Commit f17ccd7f authored by Elwasif, Wael's avatar Elwasif, Wael
Browse files

Nerge done

parents 093b948f b8515846

README.md

+60 −0
Original line number Diff line number Diff line 


# DMRG++PluginSc



## How to build



Download and build PsimagLite

```BASH

$ git clone https://code.ornl.gov/gonzalo_3/PsimagLite

$ cd PsimagLite/lib

$ git checkout features

$ ./configure.pl



# to take advantage of 4 CPU cores 

$ make -j 4

```



Download and build DMRG++PluginSc (this repo)

```BASH

$ git clone https://code.ornl.gov/e6d/dmrgppPluginSc

$ cd dmrgppPluginSc/src



# create a make.inc file by modifying one of the existing ones

# for example to take advantage of OpenMP on multiple CPU cores

# cp make.inc.openmp   make.inc

# or take advantage of MAGMA library on GPU

# cp make.inc.magma   make.inc

$ cp make.inc.openmp  make.inc





$ make -j 4

```



Download DMRG++

```BASH

$ cd ../../

$ git clone https://code.ornl.gov/gonzalo_3/dmrgpp

$ cd dmrgpp/

$ git checkout features

$ cd src

```

Create a file myconfig.psiTag with the following content



```

flavor myflavor = (

< flavor production

< dependency PluginSc

)



default flavor = myflavor

```

and then

```BASH

$ ./configure.pl -c myconfig.psiTag

$ make -j 4

```



To DMRG++ inputs, please add BatchedGemm to SolverOptions, like

```

SolverOptions=BatchedGemm,...

```

src/BatchedGemm.cpp

+7 −0
Original line number Diff line number Diff line
@@ -9,7 +9,10 @@ 
#include <math.h>

#include <stdio.h>



#ifdef SHARE_GPU

sem_t* BatchedGemm<FpType>:: gpu_access = NULL;       

#endif



extern "C" void get_current_time_with_ms (time_t *sec , long *millisec);
@@ -37,6 +40,7 @@ BatchedGemm<FpType>::BatchedGemm(int noperator, 
	xy_patch_start_ = 0;





#ifdef SHARE_GPU    

        if (gpu_access == NULL){

            gpu_access = sem_open("/dmrg_gpu", O_CREAT, 0666, 1); 

            if(gpu_access == SEM_FAILED) {
@@ -50,6 +54,7 @@ BatchedGemm<FpType>::BatchedGemm(int noperator, 
        sem_wait(gpu_access);

        get_current_time_with_ms (&sec , &ms);

        printf(" %"PRIdMAX".%03ld : acuired GPU\n", (intmax_t)sec, ms);

#endif



        if (use_sparse) {

          nC_ =  0;
@@ -132,11 +137,13 @@ BatchedGemm<FpType>::~BatchedGemm() { 
	               &Bbatch_

                       );

    };

#ifdef SHARE_GPU

    sem_post(gpu_access);

    time_t sec;

    long  ms;

    get_current_time_with_ms (&sec , &ms);

    printf(" %"PRIdMAX".%03ld : released GPU\n", (intmax_t)sec, ms);

#endif



}

src/apply_Htarget_vbatch.c

+4 −4
Original line number Diff line number Diff line
@@ -30,10 +30,10 @@ void apply_Htarget_vbatch( 
 const int ialign = 32;

 const double giga = 1000.0*1000.0*1000.0;



 double gflops1 = (FpType) 0.0;

 double gflops2 = (FpType) 0.0;

 double time_1st_vbatch = (FpType) 0.0;

 double time_2nd_vbatch = (FpType) 0.0;

 double gflops1 = (double) 0.0;

 double gflops2 = (double) 0.0;

 double time_1st_vbatch = (double) 0.0;

 double time_2nd_vbatch = (double) 0.0;



 size_t nbytes_BX = 0;

src/dmrg_magma.h

+21 −0
Original line number Diff line number Diff line
@@ -10,6 +10,26 @@ 


#include "magma_types.h"

#include "magma_v2.h"

#include "magma_operators.h"





#if defined(USE_COMPLEX_Z)



#define MAGMA_FpType  magmaDoubleComplex



#elif defined(USE_COMPLEX_C)



#define MAGMA_FpType  magmaFloatComplex



#elif defined(USE_FLOAT)



#define MAGMA_FpType  float



#else



#define MAGMA_FpType double



#endif





/*
@@ -562,6 +582,7 @@ void magma_Xsetvector( magma_int_t n, 
    #define magmablas_Xgemm_vbatched_max_nocheck magmablas_zgemm_vbatched_max_nocheck

    #define magmablas_Xgemm_vbatched_max magmablas_zgemm_vbatched_max



  

  #elif defined(USE_COMPLEX_C)



    #define magma_Xsetmatrix magma_csetmatrix

src/dmrg_malloc.c

+11 −3
Original line number Diff line number Diff line 
#include <stdlib.h>

#include <stdio.h>

#include "dmrg_vbatch.h"



#ifdef USE_MAGMA
@@ -6,16 +7,23 @@ 
#include "cuda_runtime.h"

#endif



#include "dmrg_lapack.h"



int dmrg_is_managed( const void *ptr )

{

  const int false = (0 == 1);

  int is_managed = false;

  const int lfalse = (0 == 1);

  int is_managed = lfalse;

 

#ifdef USE_MAGMA

  struct cudaPointerAttributes attribute;

  cudaError_t ierr = cudaPointerGetAttributes( &attribute, ptr );





 #if defined(CUDART_VERSION) && (CUDART_VERSION >= 10000)

  is_managed = (ierr == cudaSuccess) && (attribute.type == cudaMemoryTypeManaged);

 #else

  is_managed = (ierr == cudaSuccess) && attribute.isManaged;

 #endif

#endif

  return( is_managed );

}