minor fix

include/backend/hpl_backendCPU.h

@@ -11,7 +11,7 @@ extern "C" {
 #include "hpl_panel.h"
 }
 
-#define HPL_PRINT_INFO
+//#define HPL_PRINT_INFO
 #ifdef HPL_PRINT_INFO
 #define CPUInfo(msg, ...)                           \
 {                                                   \

include/backend/hpl_backendCommon.h

@@ -11,7 +11,7 @@ namespace HPL
 {
   template<typename Func, typename... Args> 
   constexpr auto dispatch(Func&& func, Args... args) {
-    func(args...);
+    return func(args...);
   }
 };
 

include/backend/hpl_backendHIP.h

@@ -51,11 +51,18 @@ extern "C" {
         exit(EXIT_FAILURE);                           \
     }
 
+
 #define GPUError(...) GPUInfo(__VA_ARGS__)
+
+//#define HPL_PRINT_INFO
+#ifdef HPL_PRINT_INFO
 #define GPUInfo(msg, ...)                             \
 {                                                     \
   printf("INFO\t %-15s\t" msg "\n", ##__VA_ARGS__);   \
 }
+#else
+#define GPUInfo(msg, ...)
+#endif
 
 
 

include/backend/hpl_backendWrapper.h

@@ -14,7 +14,7 @@ extern "C" {
 #include "hpl_blas.h"
 #include "hpl_panel.h"
 
-enum HPL_TARGET {T_DEFAULT, T_CPU, T_HIP};
+enum HPL_TARGET {T_DEFAULT, T_CPU, T_HIP, T_TEMPO};
 
 void HPL_BE_init(size_t, enum HPL_TARGET);
 

src/panel/HPL_pdpanel_init.cpp

@@ -210,7 +210,7 @@ void HPL_pdpanel_init
       { nu = nq - JB; lwork += JB * Mmax( 0, nu ); }
 
       //Adil
-      HPL_BE_malloc((void**)&PANEL->WORK, (size_t)(lwork) * sizeof( double ), T_DEFAULT);
+      HPL_BE_malloc((void**)&PANEL->WORK, (size_t)(lwork) * sizeof( double ), T_TEMPO);
       //if( !( PANEL->WORK = (void *)malloc( (size_t)(lwork) * sizeof( double ) ) ) )
       if(!( PANEL->WORK ))
       {
@@ -244,7 +244,7 @@ void HPL_pdpanel_init
       }
 
       //Adil
-      HPL_BE_malloc((void**)&PANEL->WORK, (size_t)(lwork) * sizeof( double ), T_DEFAULT);
+      HPL_BE_malloc((void**)&PANEL->WORK, (size_t)(lwork) * sizeof( double ), T_TEMPO);
       //if( !( PANEL->WORK = (void *)malloc( (size_t)(lwork) * sizeof( double ) ) ) )
       if(!(PANEL->WORK))
       {
@@ -341,7 +341,7 @@ void HPL_pdpanel_init
    }
 
    //Adil
-   HPL_BE_malloc((void**)&PANEL->IWORK, (size_t)(lwork) * sizeof( int ), T_DEFAULT);
+   HPL_BE_malloc((void**)&PANEL->IWORK, (size_t)(lwork) * sizeof( int ), T_TEMPO);
    //PANEL->IWORK = (int *)malloc( (size_t)(lwork) * sizeof( int ) );
 
    if( PANEL->IWORK == NULL )

src/panel/HPL_pdpanel_new.cpp

@@ -140,7 +140,7 @@ void HPL_pdpanel_new
  * Allocate the panel structure - Check for enough memory
  */
    //Adil
-   HPL_BE_malloc((void**)&p, sizeof( HPL_T_panel ), T_DEFAULT);
+   HPL_BE_malloc((void**)&p, sizeof( HPL_T_panel ), T_TEMPO);
    /*if( !( p = (HPL_T_panel *)malloc( sizeof( HPL_T_panel ) ) ) )*/
    if( !( p ))
    {

src/pauxil/HPL_pdlange.cpp

@@ -204,7 +204,7 @@ double HPL_pdlange
       if( mp > 0 )
       {
          //Adil
-         HPL_BE_malloc((void**)&work, (size_t)(mp) * sizeof( double ), T_DEFAULT);
+         HPL_BE_malloc((void**)&work, (size_t)(mp) * sizeof( double ), T_TEMPO);
          /*work = (double*)malloc( (size_t)(mp) * sizeof( double ) );*/
          if( work == NULL )
          { HPL_pabort( __LINE__, "HPL_pdlange", "Memory allocation failed" ); }
@@ -225,9 +225,11 @@ double HPL_pdlange
 /*       
  * Find maximum sum of rows for inf-norm
  */      
-         if( mycol == 0 )
-         { v0 = work[HPL_idamax( mp, work, 1 )]; v0 = Mabs( v0 ); }
-         if( work ) HPL_BE_free((void**)&work, T_DEFAULT);
+         if( mycol == 0 ){
+            v0 = work[HPL_BE_idamax( mp, work, 1, T_TEMPO)];
+            v0 = Mabs( v0 ); 
+         }
+         if( work ) HPL_BE_free((void**)&work, T_TEMPO);
          /*if( work ) free( work );*/
       }
 /*

src/pfact/HPL_pdfact.cpp

@@ -120,7 +120,7 @@ void HPL_pdfact
    align = PANEL->algo->align;
 
    //Adil
-   HPL_BE_malloc((void**)&vptr, ( (size_t)(align) + (size_t)(((4+((unsigned int)(jb) << 1)) << 1) )) * sizeof(double), T_DEFAULT);
+   HPL_BE_malloc((void**)&vptr, ( (size_t)(align) + (size_t)(((4+((unsigned int)(jb) << 1)) << 1) )) * sizeof(double), T_TEMPO);
    /*vptr  = (void *)malloc( ( (size_t)(align) + (size_t)(((4+((unsigned int)(jb) << 1)) << 1) )) * sizeof(double) );*/
    if( vptr == NULL )
    { HPL_pabort( __LINE__, "HPL_pdfact", "Memory allocation failed" ); }
@@ -131,7 +131,7 @@ void HPL_pdfact
                        ((size_t)(align) * sizeof(double) ) ) );
 
    //Adil
-   if( vptr ) HPL_BE_free((void**)&vptr, T_DEFAULT);
+   if( vptr ) HPL_BE_free((void**)&vptr, T_TEMPO);
    //if( vptr ) free( vptr );
 
    PANEL->A   = Mptr( PANEL->A, 0, jb, PANEL->lda );

src/pfact/HPL_pdmxswp.cpp

@@ -163,13 +163,13 @@ void HPL_pdmxswp
       lda = PANEL->lda;
       //Adil
       HPL_BE_dcopy( n0, Mptr( PANEL->A, II+(int)(WORK[1]), 0, lda ), lda,
-                 Wmx, 1, T_DEFAULT);
+                 Wmx, 1, T_TEMPO);
       /*HPL_dcopy( n0, Mptr( PANEL->A, II+(int)(WORK[1]), 0, lda ), lda,
                  Wmx, 1 );*/
       if( myrow == icurrow )
       {
          //Adil
-         HPL_BE_dcopy( n0, Mptr( PANEL->A, II, 0, lda ), lda, A0, 1, T_DEFAULT); 
+         HPL_BE_dcopy( n0, Mptr( PANEL->A, II, 0, lda ), lda, A0, 1, T_TEMPO); 
          /*HPL_dcopy( n0, Mptr( PANEL->A, II, 0, lda ), lda, A0, 1 ); */
       }
    }

src/pfact/HPL_pdpanrlT.cpp

@@ -182,11 +182,11 @@ void HPL_pdpanrlT
  */
       if( WORK[0] != HPL_rzero )
          //Adil
-         HPL_BE_dscal(  Mm1, HPL_rone / WORK[0], Acur, 1 , T_DEFAULT);
+         HPL_BE_dscal(  Mm1, HPL_rone / WORK[0], Acur, 1 , T_TEMPO);
          /*HPL_dscal( Mm1, HPL_rone / WORK[0], Acur, 1 );*/
 
       //Adil
-      HPL_BE_daxpy( Mm1, -(*(Mptr( L1, jj+1, jj, n0 ))), Acur, 1, Anxt, 1, T_DEFAULT);
+      HPL_BE_daxpy( Mm1, -(*(Mptr( L1, jj+1, jj, n0 ))), Acur, 1, Anxt, 1, T_TEMPO);
       /*HPL_daxpy( Mm1, -(*(Mptr( L1, jj+1, jj, n0 ))), Acur, 1, Anxt, 1 );*/
       HPL_dlocmax( PANEL, Mm1, iip1, jj+1, WORK );
 
@@ -214,7 +214,7 @@ void HPL_pdpanrlT
          //Adil
          HPL_BE_dger( HplColumnMajor, Mm1, Nm1-1, -HPL_rone, Acur, 1,
                    Mptr( L1, jj+2, jj, n0 ), 1, Mptr( Anxt, 0, 1, lda ),
-                   lda, T_DEFAULT);
+                   lda, T_TEMPO);
          /*HPL_dger( HplColumnMajor, Mm1, Nm1-1, -HPL_rone, Acur, 1,
                    Mptr( L1, jj+2, jj, n0 ), 1, Mptr( Anxt, 0, 1, lda ),
                    lda );*/
@@ -232,7 +232,7 @@ void HPL_pdpanrlT
    HPL_dlocswpT( PANEL,    ii, jj, WORK );
    if( WORK[0] != HPL_rzero )
       //Adil
-      HPL_BE_dscal( Mm1, HPL_rone / WORK[0], Mptr( A, iip1, jj, lda ), 1, T_DEFAULT);
+      HPL_BE_dscal( Mm1, HPL_rone / WORK[0], Mptr( A, iip1, jj, lda ), 1, T_TEMPO);
       /*HPL_dscal( Mm1, HPL_rone / WORK[0], Mptr( A, iip1, jj, lda ), 1 );*/
 #ifdef HPL_CALL_VSIPL
 /*

src/pgesv/HPL_pdgesv0.cpp

@@ -115,7 +115,7 @@ void HPL_pdgesv0
  * Allocate a panel list of length 1 - Allocate panel[0] resources
  */
    //Adil
-   HPL_BE_malloc((void**)&panel, sizeof( HPL_T_panel * ), T_DEFAULT);
+   HPL_BE_malloc((void**)&panel, sizeof( HPL_T_panel * ), T_TEMPO);
    /*panel = (HPL_T_panel **)malloc( sizeof( HPL_T_panel * ) );*/
    if( panel == NULL )
    { HPL_pabort( __LINE__, "HPL_pdgesv0", "Memory allocation failed" ); }
@@ -141,7 +141,7 @@ void HPL_pdgesv0
  * Release panel resources - re-initialize panel data structure
  */
       //Adil
-      (void) HPL_BE_panel_free( panel[0], T_DEFAULT);
+      (void) HPL_BE_panel_free( panel[0], T_TEMPO);
       //(void) HPL_pdpanel_free( panel[0] );
       HPL_pdpanel_init( GRID, ALGO, n, n+1, jb, A, j, j, tag, panel[0] );
 /*
@@ -163,11 +163,11 @@ void HPL_pdgesv0
  * Release panel resources and panel list
  */
    //Adil
-   (void) HPL_BE_panel_disp((void**)&panel[0], T_DEFAULT);
+   (void) HPL_BE_panel_disp((void**)&panel[0], T_TEMPO);
    /*(void) HPL_pdpanel_disp( &panel[0] );*/
 
    //Adil
-   if( panel ) HPL_BE_free((void**)&panel, T_DEFAULT);
+   //if( panel ) HPL_BE_free((void**)&panel, T_TEMPO);
    //if( panel ) free( panel);
 /*
  * End of HPL_pdgesv0

src/pgesv/HPL_pdtrsv.cpp

@@ -160,7 +160,7 @@ void HPL_pdtrsv
    if( Anp > 0 )
    {
       //Adil
-      HPL_BE_malloc((void**)&W, (size_t)(Mmin( n1, Anp )) * sizeof( double ), T_DEFAULT);
+      HPL_BE_malloc((void**)&W, (size_t)(Mmin( n1, Anp )) * sizeof( double ), T_TEMPO);
       /*W = (double*)malloc( (size_t)(Mmin( n1, Anp )) * sizeof( double ) );*/
       if( W == NULL )
       { HPL_pabort( __LINE__, "HPL_pdtrsv", "Memory allocation failed" ); }
@@ -179,12 +179,12 @@ void HPL_pdtrsv
       {
          //Adil
          HPL_BE_dtrsv(HplColumnMajor, HplUpper, HplNoTrans, HplNonUnit,
-                    kb, Aptr+Anp, lda, XC+Anp, 1, T_DEFAULT);
+                    kb, Aptr+Anp, lda, XC+Anp, 1, T_TEMPO);
          /*HPL_dtrsv( HplColumnMajor, HplUpper, HplNoTrans, HplNonUnit,
                     kb, Aptr+Anp, lda, XC+Anp, 1 );*/
 
          //Adil
-         HPL_BE_dcopy(kb, XC+Anp, 1, Xd, 1, T_DEFAULT);                    
+         HPL_BE_dcopy(kb, XC+Anp, 1, Xd, 1, T_TEMPO);                    
          /*HPL_dcopy( kb, XC+Anp, 1, Xd, 1 );*/
       }
    }
@@ -310,7 +310,7 @@ void HPL_pdtrsv
                             Ccomm );
 
    //Adil
-   if( Wfr ) HPL_BE_free((void**)&W, T_DEFAULT);
+   if( Wfr ) HPL_BE_free((void**)&W, T_TEMPO);
    /*if( Wfr  ) free( W  );*/
 #ifdef HPL_DETAILED_TIMING
    HPL_ptimer( HPL_TIMING_PTRSV );

src/pgesv/HPL_pdupdateTT.cpp

@@ -226,7 +226,7 @@ void HPL_pdupdateTT
 #endif
          //Adil
          HPL_BE_dtrsm(HplColumnMajor, HplLeft, HplUpper, HplTrans,
-                    HplUnit, jb, nn, HPL_rone, L1ptr, jb, Aptr, lda, T_DEFAULT);
+                    HplUnit, jb, nn, HPL_rone, L1ptr, jb, Aptr, lda, T_TEMPO);
 
          /*HPL_dtrsm( HplColumnMajor, HplLeft, HplUpper, HplTrans,
                     HplUnit, jb, nn, HPL_rone, L1ptr, jb, Aptr, lda );*/
@@ -248,7 +248,7 @@ void HPL_pdupdateTT
          //Adil
          HPL_BE_dgemm(HplColumnMajor, HplNoTrans, HplNoTrans, mp, nn,
                     jb, -HPL_rone, L2ptr, ldl2, Aptr, lda, HPL_rone,
-                    Mptr( Aptr, jb, 0, lda ), lda, T_DEFAULT);
+                    Mptr( Aptr, jb, 0, lda ), lda, T_TEMPO);
          /*HPL_dgemm( HplColumnMajor, HplNoTrans, HplNoTrans, mp, nn,
                     jb, -HPL_rone, L2ptr, ldl2, Aptr, lda, HPL_rone,
                     Mptr( Aptr, jb, 0, lda ), lda );*/

testing/backend/HPL_backendCPU.cpp

@@ -23,18 +23,31 @@ int CPU::panel_disp(HPL_T_panel **ptr)
     return HPL_pdpanel_disp( ptr );
 }
 
+void printMat(const int M, const int N, const int LDA, const double *A)
+{
+    // Last row is the vector b
+    for(int y=0;y<M+1; y++){
+        for(int x=0;x<N-1; x++){
+            int index = x+y*LDA;
+            printf("%-4d:%-8lf\t", index, A[index]);
+        }
+        printf("\n");
+    }
+}
+
 void CPU::matgen(const HPL_T_grid *GRID, const int M, const int N,
                  const int NB, double *A, const int LDA,
                  const int ISEED)
 {
     CPUInfo("%-25s %-8d%-8d \t%-5s", "[Generate matrix]", "With A of (R:C)", M, N, "CPU");
     HPL_pdmatgen(GRID, M, N, NB, A, LDA, ISEED);
+    //printMat(M,N,LDA,A);
 }
 
 int CPU::idamax(const int N, const double *DX, const int INCX)
 {
     CPUInfo("%-25s %-17d \t%-5s", "[IDAMAX]", "With X of (R)", N, "CPU");
-    HPL_idamax( N, DX, INCX );
+    return HPL_idamax( N, DX, INCX );
 }
 
 void CPU::daxpy(const int N, const double DA, const double *DX, const int INCX, double *DY, 

testing/backend/HPL_backendHIP.cpp

@@ -73,6 +73,18 @@ int HIP::panel_disp(HPL_T_panel **ptr)
     return( err );
 }
 
+void gPrintMat(const int M, const int N, const int LDA, const double *A)
+{
+    // Last row is the vector b
+    for(int y=0;y<M+1; y++){
+        for(int x=0;x<N-1; x++){
+            int index = x+y*LDA;
+            printf("%-4d:%-8lf\t", index, A[index]);
+        }
+        printf("\n");
+    }
+}
+
 void HIP::matgen(const HPL_T_grid *GRID, const int M, const int N,
                  const int NB, double *A, const int LDA,
                  const int ISEED)
@@ -94,6 +106,7 @@ void HIP::matgen(const HPL_T_grid *GRID, const int M, const int N,
     //TODO: generate numbers in this range (-0.5, 0.5]
     ROCRAND_CHECK_STATUS(rocrand_generate_normal_double(generator, A, mp*nq, 0, 0.1));
     ROCRAND_CHECK_STATUS(rocrand_destroy_generator(generator));
+    //gPrintMat(M,N,LDA,A);
 }
 
 int HIP::idamax(const int N, const double *DX, const int INCX)
@@ -189,7 +202,7 @@ __global__ void
 _dlacpy(const int M, const int N, const double *A, const int LDA,
         double *B, const int LDB)
 {
-
+ 
 }
 
 /*
@@ -225,4 +238,5 @@ void HIP::atcpy(const int M, const int N, const double *A, const int LDA,
     dim3 grid_size((M+TILE_DIM-1)/TILE_DIM, (N+TILE_DIM-1)/TILE_DIM);
     dim3 block_size(TILE_DIM, BLOCK_ROWS);
     _dlatcpy<<<grid_size, block_size, 0, 0>>>(M, N, A, LDA, B, LDB);
-}
+}
+

testing/backend/HPL_backendWrapper.cpp

@@ -4,16 +4,14 @@
 #include "backend/hpl_backendCommon.h"
 
 
-
 extern "C" {
-
    /*
    * Initialize the Target device
    */
    void HPL_BE_init(size_t num_devices, enum HPL_TARGET TR)
    {
       switch(TR) {
-         case T_CPU :
+         case T_CPU:
             DO_NOTHING();
             break;
          case T_HIP:
@@ -124,13 +122,13 @@ extern "C" {
    {
       switch(TR) {
          case T_CPU :
-            HPL::dispatch(CPU::idamax, N, DX, INCX);
+            return HPL::dispatch(CPU::idamax, N, DX, INCX);
             break;
          case T_HIP:
-            HPL::dispatch(HIP::idamax, N, DX, INCX);
+            return HPL::dispatch(HIP::idamax, N, DX, INCX);
             break;
          default:
-            HPL::dispatch(CPU::idamax, N, DX, INCX);
+            return HPL::dispatch(CPU::idamax, N, DX, INCX);
       }
    }