_ClearAllErrors call to LS build and preconditioner setup
* Fix dofmap array read for distributed runs
* Update ErrorMsgAdd to be variadic

.gitignore

@@ -4,6 +4,7 @@
 *.o
 *.lo
 *.btr
+*.csv
 .deps
 .dirstamp
 *.~HEAD~*

Makefile.am

@@ -3,9 +3,10 @@ hypredrive_SOURCES = src/main.c
 hypredrive_LDADD = libHYPREDRV.la
 
 lib_LTLIBRARIES = libHYPREDRV.la
-libHYPREDRV_la_SOURCES = src/error.c src/info.c src/stats.c src/utils.c src/containers.c src/yaml.c src/field.c\
-                         src/linsys.c src/cheby.c src/amg.c src/fsai.c src/ilu.c src/mgr.c src/precon.c src/pcg.c\
-                         src/gmres.c src/fgmres.c src/bicgstab.c src/solver.c src/args.c src/HYPREDRV.c
+libHYPREDRV_la_SOURCES = src/error.c src/info.c src/stats.c src/utils.c src/containers.c src/yaml.c\
+                         src/field.c src/vector.c src/matrix.c src/linsys.c src/cheby.c src/amg.c\
+                         src/fsai.c src/ilu.c src/mgr.c src/precon.c src/pcg.c src/gmres.c src/fgmres.c\
+                         src/bicgstab.c src/solver.c src/args.c src/HYPREDRV.c
 libHYPREDRV_la_CFLAGS = $(AM_CPPFLAGS) -fvisibility=hidden
 libHYPREDRV_la_LDFLAGS = -version-info 0:1:0
 

configure.ac

@@ -182,6 +182,10 @@ dnl Check for libHYPRE
 AC_CHECK_LIB([HYPRE], [HYPRE_IJMatrixCreate], [],
              [AC_MSG_ERROR([HYPRE library not found or not usable])])
 
+dnl Check for HYPRE headers
+AC_CHECK_HEADER([HYPRE_config.h], [],
+                [AC_MSG_ERROR([HYPRE headers not found or not usable. Please ensure the correct path is specified with --with-hypre-include and the headers are installed.])])
+
 dnl Check for minimum HYPRE version
 AC_LANG_PUSH([C])
 AC_MSG_CHECKING([for HYPRE version])
@@ -193,10 +197,6 @@ AC_COMPILE_IFELSE([AC_LANG_PROGRAM([#include <HYPRE_config.h>],
                   [AC_MSG_ERROR([HYPRE library version is too old. Use hypre >= 2.31.0])])
 AC_LANG_POP([C])
 
-dnl Check for HYPRE headers
-AC_CHECK_HEADER([HYPRE_config.h], [],
-                [AC_MSG_ERROR([HYPRE headers not found or not usable. Please ensure the correct path is specified with --with-hypre-include and the headers are installed.])])
-
 AC_CONFIG_MACRO_DIRS([m4])
 AC_CONFIG_HEADERS([HYPREDRV_config.h])
 AC_CHECK_HEADERS([stdint.h])

docs/usrman-src/input_file_structure.rst

@@ -43,6 +43,23 @@ The ``general`` section contains global settings that apply to the entire execut
   repeated. Useful for benchmarking and profiling. The default value for this parameter is
   `1`.
 
+- ``dev_pool_size`` - Initial size of the umpire's device memory pool. This parameter is
+  neglected when hypre is *not* configured with umpire support. The default value for this
+  parameter is `8 GB`.
+
+- ``uvm_pool_size`` - Initial size of the umpire's unified virtual memory pool. This
+  parameter is neglected when hypre is *not* configured with umpire support. The default
+  value for this parameter is `8 GB`.
+
+- ``host_pool_size`` - Initial size of the umpire's host memory pool. This parameter is
+  neglected when hypre is *not* configured with umpire support. The default value for this
+  parameter is `8 GB`.
+
+- ``pinned_pool_size`` - Initial size of the umpire's host memory pool. This parameter is
+  neglected when hypre is *not* configured with umpire support. The default value for this
+  parameter is `512 MB`.
+
+
 An example code block for the ``general`` section is given below:
 
 .. code-block:: yaml
@@ -52,6 +69,10 @@ An example code block for the ``general`` section is given below:
       statistics: yes
       use_millisec: no
       num_repetitions: 1
+      dev_pool_size: 8.0
+      uvm_pool_size: 8.0
+      host_pool_size: 8.0
+      pinned_pool_size: 0.5
 
 Linear System
 -------------

include/args.h

@@ -23,6 +23,10 @@ typedef struct input_args_struct {
    int           warmup;
    int           statistics;
    int           num_repetitions;
+   double        dev_pool_size;
+   double        uvm_pool_size;
+   double        host_pool_size;
+   double        pinned_pool_size;
 
    LS_args       ls;
 

include/containers.h

@@ -42,7 +42,12 @@ typedef struct IntArray_struct
 {
    int      *data;
    size_t    size;
-   size_t    num_unique_entries;
+
+   int      *unique_data;
+   size_t    unique_size;
+
+   int      *g_unique_data;
+   size_t    g_unique_size;
 } IntArray;
 
 IntArray* IntArrayCreate(size_t);

include/error.h

@@ -12,6 +12,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <stdarg.h>
 #include <stdint.h>
 #include <stdbool.h>
 
@@ -33,7 +34,8 @@ typedef enum ErrorCode_enum
    ERROR_INVALID_SOLVER         = 0x00001000, // 13th bit
    ERROR_INVALID_PRECON         = 0x00002000, // 14th bit
    ERROR_FILE_NOT_FOUND         = 0x00004000, // 15th bit
-   ERROR_UNKNOWN_HYPREDRV_OBJ   = 0x00008000, // 16th bit
+   ERROR_FILE_UNEXPECTED_ENTRY  = 0x00008000, // 16th bit
+   ERROR_UNKNOWN_HYPREDRV_OBJ   = 0x00010000, // 17th bit
    ERROR_UNKNOWN_TIMING         = 0x20000000, // 29th bit
    ERROR_UNKNOWN                = 0x40000000  // 30th bit
 } ErrorCode;
@@ -49,7 +51,7 @@ bool DistributedErrorCodeActive(MPI_Comm);
 /*******************************************************************************
  *******************************************************************************/
 
-void ErrorMsgAdd(const char*);
+void ErrorMsgAdd(const char*, ...);
 void ErrorMsgAddCodeWithCount(ErrorCode, const char*);
 void ErrorMsgAddMissingKey(const char*);
 void ErrorMsgAddExtraKey(const char*);

include/linsys.h

@@ -64,4 +64,7 @@ void LinearSystemComputeResidualNorm(HYPRE_IJMatrix, HYPRE_IJVector, HYPRE_IJVec
 long long int LinearSystemMatrixGetNumRows(HYPRE_IJMatrix);
 long long int LinearSystemMatrixGetNumNonzeros(HYPRE_IJMatrix);
 
+void IJVectorReadMultipartBinary(const char*, MPI_Comm, uint64_t, HYPRE_IJVector*);
+void IJMatrixReadMultipartBinary(const char*, MPI_Comm, uint64_t, HYPRE_IJMatrix*);
+
 #endif /* LINSYS_HEADER */

include/utils.h

@@ -19,20 +19,43 @@
 
 char* StrToLowerCase(char*);
 char* StrTrim(char*);
-int CheckBinaryDataExists(const char* prefix);
+int CheckBinaryDataExists(const char*);
+int CountNumberOfPartitions(const char*);
 int ComputeNumberOfDigits(int);
 void SplitFilename(const char*, char**, char**);
 void CombineFilename(const char*, const char*, char**);
 
 /*******************************************************************************
  *******************************************************************************/
 
+/* Helper macros to check if HYPRE_DEVELOP_NUMBER is defined */
+#if defined(HYPRE_DEVELOP_NUMBER)
+  #define HYPRE_DEVELOP_NUMBER_EXISTS 1
+#else
+  #define HYPRE_DEVELOP_NUMBER_EXISTS 0
+#endif
+
+/* Check if HYPRE_DEVELOP_NUMBER is greater than a given value */
+#define HYPRE_DEVELOP_NUMBER_GT(develop) \
+    (HYPRE_DEVELOP_NUMBER > (develop))
+
+/* Check if HYPRE_RELEASE_NUMBER is greater than a given value */
+#define HYPRE_RELEASE_NUMBER_GT(release) \
+    (HYPRE_RELEASE_NUMBER > (release))
+
+/* Check for a specific hypre release number and whether
+   HYPRE_DEVELOP_NUMBER is defined and greater than a given value */
+#define HYPRE_RELEASE_NUMBER_EQ_AND_DEVELOP_NUMBER_GT(release, develop) \
+    (HYPRE_RELEASE_NUMBER == (release) &&\
+     HYPRE_DEVELOP_NUMBER_EXISTS &&\
+     HYPRE_DEVELOP_NUMBER_GT(develop))
+
 /* Check for minimum HYPRE version in order to allow certain features */
 #define HYPRE_CHECK_MIN_VERSION(release, develop) \
-    (HYPRE_RELEASE_NUMBER > (release) || \
-    (HYPRE_RELEASE_NUMBER == (release) && \
-     (defined(HYPRE_DEVELOP_NUMBER) && HYPRE_DEVELOP_NUMBER > (develop))))
+    (HYPRE_RELEASE_NUMBER_GT(release) ||\
+     HYPRE_RELEASE_NUMBER_EQ_AND_DEVELOP_NUMBER_GT(release, develop))
 
+#define GB_TO_BYTES (1 << 30)
 #define MAX_DIVISOR_LENGTH 84
 #define PRINT_DASHED_LINE(_l) for (size_t i = 0; i < (_l); i++) { printf("-"); } printf("\n");
 #define PRINT_EQUAL_LINE(_l)  for (size_t i = 0; i < (_l); i++) { printf("="); } printf("\n");

include/yaml.h

@@ -99,7 +99,7 @@ char* YAMLnodeFindChildValueByKey(YAMLnode*, const char*);
          { \
             _node->mapped_val = (char*) malloc(_length * sizeof(char)); \
          } \
-         else if (_length > strlen(_node->mapped_val)) \
+         else if (_length > (int) strlen(_node->mapped_val)) \
          { \
             _node->mapped_val = (char*) realloc(_node->mapped_val, _length * sizeof(char)); \
          } \

src/HYPREDRV.c

@@ -189,7 +189,23 @@ HYPREDRV_SetGlobalOptions(HYPREDRV_t obj)
       {
          HYPRE_SetMemoryLocation(HYPRE_MEMORY_DEVICE);
          HYPRE_SetExecutionPolicy(HYPRE_EXEC_DEVICE);
-         HYPRE_SetSpGemmUseVendor(0);
+         HYPRE_SetSpGemmUseVendor(0); // TODO: Control this via input option
+         HYPRE_SetSpMVUseVendor(0);   // TODO: Control this via input option
+
+#if defined(HYPRE_USING_UMPIRE)
+        /* Setup Umpire pools */
+        HYPRE_SetUmpireDevicePoolName("DEVICE_POOL");
+        HYPRE_SetUmpireDevicePoolSize(obj->iargs->dev_pool_size);
+
+        HYPRE_SetUmpireUMPoolName("UM_POOL");
+        HYPRE_SetUmpireUMPoolSize(obj->iargs->uvm_pool_size);
+
+        HYPRE_SetUmpireHostPoolName("HOST_POOL");
+        HYPRE_SetUmpireHostPoolSize(obj->iargs->host_pool_size);
+
+        HYPRE_SetUmpirePinnedPoolName("PINNED_POOL");
+        HYPRE_SetUmpirePinnedPoolSize(obj->iargs->pinned_pool_size);
+#endif
       }
       else
       {
@@ -261,6 +277,7 @@ HYPREDRV_LinearSystemBuild(HYPREDRV_t obj)
          printf("Solving linear system #%d ", StatsGetLinearSystemID());
          printf("with %lld rows and %lld nonzeros...\n", num_rows, num_nonzeros);
       }
+      HYPRE_ClearAllErrors();
    }
    else
    {
@@ -454,6 +471,7 @@ HYPREDRV_LinearSolverSetup(HYPREDRV_t obj)
          SolverSetup(obj->iargs->precon_method, obj->iargs->solver_method,
                      obj->precon, obj->solver, obj->mat_M, obj->vec_b, obj->vec_x);
       }
+      HYPRE_ClearAllErrors();
    }
    else
    {

src/args.c

@@ -20,9 +20,13 @@ InputArgsCreate(input_args **iargs_ptr)
    input_args *iargs = (input_args*) malloc(sizeof(input_args));
 
    /* Set general default options */
-   iargs->warmup = 0;
-   iargs->num_repetitions = 1;
-   iargs->statistics = 1;
+   iargs->warmup           = 0;
+   iargs->num_repetitions  = 1;
+   iargs->statistics       = 1;
+   iargs->dev_pool_size    = 8.0 * GB_TO_BYTES;
+   iargs->uvm_pool_size    = 8.0 * GB_TO_BYTES;
+   iargs->host_pool_size   = 8.0 * GB_TO_BYTES;
+   iargs->pinned_pool_size = 0.5 * GB_TO_BYTES;
 
    /* Set default Linear System options */
    LinearSystemSetDefaultArgs(&iargs->ls);
@@ -126,7 +130,23 @@ InputArgsParseGeneral(input_args *iargs, YAMLtree *tree)
       }
       else if (!strcmp(child->key, "num_repetitions"))
       {
-         iargs->num_repetitions = (HYPRE_Int) atoi(child->val);
+         iargs->num_repetitions = atoi(child->val);
+      }
+      else if (!strcmp(child->key, "dev_pool_size"))
+      {
+         iargs->dev_pool_size = GB_TO_BYTES * atof(child->val);
+      }
+      else if (!strcmp(child->key, "uvm_pool_size"))
+      {
+         iargs->uvm_pool_size = GB_TO_BYTES * atof(child->val);
+      }
+      else if (!strcmp(child->key, "host_pool_size"))
+      {
+         iargs->host_pool_size = GB_TO_BYTES * atof(child->val);
+      }
+      else if (!strcmp(child->key, "pinned_pool_size"))
+      {
+         iargs->pinned_pool_size = GB_TO_BYTES * atof(child->val);
       }
       else
       {