diogo/yap-6.3
Archived
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

eliminate leak and fix memory tracking

Browse Source
This commit is contained in:
Vítor Santos Costa 2013-10-16 17:33:49 +01:00
parent 5ad10a1057
commit bd541bc57f
5 changed files with 46 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
packages/cuda/bpreds.cu
View File

@ -52,7 +52,7 @@ int bpreds(int *dop1, int rows, int cols, int *bin, int3 numpreds, int **ret)
int tmplen = rows + 1;
int size = tmplen * sizeof(int);
reservar(&temp, size);
// cerr << "+ " << temp << " temp bpreds " << size << endl;
// DEBUG_MEM cerr << "+ " << temp << " temp bpreds " << size << endl;
cudaMemset(temp, 0, size);
#if TIMER
@ -68,7 +68,7 @@ int bpreds(int *dop1, int rows, int cols, int *bin, int3 numpreds, int **ret)
else
hsize = sproj;
reservar(&dhead, hsize);
// cerr << "+ " << dhead << " dhead " << hsize << endl;
// DEBUG_MEM cerr << "+ " << dhead << " dhead " << hsize << endl;
cudaMemcpy(dhead, bin, spredn, cudaMemcpyHostToDevice);
int blockllen = rows / 1024 + 1;
@ -113,7 +113,7 @@ int bpreds(int *dop1, int rows, int cols, int *bin, int3 numpreds, int **ret)
int *fres;
reservar(&fres, num * sproj);
// cerr << "+ " << fres << " fres " << num * sproj << endl;
// DEBUG_MEM cerr << "+ " << fres << " fres " << num * sproj << endl;
cudaMemcpy(dhead, bin + predn, sproj, cudaMemcpyHostToDevice);
llenarproyectar<<<blockllen, numthreads, sproj>>>(dop1, rows, numpreds.y, temp, dhead, numpreds.z, fres);

14
packages/cuda/lista.cu
View File

@ -859,6 +859,7 @@ extern "C"
void Cuda_Statistics(void)
{
cerr << "GPU Statistics" << endl;
#if TIMER
cerr << "Called " << cuda_stats.calls << "times." << endl;
cerr << "GPU time " << cuda_stats.total_time << "msec." << endl;
cerr << "Longest call " << cuda_stats.max_time << "msec." << endl;
@ -875,6 +876,7 @@ extern "C"
cerr << " Selects/Projects: " << cuda_stats.selects << "." << endl;
cerr << " Unions: " << cuda_stats.unions << "." << endl;
cerr << " Built-ins: " << cuda_stats.builtins << "." << endl << endl;
#endif
}
extern "C"
@ -946,11 +948,13 @@ int Cuda_Eval(predicate **inpfacts, int ninpf, predicate **inprules, int ninpr,
vector<gpunode>::iterator qposf;
vector<rulenode>::iterator qposr;
#if TIMER
cudaEvent_t start, stop;
float time;
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start, 0);
#endif
while(reglas.size()) /*Here's the main loop*/
{
@ -996,7 +1000,7 @@ int Cuda_Eval(predicate **inpfacts, int ninpf, predicate **inprules, int ninpr,
{
num_refs = rows1 * cols1 * sizeof(int);
reservar(&res, num_refs);
// cerr << "+ " << res << " Res " << num_refs << endl;
// DEBUG_MEM cerr << "+ " << res << " Res " << num_refs << endl;
cudaMemcpyAsync(res, dop1, num_refs, cudaMemcpyDeviceToDevice);
registrar(rul_act->name, cols1, res, rows1, itr, 1);
rul_act->gen_ant = rul_act->gen_act;
@ -1285,7 +1289,7 @@ int Cuda_Eval(predicate **inpfacts, int ninpf, predicate **inprules, int ninpr,
res_rows = selectproyect(dop1, rows1, cols1, tmprule.num_columns, tmprule.select[0], tmprule.numsel[0], tmprule.selfjoin[0], tmprule.numselfj[0], tmprule.project[0], &res);
if(qposr != fin && qposr->name == qname) {
cudaFree(dop1);
// cerr << "- " << dop1 << " dop1" << endl;
// DEBUG_MEM cerr << "- " << dop1 << " dop1" << endl;
}
}
@ -1293,14 +1297,15 @@ int Cuda_Eval(predicate **inpfacts, int ninpf, predicate **inprules, int ninpr,
tipo = res_rows * cols1 * sizeof(int);
hres = (int *)malloc(tipo);
cudaMemcpy(hres, res, tipo, cudaMemcpyDeviceToHost);
if(res_rows > 0 && tmprule.numsel[0] != 0 && tmprule.numselfj[0] != 0) {
if(res_rows > 0 /*&& tmprule.numsel[0] != 0 && tmprule.numselfj[0] != 0 */) {
cudaFree(res);
// cerr << "- " << res << " res" << endl;
// DEBUG_MEM cerr << "- " << res << " res" << endl;
}
}
else
res_rows = 0;
#if TIMER
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
cudaEventElapsedTime(&time, start, stop);
@ -1311,6 +1316,7 @@ int Cuda_Eval(predicate **inpfacts, int ninpf, predicate **inprules, int ninpr,
cuda_stats.min_time = time;
cudaEventDestroy(start);
cudaEventDestroy(stop);
#endif
if(showr == 1)
{

20
packages/cuda/memory.cu
View File

@ -112,7 +112,7 @@ void limpiar(const char s[])
if(GPUmem.size() == 0)
{
// cerr << s << ": not enough GPU memory: have " << avmem << endl;
// DEBUG_MEM cerr << s << ": not enough GPU memory: have " << avmem << endl;
exit(1);
}
@ -176,7 +176,7 @@ void liberar(int *ptr, int size)
//cout << "L " << avmem << " " << size;
cudaFree(ptr);
// cerr << "- " << ptr << " " << size << endl;
// DEBUG_MEM cerr << "- " << ptr << " " << size << endl;
avmem += size;
//cout << " " << avmem << endl;
@ -186,7 +186,7 @@ void reservar(int **ptr, int size)
{
//size_t free, total;
//cudaMemGetInfo( &free, &total );
// // cerr << "- " << free << " " << size << endl;
// cerr << "? " << free << " " << size << endl;
if (size == 0) {
*ptr = NULL;
@ -199,8 +199,8 @@ void reservar(int **ptr, int size)
if (! *ptr ) {
size_t free, total;
cudaMemGetInfo( &free, &total );
// cerr << "Could not allocate " << size << " bytes, only " << free << " avaliable from total of " << total << " !!!" << endl;
// cerr << "Exiting CUDA...." << endl;
cerr << "Could not allocate " << size << " bytes, only " << free << " avaliable from total of " << total << " !!!" << endl;
cerr << "Exiting CUDA...." << endl;
exit(1);
}
avmem -= size;
@ -277,7 +277,7 @@ int cargar(int name, int num_rows, int num_columns, int is_fact, int *address_ho
}
size = num_rows * num_columns * sizeof(int);
reservar(&temp, size);
// cerr << "+ " << temp << " temp " << size << endl;
// DEBUG_MEM cerr << "+ " << temp << " temp " << size << endl;
cudaMemcpyAsync(temp, address_host_table, size, cudaMemcpyHostToDevice);
registrar(name, num_columns, temp, num_rows, itr, 0);
*ptr = temp;
@ -297,7 +297,7 @@ int cargar(int name, int num_rows, int num_columns, int is_fact, int *address_ho
}
size = totalrows * num_columns * sizeof(int);
reservar(&temp, size);
// cerr << "+ " << temp << " temp 2 " << size << endl;
// DEBUG_MEM cerr << "+ " << temp << " temp 2 " << size << endl;
for(x = 1; x < numgpu; x++)
{
cudaMemcpyAsync(temp + temp_storage[x-1].size, temp_storage[x].dev_address, temp_storage[x].size, cudaMemcpyDeviceToDevice);
@ -340,7 +340,7 @@ int cargafinal(int name, int cols, int **ptr)
}
reservar(&temp, cont * cols * sizeof(int));
// cerr << "+ " << temp << " temp 3 " << cont * cols * sizeof(int) << endl;
// DEBUG_MEM cerr << "+ " << temp << " temp 3 " << cont * cols * sizeof(int) << endl;
ini = temp;
pos = gpu;
@ -463,7 +463,7 @@ void resultados(vector<rulenode>::iterator first, vector<rulenode>::iterator las
cout << endl;
}
cudaFree(gpu->dev_address);
// cerr << "- " << gpu->dev_address << " gpu->dev_address" << endl;
// DEBUG_MEM cerr << "- " << gpu->dev_address << " gpu->dev_address" << endl;
free(temp);
gpu++;
}
@ -495,7 +495,7 @@ void clear_memory()
{
if (ini->isrule) {
cudaFree(ini->dev_address);
// cerr << "- " << ini->dev_address << " ini->dev_address" << endl;
// DEBUG_MEM cerr << "- " << ini->dev_address << " ini->dev_address" << endl;
ini = GPUmem.erase(ini);
} else {
ini++;

12
packages/cuda/selectproyect.cu
View File

@ -211,7 +211,7 @@ int selectproyect(int *dop1, int rows, int cols, int head_size, int *select, int
#endif
int head_bytes = mayor(numselect, numselfj, head_size) * sizeof(int);
reservar(&dhead, head_bytes);
// cerr << "+ " << dhead << " dhead " << head_bytes << endl;
// DEBUG_MEM cerr << "+ " << dhead << " dhead " << head_bytes << endl;
int blockllen = rows / 1024 + 1;
int numthreads = 1024;
@ -222,7 +222,7 @@ int selectproyect(int *dop1, int rows, int cols, int head_size, int *select, int
tmplen = rows + 1;
size2 = tmplen * sizeof(int);
reservar(&temp, size2);
// cerr << "+ " << temp << " temp select " << size2 << endl;
// DEBUG_MEM cerr << "+ " << temp << " temp select " << size2 << endl;
cudaMemset(temp, 0, size2);
size = numselect * sizeof(int);
@ -245,7 +245,7 @@ int selectproyect(int *dop1, int rows, int cols, int head_size, int *select, int
size = head_size * sizeof(int);
reservar(&fres, num * size);
// cerr << "+ " << fres << " fres select " << num*size << endl;
// DEBUG_MEM cerr << "+ " << fres << " fres select " << num*size << endl;
cudaMemcpy(dhead, project, size, cudaMemcpyHostToDevice);
llenarproyectar<<<blockllen, numthreads, size>>>(dop1, rows, cols, temp, dhead, head_size, fres);
liberar(dhead, head_bytes);
@ -260,7 +260,7 @@ int selectproyect(int *dop1, int rows, int cols, int head_size, int *select, int
tmplen = rows + 1;
size2 = tmplen * sizeof(int);
reservar(&temp, size2);
// cerr << "+ " << temp << " temp select " << size2 << endl;
// DEBUG_MEM cerr << "+ " << temp << " temp select " << size2 << endl;
cudaMemset(temp, 0, size2);
size = numselfj * sizeof(int);
@ -275,7 +275,7 @@ int selectproyect(int *dop1, int rows, int cols, int head_size, int *select, int
size = head_size * sizeof(int);
reservar(&fres, num * size);
// cerr << "+ " << fres << " fres select again " << num*size << endl;
// DEBUG_MEM cerr << "+ " << fres << " fres select again " << num*size << endl;
cudaMemcpy(dhead, project, size, cudaMemcpyHostToDevice);
llenarproyectar<<<blockllen, numthreads, size>>>(dop1, rows, cols, temp, dhead, head_size, fres);
liberar(dhead, head_bytes);
@ -287,7 +287,7 @@ int selectproyect(int *dop1, int rows, int cols, int head_size, int *select, int
{
size = head_size * sizeof(int);
reservar(&fres, rows * size);
// cerr << "+ " << fres << " fres select third " << rows*size << endl;
// DEBUG_MEM cerr << "+ " << fres << " fres select third " << rows*size << endl;
cudaMemcpy(dhead, project, size, cudaMemcpyHostToDevice);
proyectar<<<blockllen, numthreads, size>>>(dop1, rows, cols, dhead, head_size, fres);
liberar(dhead, head_bytes);

34
packages/cuda/treeb.cu
View File

@ -780,9 +780,9 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
int dconsize = sizet * 2;*/
reservar(&dcons, sizet);
// cerr << "+ " << dcons << " dcons tree " << sizet << endl;
// DEBUG_MEM cerr << "+ " << dcons << " dcons tree " << sizet << endl;
reservar(&temp, sizet2);
// cerr << "+ " << temp << " temp tree " << sizet2 << endl;
// DEBUG_MEM cerr << "+ " << temp << " temp tree " << sizet2 << endl;
thrust::device_ptr<int> res = thrust::device_pointer_cast(temp);
numthreads = 1024;
@ -857,9 +857,9 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
memSizeS = newLen * sizeof(int);
reservar(&d_S, memSizeS);
// cerr << "+ " << d_S << " d_S " << memSizeS << endl;
// DEBUG_MEM cerr << "+ " << d_S << " d_S " << memSizeS << endl;
reservar(&posS, memSizeS);
// cerr << "+ " << posS << " posS " << memSizeS << endl;
// DEBUG_MEM cerr << "+ " << posS << " posS " << memSizeS << endl;
llenar<<<blockllen, numthreads>>>(p2, d_S, sLen, of2, wherej[1], temp, posS);
sLen = newLen;
}
@ -880,9 +880,9 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
memSizeS = newLen * sizeof(int);
reservar(&d_S, memSizeS);
// cerr << "+ " << d_S << " d_S m " << memSizeS << endl;
// DEBUG_MEM cerr << "+ " << d_S << " d_S m " << memSizeS << endl;
reservar(&posS, memSizeS);
// cerr << "+ " << posS << " posS m " << memSizeS << endl;
// DEBUG_MEM cerr << "+ " << posS << " posS m " << memSizeS << endl;
llenar<<<blockllen, numthreads>>>(p2, d_S, sLen, of2, wherej[1], temp, posS);
sLen = newLen;
}
@ -890,7 +890,7 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
{
memSizeS = sLen * sizeof(int);
reservar(&d_S, memSizeS);
// cerr << "+ " << d_S << " d_S n " << memSizeS << endl;
// DEBUG_MEM cerr << "+ " << d_S << " d_S n " << memSizeS << endl;
llenarnosel<<<blockllen, numthreads>>>(p2, d_S, sLen, of2, wherej[1]);
}
}
@ -938,9 +938,9 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
m32rLen = newLen + extraspace;
sizem32 = m32rLen * sizeof(int);
reservar(&d_R, sizem32);
// cerr << "+ " << d_R << " d_R m " << sizem32 << endl;
// DEBUG_MEM cerr << "+ " << d_R << " d_R m " << sizem32 << endl;
reservar(&posR, sizem32);
// cerr << "+ " << posR << " posR m " << sizem32 << endl;
// DEBUG_MEM cerr << "+ " << posR << " posR m " << sizem32 << endl;
cudaMemsetAsync(d_R + newLen, 0x7f, sizextra);
cudaMemsetAsync(posR + newLen, 0x7f, sizextra);
llenar<<<blockllen, numthreads>>>(p1, d_R, rLen, of1, wherej[0], temp, posR);
@ -966,9 +966,9 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
m32rLen = newLen + extraspace;
sizem32 = m32rLen * sizeof(int);
reservar(&d_R, sizem32);
// cerr << "+ " << d_R << " d_R n " << sizem32 << endl;
// DEBUG_MEM cerr << "+ " << d_R << " d_R n " << sizem32 << endl;
reservar(&posR, sizem32);
// cerr << "+ " << posR << " posR n " << sizem32 << endl;
// DEBUG_MEM cerr << "+ " << posR << " posR n " << sizem32 << endl;
cudaMemsetAsync(d_R + newLen, 0x7f, sizextra);
cudaMemsetAsync(posR + newLen, 0x7f, sizextra);
llenar<<<blockllen, numthreads>>>(p1, d_R, rLen, of1, wherej[0], temp, posR);
@ -978,7 +978,7 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
{
sizem32 = m32rLen * sizeof(int);
reservar(&d_R, sizem32);
// cerr << "+ " << d_R << " d_R sizem32 " << sizem32 << endl;
// DEBUG_MEM cerr << "+ " << d_R << " d_R sizem32 " << sizem32 << endl;
cudaMemsetAsync(d_R + rLen, 0x7f, extraspace * sizeof(int));
llenarnosel<<<blockllen, numthreads>>>(p1, d_R, rLen, of1, wherej[0]);
}
@ -989,7 +989,7 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
{
sizem32 = m32rLen * sizeof(int);
reservar(&d_R, sizem32);
// cerr << "+ " << d_R << " d_R sz " << sizem32 << endl;
// DEBUG_MEM cerr << "+ " << d_R << " d_R sz " << sizem32 << endl;
cudaMemsetAsync(d_R + rLen, 0x7f, extraspace * sizeof(int));
llenarnosel<<<blockllen, numthreads>>>(p1, d_R, rLen, of1, wherej[0]);
}
@ -1025,7 +1025,7 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
if(posR == NULL)
{
reservar(&posR, sizem32);
// cerr << "+ " << posR << " posR m32 " << sizem32 << endl;
// DEBUG_MEM cerr << "+ " << posR << " posR m32 " << sizem32 << endl;
permutation = thrust::device_pointer_cast(posR);
thrust::sequence(permutation, permutation + m32rLen);
}
@ -1093,7 +1093,7 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
int *d_locations;
reservar(&d_locations, memSizeS);
// cerr << "+ " << d_locations << " d_locs n " << memSizeS << endl;
// DEBUG_MEM cerr << "+ " << d_locations << " d_locs n " << memSizeS << endl;
dim3 Dbs(THRD_PER_BLCK_search, 1, 1);
dim3 Dgs(BLCK_PER_GRID_search, 1, 1);
@ -1133,7 +1133,7 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
cudaMemcpy(dcons, proj, sizepro, cudaMemcpyHostToDevice);
resSize = sum * sizepro;
reservar(&d_Rout, resSize);
// cerr << "+ " << d_Rout << " d_Rout n " << resSize << endl;
// DEBUG_MEM cerr << "+ " << d_Rout << " d_Rout n " << resSize << endl;
if(numj > 2)
{
cudaMemcpy(dcons + rule->num_columns, wherej + 2, muljoinsize, cudaMemcpyHostToDevice);
@ -1148,7 +1148,7 @@ int join(int *p1, int *p2, int rLen, int sLen, int of1, int of2, list<rulenode>:
cudaMemcpy(dcons, proj, sizepro, cudaMemcpyHostToDevice);
resSize = sum * sizepro;
reservar(&d_Rout, resSize);
// cerr << "+ " << d_Rout << " d_Rout 2 " << resSize << endl;
// DEBUG_MEM cerr << "+ " << d_Rout << " d_Rout 2 " << resSize << endl;
if(numj > 2)
{
cudaMemcpy(dcons + projp.y, wherej + 2, muljoinsize, cudaMemcpyHostToDevice);