Skip to content

Commit

Permalink
cleanup simple timer
Browse files Browse the repository at this point in the history
  • Loading branch information
@cjknight
cjknight committed May 31, 2024
1 parent 285ae2e commit bf3dff4
Show file tree
Hide file tree
Showing 5 changed files with 129 additions and 294 deletions.
140 changes: 28 additions & 112 deletions gpu/src/device.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@

#include "device.h"

#define _NUM_TIMER_JK 13
#define _NUM_SIMPLE_TIMER 4

#define _DEBUG_OPENMP

Expand Down Expand Up @@ -102,13 +102,8 @@ Device::Device()
#endif

#ifdef _SIMPLE_TIMER
t_array_count = 0;
t_array = (double *) malloc(14 * sizeof(double));
for(int i=0; i<14; ++i) t_array[i] = 0.0;

t_array_jk_count = 0;
t_array_jk = (double* ) malloc(_NUM_TIMER_JK * sizeof(double));
for(int i=0; i<_NUM_TIMER_JK; ++i) t_array_jk[i] = 0.0;
t_array = (double* ) malloc(_NUM_SIMPLE_TIMER * sizeof(double));
for(int i=0; i<_NUM_SIMPLE_TIMER; ++i) t_array[i] = 0.0;
#endif
}

Expand All @@ -117,10 +112,6 @@ Device::Device()
Device::~Device()
{
printf("LIBGPU: destroying device\n");

#ifdef _SIMPLE_TIMER
double t0 = omp_get_wtime();
#endif

pm->dev_free_host(rho);
//pm->dev_free_host(vj);
Expand All @@ -144,38 +135,34 @@ Device::~Device()
// d_tril_map.clear();

// pm->dev_free_host(tril_map);

#ifdef _SIMPLE_TIMER
t_array_jk[11] += omp_get_wtime() - t0;
#endif

#ifdef _SIMPLE_TIMER
printf("LIBGPU::orbital_response\n");
double total = 0.0;
for(int i=0; i<14; ++i) {
total += t_array[i];
printf("i= %i t_array= %f s\n",i,t_array[i]);
}
printf(" total= %f s count= %i\n",total,t_array_count);

printf("LIBGPU::get_jk\n");
total = 0.0;
for(int i=0; i<_NUM_TIMER_JK; ++i) {
total += t_array_jk[i];
printf("i= %i t_array= %f s\n",i,t_array_jk[i]);
}
printf(" total= %f s count= %i\n",total,t_array_jk_count);
for(int i=0; i<_NUM_SIMPLE_TIMER; ++i) total += t_array[i];

printf("\nLIBGPU :: SIMPLE_TIMER\n");
printf("\nLIBGPU :: SIMPLE_TIMER :: get_jk\n");
printf("LIBGPU :: SIMPLE_TIMER :: i= %i name= init_get_jk() time= %f s\n",0,t_array[0]);
printf("LIBGPU :: SIMPLE_TIMER :: i= %i name= pull_get_jk() time= %f s\n",1,t_array[1]);
printf("LIBGPU :: SIMPLE_TIMER :: i= %i name= get_jk() time= %f s\n",2,t_array[2]);

printf("\nLIBGPU :: SIMPLE_TIMER :: hessop\n");
printf("LIBGPU :: SIMPLE_TIMER :: i= %i name= hessop_get_veff() time= %f s\n",3,t_array[3]);

printf("\nLIBGPU :: SIMPLE_TIMER :: orbital_response\n");
printf("LIBGPU :: SIMPLE_TIMER :: i= %i name= orbital_response() time= %f s\n",4,t_array[4]);

printf("LIBGPU :: SIMPLE_TIMER :: total= %f s\n",total);

free(t_array);
free(t_array_jk);
#endif

// print summary of cached eri blocks

if(use_eri_cache) {
printf("LIBGPU::eri cache :: size= %i\n",eri_list.size());
printf("\nLIBGPU :: eri cache statistics :: count= %i\n",eri_list.size());
for(int i=0; i<eri_list.size(); ++i)
printf("%i : eri= %p Mbytes= %f count= %i update= %i device= %i\n", i, eri_list[i],
printf("LIBGPU :: %i : eri= %p Mbytes= %f count= %i update= %i device= %i\n", i, eri_list[i],
eri_size[i]*sizeof(double)/1024./1024., eri_count[i], eri_update[i], eri_device[i]);

eri_count.clear();
Expand Down Expand Up @@ -391,7 +378,12 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
py::array_t<double> _ppaa, py::array_t<double> _papa, py::array_t<double> _eri_paaa,
py::array_t<double> _ocm2, py::array_t<double> _tcm2, py::array_t<double> _gorb,
int ncore, int nocc, int nmo)
{
{

#ifdef _SIMPLE_TIMER
double t0 = omp_get_wtime();
#endif

py::buffer_info info_ppaa = _ppaa.request(); // 4D array (26, 26, 2, 2)
py::buffer_info info_papa = _papa.request(); // 4D array (26, 2, 26, 2)
py::buffer_info info_paaa = _eri_paaa.request();
Expand Down Expand Up @@ -457,10 +449,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
// printf("f1 += paaa{%i, %i, %i} X cm{%i, %i, %i, %i}\n",
// nocc-ncore,info_ppaa.shape[2],info_ppaa.shape[3],
// info_ocm2.shape[0],info_ocm2.shape[1],info_ocm2.shape[2],ncore);

#ifdef _SIMPLE_TIMER
double t0 = omp_get_wtime();
#endif

for(int i=0; i<ncore; ++i) {

Expand All @@ -477,10 +465,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
f1[i] += val;
}

#ifdef _SIMPLE_TIMER
double t1 = omp_get_wtime();
#endif

// tensordot(ra, cm, axes=((0,1,2), (3,0,1)))

// printf("f1 += ra{%i, %i, %i} X cm{%i, %i, %i, %i}\n",
Expand All @@ -502,10 +486,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
f1[ncore+i] += val;
}

#ifdef _SIMPLE_TIMER
double t2 = omp_get_wtime();
#endif

// tensordot(ar, cm, axes=((0,1,2), (0,3,2)))

// printf("f1 += ar{%i, %i, %i} X cm{%i, %i, %i, %i}\n",
Expand All @@ -527,10 +507,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
f1[ncore+i] += val;
}

#ifdef _SIMPLE_TIMER
double t3 = omp_get_wtime();
#endif

// tensordot(ar, cm, axes=((0,1,2), (1,3,2)))

// printf("f1 += ar{%i, %i, %i} X cm{%i, %i, %i, %i}\n",
Expand All @@ -552,10 +528,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
f1[ncore+i] += val;
}

#ifdef _SIMPLE_TIMER
double t4 = omp_get_wtime();
#endif

// ====================================================================
// iteration (nocc, nmo)
// ====================================================================
Expand Down Expand Up @@ -587,10 +559,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
f1[i] += val;
}

#ifdef _SIMPLE_TIMER
double t5 = omp_get_wtime();
#endif

// tensordot(ra, cm, axes=((0,1,2), (3,0,1)))

// printf("f1 += ra{%i, %i, %i} X cm{%i, %i, %i, %i}\n",
Expand All @@ -612,10 +580,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
f1[ncore+i] += val;
}

#ifdef _SIMPLE_TIMER
double t6 = omp_get_wtime();
#endif

// tensordot(ar, cm, axes=((0,1,2), (0,3,2)))

// printf("f1 += ar{%i, %i, %i} X cm{%i, %i, %i, %i}\n",
Expand All @@ -637,10 +601,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
f1[ncore+i] += val;
}

#ifdef _SIMPLE_TIMER
double t7 = omp_get_wtime();
#endif

// tensordot(ar, cm, axes=((0,1,2), (1,3,2)))

// printf("f1 += ar{%i, %i, %i} X cm{%i, %i, %i, %i}\n",
Expand All @@ -661,18 +621,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,

f1[ncore+i] += val;
}
#ifdef _SIMPLE_TIMER
double t8 = omp_get_wtime();

t_array[0] += t1 - t0;
t_array[1] += t2 - t1;
t_array[2] += t3 - t2;
t_array[3] += t4 - t3;
t_array[4] += t5 - t4;
t_array[5] += t6 - t5;
t_array[6] += t7 - t6;
t_array[7] += t8 - t7;
#endif
} // for(p<nmo)

// # (H.x_aa)_va, (H.x_aa)_ac
Expand All @@ -687,10 +635,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,

// ocm2 = ocm2[:,:,:,ncore:nocc] + ocm2[:,:,:,ncore:nocc].transpose (1,0,3,2)

#ifdef _SIMPLE_TIMER
double t8 = omp_get_wtime();
#endif

int indx = 0;
double * _ocm2_tmp = ecm2;
for(int i=0; i<info_ocm2.shape[0]; ++i)
Expand All @@ -703,10 +647,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
_ocm2_tmp[indx++] = ocm2[indx1] + ocm2[indx2];
}

#ifdef _SIMPLE_TIMER
double t9 = omp_get_wtime();
#endif

// ocm2 += ocm2.transpose (2,3,0,1)

_ocm2_size_3d = info_ocm2.shape[1] * _ocm2_size_2d;
Expand All @@ -722,18 +662,10 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
_ocm2t[indx] = _ocm2_tmp[indx1] + _ocm2_tmp[indx2];
indx++;
}

#ifdef _SIMPLE_TIMER
double t10 = omp_get_wtime();
#endif

// ecm2 = ocm2 + tcm2

for(int i=0; i<size_ecm; ++i) ecm2[i] = _ocm2t[i] + tcm2[i];

#ifdef _SIMPLE_TIMER
double t11 = omp_get_wtime();
#endif

// f1_prime[:ncore,ncore:nocc] += np.tensordot (self.eri_paaa[:ncore], ecm2, axes=((1,2,3),(1,2,3)))

Expand All @@ -756,10 +688,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,

f1_prime[i*nmo+ncore+j] += val;
}

#ifdef _SIMPLE_TIMER
double t12 = omp_get_wtime();
#endif

// f1_prime[nocc:,ncore:nocc] += np.tensordot (self.eri_paaa[nocc:], ecm2, axes=((1,2,3),(1,2,3)))

Expand All @@ -779,10 +707,6 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
f1_prime[i*nmo+ncore+j] += val;
}

#ifdef _SIMPLE_TIMER
double t13 = omp_get_wtime();
#endif

// return gorb + (f1_prime - f1_prime.T)

double * g_f1_prime = (double *) pm->dev_malloc_host(nmo*nmo*sizeof(double));
Expand All @@ -802,16 +726,8 @@ void Device::orbital_response(py::array_t<double> _f1_prime,
for(int i=0; i<nmo*nmo; ++i) res[i] = g_f1_prime[i];

#ifdef _SIMPLE_TIMER
double t14 = omp_get_wtime();

t_array[8] += t9 - t8;
t_array[9] += t10 - t9;
t_array[10] += t11 - t10;
t_array[11] += t12 - t11;
t_array[12] += t13 - t12;
t_array[13] += t14 - t13;

t_array_count++;
double t1 = omp_get_wtime();
t_array[4] += t1 - t0;
#endif

#if 0
Expand Down
4 changes: 0 additions & 4 deletions gpu/src/device.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -213,11 +213,7 @@ public :
void NPdunpack_tril(int, double *, double *, int);

#ifdef _SIMPLE_TIMER
int t_array_count;
double * t_array;

int t_array_jk_count;
double * t_array_jk;
#endif

int num_threads;
Expand Down
Loading

0 comments on commit bf3dff4

Please sign in to comment.