scemama
/
qp_plugins_scemama
mirror of https://gitlab.com/scemama/qp_plugins_scemama.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Compare commits

base: scemama:0c5b6834b17c64e9ef091f669f95a88de0b9d06f
Branches Tags
scemama:master
...
compare: scemama:29903111f4169e0d019f56e0cdc242e32d931ce7
Branches Tags
scemama:master

2 Commits
0c5b6834b1 ... 29903111f4

Author SHA1 Message Date
Anthony Scemama 29903111f4 Allocate managed memory only if full 2023-12-06 10:42:52 +01:00
Anthony Scemama a08342f60c memory leak 2023-12-05 18:28:26 +01:00
3 changed files with 127 additions and 112 deletions
Show Stats Expand all files Collapse all files

130
devel/ccsd_gpu/gpu.c
View File

@ -138,15 +138,15 @@ void compute_h_oo_chol_gpu(gpu_data* data, int igpu)
double* d_cc_space_v_ov_chol = data[igpu].cc_space_v_ov_chol;
double* d_tau_kau;
cudaStat = cudaMallocManaged((void **)&d_tau_kau, cholesky_mo_num*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tau_kau, cholesky_mo_num*nV*nO * sizeof(double));
assert(cudaStat == cudaSuccess);
double* d_tmp_ovv;
cudaStat = cudaMallocManaged((void **)&d_tmp_ovv, nO*nV*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tmp_ovv, nO*nV*nV * sizeof(double));
assert(cudaStat == cudaSuccess);
double* d_tmp_vov;
cudaStat = cudaMallocManaged((void **)&d_tmp_vov, nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tmp_vov, nV*nO*nV * sizeof(double));
assert(cudaStat == cudaSuccess);
for (size_t i=0 ; i<nV ; ++i) {
@ -239,7 +239,7 @@ void compute_h_vo_chol_gpu(gpu_data* data, int igpu)
cublasDcopy(handle, nV*nO, d_cc_space_f_vo, 1, d_H_vo, 1);
double* d_tmp_k;
cudaStat = cudaMallocManaged((void **)&d_tmp_k, cholesky_mo_num * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tmp_k, cholesky_mo_num * sizeof(double));
assert(cudaStat == cudaSuccess);
alpha = 2.0;
@ -261,7 +261,7 @@ void compute_h_vo_chol_gpu(gpu_data* data, int igpu)
cudaFree(d_tmp_k);
double* d_tmp;
cudaStat = cudaMallocManaged((void **)&d_tmp, cholesky_mo_num*nO*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tmp, cholesky_mo_num*nO*nO * sizeof(double));
assert(cudaStat == cudaSuccess);
alpha = 1.0;
@ -273,7 +273,7 @@ void compute_h_vo_chol_gpu(gpu_data* data, int igpu)
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_T, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_tmp2;
cudaStat = cudaMallocManaged((void **)&d_tmp2, cholesky_mo_num*nO*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tmp2, cholesky_mo_num*nO*nO * sizeof(double));
assert(cudaStat == cudaSuccess);
for (size_t i=0 ; i<nV ; ++i) {
@ -290,6 +290,7 @@ void compute_h_vo_chol_gpu(gpu_data* data, int igpu)
cudaStreamDestroy(stream[i]);
}
cublasSetStream(handle, NULL);
cudaFree(d_tmp);
alpha = -1.0;
beta = 1.0;
@ -298,6 +299,7 @@ void compute_h_vo_chol_gpu(gpu_data* data, int igpu)
B=d_tmp2; ldb=cholesky_mo_num*nO;
C=d_H_vo; ldc=nV;
cublasDgemm(handle, CUBLAS_OP_T, CUBLAS_OP_N, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
cudaFree(d_tmp2);
double* H_vo = malloc(nV*nO*sizeof(double));
cublasGetMatrix(nV, nO, sizeof(double), d_H_vo, nV, H_vo, nV);
@ -344,11 +346,11 @@ void compute_h_vv_chol_gpu(gpu_data* data, int igpu)
double* d_cc_space_v_ov_chol = data[igpu].cc_space_v_ov_chol;
double* d_tau_kia;
cudaStat = cudaMallocManaged((void **)&d_tau_kia, cholesky_mo_num*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tau_kia, cholesky_mo_num*nO*nV * sizeof(double));
assert(cudaStat == cudaSuccess);
double* d_tmp_oov;
cudaStat = cudaMallocManaged((void **)&d_tmp_oov, nO*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tmp_oov, nO*nO*nV * sizeof(double));
assert(cudaStat == cudaSuccess);
alpha = 1.0;
@ -433,7 +435,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
double* d_r2;
lda = nO * nO;
cudaStat = cudaMallocManaged((void **)&d_r2, lda * nV * nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_r2, lda * nV * nV * sizeof(double));
assert (cudaStat == cudaSuccess);
cudaMemset(d_r2, 0, nO*nO*nV*nV*sizeof(double));
memset(r2, 0, nO*nO*nV*nV*sizeof(double));
@ -458,7 +460,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
double* d_H_vv = data[igpu].H_vv;
double* d_K1;
cudaStat = cudaMallocManaged((void **)&d_K1, nO*nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_K1, nO*nV*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
#pragma omp sections
@ -467,7 +469,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_J1;
cudaStat = cudaMallocManaged((void **)&d_J1, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_J1, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -479,7 +481,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
double* d_X_ovoo;
cudaStat = cudaMallocManaged((void **)&d_X_ovoo, nO*nV*nO*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_ovoo, nO*nV*nO*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 0.0;
beta = 1.0;
@ -502,7 +504,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
double* d_Y_ovov;
cudaStat = cudaMallocManaged((void **)&d_Y_ovov, nO*nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Y_ovov, nO*nV*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -534,7 +536,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasSetStream(handle, NULL);
double* d_tmp_cc;
cudaStat = cudaMallocManaged((void **)&d_tmp_cc, cholesky_mo_num*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tmp_cc, cholesky_mo_num*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -546,7 +548,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_T, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_J1_tmp;
cudaStat = cudaMallocManaged((void **)&d_J1_tmp, nV*nO*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_J1_tmp, nV*nO*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -578,7 +580,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cudaFree(d_J1_tmp);
double* d_X_voov;
cudaStat = cudaMallocManaged((void **)&d_X_voov, nV*nO*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_voov, nV*nO*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
for (size_t i=0 ; i<nV ; ++i) {
@ -612,7 +614,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasSetStream(handle, NULL);
double* d_Z_ovvo;
cudaStat = cudaMallocManaged((void **)&d_Z_ovvo, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Z_ovvo, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = -1.0;
@ -641,7 +643,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
}
double* d_Y_vovo;
cudaStat = cudaMallocManaged((void **)&d_Y_vovo, nV*nO*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Y_vovo, nV*nO*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -657,7 +659,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
}
double* d_X_ovvo;
cudaStat = cudaMallocManaged((void **)&d_X_ovvo, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_ovvo, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -732,11 +734,11 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_X;
cudaStat = cudaMallocManaged((void **)&d_X, nV*nO*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X, nV*nO*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_Y;
cudaStat = cudaMallocManaged((void **)&d_Y, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Y, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
for (size_t i=0 ; i<nV ; ++i) {
@ -771,7 +773,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasSetStream(handle, NULL);
double* d_Z;
cudaStat = cudaMallocManaged((void **)&d_Z, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Z, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -787,7 +789,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
double* d_t1v;
cudaStat = cudaMallocManaged((void **)&d_t1v, cholesky_mo_num*nO*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_t1v, cholesky_mo_num*nO*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -799,7 +801,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_T, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_K1tmp;
cudaStat = cudaMallocManaged((void **)&d_K1tmp, nO*nO*nV*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_K1tmp, nO*nO*nV*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -852,7 +854,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_Y_oooo;
cudaStat = cudaMallocManaged((void**)&d_Y_oooo, nO*nO*nO*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_Y_oooo, nO*nO*nO*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -864,7 +866,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_A1;
cudaStat = cudaMallocManaged((void**)&d_A1, nO*nO*nO*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_A1, nO*nO*nO*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -916,7 +918,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_g_vir;
cudaStat = cudaMallocManaged((void**)&d_g_vir, nV*nV*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_g_vir, nV*nV*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasDcopy(handle, nV*nV, d_H_vv, 1, d_g_vir, 1);
@ -929,7 +931,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_tmp_k;
cudaStat = cudaMallocManaged((void**)&d_tmp_k, cholesky_mo_num*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_tmp_k, cholesky_mo_num*sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
beta = 0.0;
@ -949,7 +951,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cudaFree(d_tmp_k);
double* d_tmp_vo;
cudaStat = cudaMallocManaged((void**)&d_tmp_vo, cholesky_mo_num*nV*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_tmp_vo, cholesky_mo_num*nV*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
beta = 0.0;
@ -960,7 +962,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_T, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_tmp_vo2;
cudaStat = cudaMallocManaged((void**)&d_tmp_vo2, cholesky_mo_num*nV*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_tmp_vo2, cholesky_mo_num*nV*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
for (size_t i=0 ; i<nO ; ++i) {
cudaStreamCreate(&(stream[i]));
@ -990,7 +992,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cudaFree(d_tmp_vo2);
double* d_Y_oovv;
cudaStat = cudaMallocManaged((void**)&d_Y_oovv, nO*nO*nV*nV*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_Y_oovv, nO*nO*nV*nV*sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
beta = 0.0;
@ -1035,12 +1037,12 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
{
double* d_g_occ;
lda = nO;
cudaStat = cudaMallocManaged((void **)&d_g_occ, nO*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_g_occ, nO*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasDcopy(handle, nO*nO, d_H_oo, 1, d_g_occ, 1);
double* d_X;
cudaStat = cudaMallocManaged((void **)&d_X, cholesky_mo_num*sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X, cholesky_mo_num*sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 2.0;
@ -1077,7 +1079,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_X_oovv;
cudaStat = cudaMallocManaged((void **)&d_X_oovv, nO*nO*nV*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_oovv, nO*nO*nV*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -1130,11 +1132,11 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_X_vovv;
cudaStat = cudaMallocManaged((void **)&d_X_vovv, nV*nO*nV*BLOCK_SIZE * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_vovv, nV*nO*nV*BLOCK_SIZE * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_Y_oovv;
cudaStat = cudaMallocManaged((void **)&d_Y_oovv, nO*nO*nV*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Y_oovv, nO*nO*nV*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
for (size_t iblock=0 ; iblock<nV ; iblock += BLOCK_SIZE) {
@ -1197,7 +1199,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_tcc2;
cudaStat = cudaMallocManaged((void **)&d_tcc2, cholesky_mo_num*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tcc2, cholesky_mo_num*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -1209,7 +1211,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_T, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_tcc;
cudaStat = cudaMallocManaged((void **)&d_tcc, cholesky_mo_num*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tcc, cholesky_mo_num*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -1221,7 +1223,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, m, n, k, &alpha, A, lda, B, ldb, &beta, C, ldc);
double* d_X_ovvo;
cudaStat = cudaMallocManaged((void **)&d_X_ovvo, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_ovvo, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -1265,7 +1267,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_X_oovv;
cudaStat = cudaMallocManaged((void **)&d_X_oovv, nO*nO*nV*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_oovv, nO*nO*nV*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -1296,7 +1298,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
}
double* d_X_vovo;
cudaStat = cudaMallocManaged((void **)&d_X_vovo, nV*nO*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_vovo, nV*nO*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 0.0;
@ -1316,7 +1318,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasSetStream(handle, NULL);
double* d_Y_oovo;
cudaStat = cudaMallocManaged((void **)&d_Y_oovo, nO*nO*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Y_oovo, nO*nO*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -1382,12 +1384,12 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
{
double* d_J1;
lda = nO*nV;
cudaStat = cudaMallocManaged((void **)&d_J1, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_J1, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(lda, nV*nO, sizeof(double), J1, lda, d_J1, lda);
double* d_X_ovvo;
cudaStat = cudaMallocManaged((void **)&d_X_ovvo, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_ovvo, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
for (size_t i=0 ; i<nV ; ++i) {
@ -1411,7 +1413,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cudaFree(d_J1);
double* d_Y_voov;
cudaStat = cudaMallocManaged((void **)&d_Y_voov, nV*nO*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Y_voov, nV*nO*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
for (size_t i=0 ; i<nV ; ++i) {
@ -1434,7 +1436,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasSetStream(handle, NULL);
double* d_Z_ovov;
cudaStat = cudaMallocManaged((void **)&d_Z_ovov, nO*nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Z_ovov, nO*nV*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -1479,11 +1481,11 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_X_ovov;
cudaStat = cudaMallocManaged((void **)&d_X_ovov, nO*nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_ovov, nO*nV*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_Y_ovov;
cudaStat = cudaMallocManaged((void **)&d_Y_ovov, nO*nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Y_ovov, nO*nV*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
for (size_t i=0 ; i<nV ; ++i) {
@ -1516,7 +1518,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasSetStream(handle, NULL);
double* d_Z_ovov;
cudaStat = cudaMallocManaged((void **)&d_Z_ovov, nO*nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Z_ovov, nO*nV*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -1562,11 +1564,11 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_X_ovov;
cudaStat = cudaMallocManaged((void **)&d_X_ovov, nO*nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_ovov, nO*nV*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_Y_ovov;
cudaStat = cudaMallocManaged((void **)&d_Y_ovov, nO*nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Y_ovov, nO*nV*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
@ -1600,7 +1602,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasSetStream(handle, NULL);
double* d_Z_ovov;
cudaStat = cudaMallocManaged((void **)&d_Z_ovov, nO*nV*nO*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_Z_ovov, nO*nV*nO*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 1.0;
@ -1648,7 +1650,7 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
double* d_tmp_cc;
lda = cholesky_mo_num * nV;
cudaStat = cudaMallocManaged((void **)&d_tmp_cc, lda * nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tmp_cc, lda * nV * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha=1.0; beta=0.0;
@ -1657,15 +1659,15 @@ void compute_r2_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, m, n, k, &alpha, A, m, B, k, &beta, C, m);
double* d_tmp_cc2;
cudaStat = cudaMallocManaged((void **)&d_tmp_cc2, cholesky_mo_num*nV*sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tmp_cc2, cholesky_mo_num*nV*sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_B1;
cudaStat = cudaMallocManaged((void**)&d_B1, nV*nV*BLOCK_SIZE*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_B1, nV*nV*BLOCK_SIZE*sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_tmpB1;
cudaStat = cudaMallocManaged((void**)&d_tmpB1, nV*BLOCK_SIZE*nV*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_tmpB1, nV*BLOCK_SIZE*nV*sizeof(double));
assert (cudaStat == cudaSuccess);
#pragma omp for
@ -1781,7 +1783,7 @@ void compute_r1_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
double* d_r1;
lda = nO ;
cudaStat = cudaMallocManaged((void **)&d_r1, lda * nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_r1, lda * nV * sizeof(double));
assert (cudaStat == cudaSuccess);
cudaMemset(d_r1, 0, nO*nV*sizeof(double));
memset(r1, 0, nO*nV*sizeof(double));
@ -1808,7 +1810,7 @@ void compute_r1_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDcopy(handle, nO*nV, d_cc_space_f_ov, 1, d_r1, 1);
double* d_X_oo;
cudaStat = cudaMallocManaged((void **)&d_X_oo, nO*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_oo, nO*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = -2.0;
@ -1855,7 +1857,7 @@ void compute_r1_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_X_voov;
cudaStat = cudaMallocManaged((void **)&d_X_voov, nV* nO* nO* nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_voov, nV* nO* nO* nV * sizeof(double));
assert (cudaStat == cudaSuccess);
for (size_t i=0 ; i<nV ; ++i) {
@ -1901,7 +1903,7 @@ void compute_r1_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_X_ovov;
cudaStat = cudaMallocManaged((void **)&d_X_ovov, nO* nV* nO* nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_X_ovov, nO* nV* nO* nV * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasDcopy(handle, nO*nV*nO*nV, d_cc_space_v_ovov, 1, d_X_ovov, 1);
@ -1939,7 +1941,7 @@ void compute_r1_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_T_vvoo;
cudaStat = cudaMallocManaged((void **)&d_T_vvoo, nV*nV*nO*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_T_vvoo, nV*nV*nO*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 0.0;
@ -1950,11 +1952,11 @@ void compute_r1_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
cublasDgeam(handle, CUBLAS_OP_N, CUBLAS_OP_T, nV*nV, nO*nO, &alpha, A, lda, &beta, B, ldb, C, ldc);
double* d_W_vvov;
cudaStat = cudaMallocManaged((void **)&d_W_vvov, nV*nV*nO*BLOCK_SIZE * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_W_vvov, nV*nV*nO*BLOCK_SIZE * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_W_vvov_tmp;
cudaStat = cudaMallocManaged((void **)&d_W_vvov_tmp, nV*nO*nV*BLOCK_SIZE * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_W_vvov_tmp, nV*nO*nV*BLOCK_SIZE * sizeof(double));
assert (cudaStat == cudaSuccess);
@ -2008,7 +2010,7 @@ void compute_r1_space_chol_gpu(gpu_data* data, int nO, int nV, double* t1, doubl
#pragma omp section
{
double* d_W_oovo;
cudaStat = cudaMallocManaged((void **)&d_W_oovo, nO*nO*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_W_oovo, nO*nO*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
alpha = 2.0;

13
devel/ccsd_gpu/gpu.h
View File

@ -60,3 +60,16 @@ typedef struct {
int cholesky_mo_num;
} gpu_data_sp;
static cudaError_t gpu_malloc(void** ptr, size_t size) {
size_t free, total;
cudaError_t rc = cudaMemGetInfo( &free, &total );
if (rc != cudaSuccess) return rc;
if (size < free && size < total/10) {
rc= cudaMalloc(ptr, size);
} else {
rc = cudaMallocManaged(ptr, size, cudaMemAttachGlobal);
}
return rc;
}

96
devel/ccsd_gpu/gpu_init.c
View File

@ -36,125 +36,125 @@ gpu_data* gpu_init(
double* d_cc_space_v_oo_chol;
lda = cholesky_mo_num * nO;
cudaStat = cudaMallocManaged((void **)&d_cc_space_v_oo_chol, lda * nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_cc_space_v_oo_chol, lda * nO * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(cholesky_mo_num*nO, nO, sizeof(double), cc_space_v_oo_chol, lda, d_cc_space_v_oo_chol, lda);
double* d_cc_space_v_ov_chol;
lda = cholesky_mo_num * nO;
cudaStat = cudaMallocManaged((void **)&d_cc_space_v_ov_chol, lda * nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_cc_space_v_ov_chol, lda * nV * sizeof(double) );
assert (cudaStat == cudaSuccess);
cublasSetMatrix(cholesky_mo_num*nO, nV, sizeof(double), cc_space_v_ov_chol, lda, d_cc_space_v_ov_chol, lda);
double* d_cc_space_v_vo_chol;
lda = cholesky_mo_num * nV;
cudaStat = cudaMallocManaged((void **)&d_cc_space_v_vo_chol, lda * nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_cc_space_v_vo_chol, lda * nO * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(cholesky_mo_num*nV, nO, sizeof(double), cc_space_v_vo_chol, lda, d_cc_space_v_vo_chol, lda);
double* d_cc_space_v_vv_chol;
lda = cholesky_mo_num * nV;
cudaStat = cudaMallocManaged((void **)&d_cc_space_v_vv_chol, lda * nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_cc_space_v_vv_chol, lda * nV * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(cholesky_mo_num*nV, nV, sizeof(double), cc_space_v_vv_chol, lda, d_cc_space_v_vv_chol, lda);
double* d_cc_space_v_oooo;
cudaStat = cudaMallocManaged((void**)&d_cc_space_v_oooo, nO*nO*nO*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_cc_space_v_oooo, nO*nO*nO*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(nO*nO, nO*nO, sizeof(double), cc_space_v_oooo, nO*nO, d_cc_space_v_oooo, nO*nO);
double* d_cc_space_v_vooo;
cudaStat = cudaMallocManaged((void**)&d_cc_space_v_vooo, nV*nO*nO*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_cc_space_v_vooo, nV*nO*nO*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(nV*nO, nO*nO, sizeof(double), cc_space_v_vooo, nV*nO, d_cc_space_v_vooo, nV*nO);
double* d_cc_space_v_voov;
cudaStat = cudaMallocManaged((void**)&d_cc_space_v_voov, nV*nO*nO*nV*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_cc_space_v_voov, nV*nO*nO*nV*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(nV*nO, nO*nV, sizeof(double), cc_space_v_voov, nV*nO, d_cc_space_v_voov, nV*nO);
double* d_cc_space_v_oovv;
cudaStat = cudaMallocManaged((void**)&d_cc_space_v_oovv, nO*nO*nV*nV*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_cc_space_v_oovv, nO*nO*nV*nV*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(nO*nO, nV*nV, sizeof(double), cc_space_v_oovv, nO*nO, d_cc_space_v_oovv, nO*nO);
double* d_cc_space_v_vvoo;
cudaStat = cudaMallocManaged((void**)&d_cc_space_v_vvoo, nV*nV*nO*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_cc_space_v_vvoo, nV*nV*nO*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(nV*nV, nO*nO, sizeof(double), cc_space_v_vvoo, nV*nV, d_cc_space_v_vvoo, nV*nV);
double* d_cc_space_v_oovo;
lda = nO*nO;
cudaStat = cudaMallocManaged((void **)&d_cc_space_v_oovo, nO*nO*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_cc_space_v_oovo, nO*nO*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(lda, nV*nO, sizeof(double), cc_space_v_oovo, lda, d_cc_space_v_oovo, lda);
double* d_cc_space_v_ovvo;
lda = nO*nV;
cudaStat = cudaMallocManaged((void **)&d_cc_space_v_ovvo, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_cc_space_v_ovvo, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(lda, nV*nO, sizeof(double), cc_space_v_ovvo, lda, d_cc_space_v_ovvo, lda);
double* d_cc_space_v_ovov;
lda = nO*nV;
cudaStat = cudaMallocManaged((void **)&d_cc_space_v_ovov, nO*nV*nV*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_cc_space_v_ovov, nO*nV*nV*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(lda, nV*nO, sizeof(double), cc_space_v_ovov, lda, d_cc_space_v_ovov, lda);
double* d_cc_space_v_ovoo;
lda = nO*nV;
cudaStat = cudaMallocManaged((void **)&d_cc_space_v_ovoo, nO*nV*nO*nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_cc_space_v_ovoo, nO*nV*nO*nO * sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(lda, nO*nO, sizeof(double), cc_space_v_ovoo, lda, d_cc_space_v_ovoo, lda);
double* d_cc_space_f_oo;
cudaStat = cudaMallocManaged((void**)&d_cc_space_f_oo, nO*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_cc_space_f_oo, nO*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(nO, nO, sizeof(double), cc_space_f_oo, nO, d_cc_space_f_oo, nO);
double* d_cc_space_f_vo;
cudaStat = cudaMallocManaged((void**)&d_cc_space_f_vo, nV*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_cc_space_f_vo, nV*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(nV, nO, sizeof(double), cc_space_f_vo, nV, d_cc_space_f_vo, nV);
double* d_cc_space_f_ov;
cudaStat = cudaMallocManaged((void**)&d_cc_space_f_ov, nV*nO*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_cc_space_f_ov, nV*nO*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(nO, nV, sizeof(double), cc_space_f_ov, nO, d_cc_space_f_ov, nO);
double* d_cc_space_f_vv;
cudaStat = cudaMallocManaged((void**)&d_cc_space_f_vv, nV*nV*sizeof(double), 1);
cudaStat = gpu_malloc((void**)&d_cc_space_f_vv, nV*nV*sizeof(double));
assert (cudaStat == cudaSuccess);
cublasSetMatrix(nV, nV, sizeof(double), cc_space_f_vv, nV, d_cc_space_f_vv, nV);
double* d_tau;
lda = nO * nO;
cudaStat = cudaMallocManaged((void **)&d_tau, lda * nV * nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tau, lda * nV * nV * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_tau_x;
lda = nO * nO;
cudaStat = cudaMallocManaged((void **)&d_tau_x, lda * nV * nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_tau_x, lda * nV * nV * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_t1;
cudaStat = cudaMallocManaged((void **)&d_t1, nO * nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_t1, nO * nV * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_t2;
cudaStat = cudaMallocManaged((void **)&d_t2, nO*nO*nV*nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_t2, nO*nO*nV*nV * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_H_oo;
cudaStat = cudaMallocManaged((void **)&d_H_oo, nO * nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_H_oo, nO * nO * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_H_vo;
cudaStat = cudaMallocManaged((void **)&d_H_vo, nV * nO * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_H_vo, nV * nO * sizeof(double));
assert (cudaStat == cudaSuccess);
double* d_H_vv;
cudaStat = cudaMallocManaged((void **)&d_H_vv, nV * nV * sizeof(double), 1);
cudaStat = gpu_malloc((void **)&d_H_vv, nV * nV * sizeof(double));
assert (cudaStat == cudaSuccess);
data[igpu].cc_space_v_oo_chol = d_cc_space_v_oo_chol;
@ -202,30 +202,30 @@ void gpu_deinit(gpu_data* data)
int igpu = omp_get_thread_num();
cudaSetDevice(igpu);
free(data[igpu].cc_space_v_oo_chol);
free(data[igpu].cc_space_v_ov_chol);
free(data[igpu].cc_space_v_vo_chol);
free(data[igpu].cc_space_v_vv_chol);
free(data[igpu].cc_space_v_oooo);
free(data[igpu].cc_space_v_vooo);
free(data[igpu].cc_space_v_voov);
free(data[igpu].cc_space_v_oovv);
free(data[igpu].cc_space_v_vvoo);
free(data[igpu].cc_space_v_oovo);
free(data[igpu].cc_space_v_ovvo);
free(data[igpu].cc_space_v_ovov);
free(data[igpu].cc_space_v_ovoo);
free(data[igpu].cc_space_f_oo);
free(data[igpu].cc_space_f_ov);
free(data[igpu].cc_space_f_vo);
free(data[igpu].cc_space_f_vv);
free(data[igpu].tau);
free(data[igpu].tau_x);
free(data[igpu].t1);
free(data[igpu].t2);
free(data[igpu].H_oo);
free(data[igpu].H_vo);
free(data[igpu].H_vv);
cudaFree(data[igpu].cc_space_v_oo_chol);
cudaFree(data[igpu].cc_space_v_ov_chol);
cudaFree(data[igpu].cc_space_v_vo_chol);
cudaFree(data[igpu].cc_space_v_vv_chol);
cudaFree(data[igpu].cc_space_v_oooo);
cudaFree(data[igpu].cc_space_v_vooo);
cudaFree(data[igpu].cc_space_v_voov);
cudaFree(data[igpu].cc_space_v_oovv);
cudaFree(data[igpu].cc_space_v_vvoo);
cudaFree(data[igpu].cc_space_v_oovo);
cudaFree(data[igpu].cc_space_v_ovvo);
cudaFree(data[igpu].cc_space_v_ovov);
cudaFree(data[igpu].cc_space_v_ovoo);
cudaFree(data[igpu].cc_space_f_oo);
cudaFree(data[igpu].cc_space_f_ov);
cudaFree(data[igpu].cc_space_f_vo);
cudaFree(data[igpu].cc_space_f_vv);
cudaFree(data[igpu].tau);
cudaFree(data[igpu].tau_x);
cudaFree(data[igpu].t1);
cudaFree(data[igpu].t2);
cudaFree(data[igpu].H_oo);
cudaFree(data[igpu].H_vo);
cudaFree(data[igpu].H_vv);
}
}