CUresult
TestSAXPY( chCUDADevice *chDevice, size_t N, float alpha )
{
CUresult status;
CUdeviceptr dptrOut = 0;
CUdeviceptr dptrIn = 0;
float *hostOut = 0;
float *hostIn = 0;
CUDA_CHECK( cuCtxPushCurrent( chDevice->context() ) );
CUDA_CHECK( cuMemAlloc( &dptrOut, N*sizeof(float) ) );
CUDA_CHECK( cuMemsetD32( dptrOut, 0, N ) );
CUDA_CHECK( cuMemAlloc( &dptrIn, N*sizeof(float) ) );
CUDA_CHECK( cuMemHostAlloc( (void **) &hostOut, N*sizeof(float), 0 ) );
CUDA_CHECK( cuMemHostAlloc( (void **) &hostIn, N*sizeof(float), 0 ) );
for ( size_t i = 0; i < N; i++ ) {
hostIn[i] = (float) rand() / (float) RAND_MAX;
}
CUDA_CHECK( cuMemcpyHtoDAsync( dptrIn, hostIn, N*sizeof(float ), NULL ) );
{
CUmodule moduleSAXPY;
CUfunction kernelSAXPY;
void *params[] = { &dptrOut, &dptrIn, &N, &alpha };
moduleSAXPY = chDevice->module( "saxpy.ptx" );
if ( ! moduleSAXPY ) {
status = CUDA_ERROR_NOT_FOUND;
goto Error;
}
CUDA_CHECK( cuModuleGetFunction( &kernelSAXPY, moduleSAXPY, "saxpy" ) );
CUDA_CHECK( cuLaunchKernel( kernelSAXPY, 1500, 1, 1, 512, 1, 1, 0, NULL, params, NULL ) );
}
CUDA_CHECK( cuMemcpyDtoHAsync( hostOut, dptrOut, N*sizeof(float), NULL ) );
CUDA_CHECK( cuCtxSynchronize() );
for ( size_t i = 0; i < N; i++ ) {
if ( fabsf( hostOut[i] - alpha*hostIn[i] ) > 1e-5f ) {
status = CUDA_ERROR_UNKNOWN;
goto Error;
}
}
status = CUDA_SUCCESS;
printf( "Well it worked!\n" );
Error:
cuCtxPopCurrent( NULL );
cuMemFreeHost( hostOut );
cuMemFreeHost( hostIn );
cuMemFree( dptrOut );
cuMemFree( dptrIn );
return status;
}
void* GPUInterface::AllocatePinnedHostMemory(size_t memSize, bool writeCombined, bool mapped) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tEntering GPUInterface::AllocatePinnedHostMemory\n");
#endif
void* ptr;
unsigned int flags = 0;
if (writeCombined)
flags |= CU_MEMHOSTALLOC_WRITECOMBINED;
if (mapped)
flags |= CU_MEMHOSTALLOC_DEVICEMAP;
SAFE_CUPP(cuMemHostAlloc(&ptr, memSize, flags));
#ifdef BEAGLE_DEBUG_VALUES
fprintf(stderr, "Allocated pinned host (CPU) memory %ld to %lu .\n", (long)ptr, ((long)ptr + memSize));
#endif
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tLeaving GPUInterface::AllocatePinnedHostMemory\n");
#endif
return ptr;
}
int gib_alloc ( void **buffers, int buf_size, int *ld, gib_context c ) {
ERROR_CHECK_FAIL(cuCtxPushCurrent(((gpu_context)(c->acc_context))->pCtx));
#if GIB_USE_MMAP
ERROR_CHECK_FAIL(cuMemHostAlloc(buffers, (c->n+c->m)*buf_size,
CU_MEMHOSTALLOC_DEVICEMAP));
#else
ERROR_CHECK_FAIL(cuMemAllocHost(buffers, (c->n+c->m)*buf_size));
#endif
*ld = buf_size;
ERROR_CHECK_FAIL(cuCtxPopCurrent(&((gpu_context)(c->acc_context))->pCtx));
return GIB_SUC;
}
int main(int argc, char *argv[])
{
char c;
CUcontext ctx;
CUdevice dev = 0;
void *toSpace;
int status, free, total;
CUdeviceptr ptr = (CUdeviceptr)NULL;
int size;
if(argc != 2){
fprintf(stderr,"Usage: mem_alloc.exe [MEMORY TO ALLOCATE IN MB]\n");
exit(1);
}
printf("All status results should be 0, if not an error has occured.\nIf 2 is reported an out of memory error has occured for\nwhich you should decrease the memory input\n");
size = atoi(argv[1]);
printf("\nTrying to allocate %iMB of memory on host and GPU\n",size);
if(size <= 0){
fprintf(stderr,"\nERROR: Memory must be greater than 0\n");
exit(1);
}
status = cuInit(0);
printf("Init status: %i\n",status);
status = cuCtxCreate(&ctx, 0, dev);
printf("Context creation status: %i\n",status);
cuMemGetInfo(&free, &total);
printf("Get memory info status: %i\n",status);
printf("\n%.1f/%.1f (Free/Total) MB\n", free/1024.0/1024.0, total/1024.0/1024.0);
status = cuMemHostAlloc(&toSpace, size*1024*1024, 0);
printf("Host allocation status: %i %s\n",status, (status==CUDA_SUCCESS) ? "SUCCESS" : "FAILED");
status = cuMemAlloc(&ptr, size*1024*1024);
printf("GPU allocation status: %i %s\n",status, (status==CUDA_SUCCESS) ? "SUCCESS" : "FAILED");
printf("\nPress any key to exit...");
scanf("%c", &c);
status = cuCtxDestroy(ctx);
printf("Context destroy status: %i\n",status);
return 0;
}
SEXP
R_auto_cuMemHostAlloc(SEXP r_bytesize, SEXP r_Flags)
{
SEXP r_ans = R_NilValue;
void * pp;
size_t bytesize = REAL(r_bytesize)[0];
unsigned int Flags = REAL(r_Flags)[0];
CUresult ans;
ans = cuMemHostAlloc(& pp, bytesize, Flags);
if(ans)
return(R_cudaErrorInfo(ans));
r_ans = R_createRef(pp, "voidPtr") ;
return(r_ans);
}
void *swanMallocHost( size_t len ) {
CUresult err;
void *ptr;
try_init();
#ifdef DEV_EXTENSIONS
err= cuMemHostAlloc( &ptr, len, CU_MEMHOSTALLOC_PORTABLE ); //| CU_MEMHOSTALLOC_DEVICEMAP ); //| CU_MEMHOSTALLOC_WRITECOMBINED );
#else
err = cuMemAllocHost( &ptr, len );
#endif
if ( err != CUDA_SUCCESS ) {
fprintf( stderr, "swanMallocHost error: %d\n", err );
error("swanMallocHost failed\n" );
}
// printf("MallocHost %p\n", ptr );
memset( ptr, 0, len );
return ptr;
}
static void calc_a_score_GPU(FLOAT *ac_score, FLOAT **score,
int *ssize_start, Model_info *MI,
FLOAT scale, int *size_score_array,
int NoC)
{
CUresult res;
const int IHEI = MI->IM_HEIGHT;
const int IWID = MI->IM_WIDTH;
int pady_n = MI->pady;
int padx_n = MI->padx;
int block_pad = (int)(scale/2.0);
struct timeval tv;
int *RY_array, *RX_array;
res = cuMemHostAlloc((void**)&RY_array, NoC*sizeof(int), CU_MEMHOSTALLOC_DEVICEMAP);
if(res != CUDA_SUCCESS) {
printf("cuMemHostAlloc(RY_array) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemHostAlloc((void**)&RX_array, NoC*sizeof(int), CU_MEMHOSTALLOC_DEVICEMAP);
if(res != CUDA_SUCCESS) {
printf("cuMemHostAlloc(RX_array) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
for(int i = 0; i < NoC; i++) {
int rsize[2] = {MI->rsize[i*2], MI->rsize[i*2+1]};
RY_array[i] = (int)((FLOAT)rsize[0]*scale/2.0-1.0+block_pad);
RX_array[i] = (int)((FLOAT)rsize[1]*scale/2.0-1.0+block_pad);
}
CUdeviceptr ac_score_dev, score_dev;
CUdeviceptr ssize_dev, size_score_dev;
CUdeviceptr RY_dev, RX_dev;
int size_score=0;
for(int i = 0; i < NoC; i++) {
size_score += size_score_array[i];
}
/* allocate GPU memory */
res = cuMemAlloc(&ac_score_dev, gpu_size_A_SCORE);
if(res != CUDA_SUCCESS) {
printf("cuMemAlloc(ac_score) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemAlloc(&score_dev, size_score);
if(res != CUDA_SUCCESS) {
printf("cuMemAlloc(score) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemAlloc(&ssize_dev, NoC*sizeof(int));
if(res != CUDA_SUCCESS) {
printf("cuMemAlloc(ssize) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemAlloc(&size_score_dev, NoC*sizeof(int));
if(res != CUDA_SUCCESS) {
printf("cuMemAlloc(size_score) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemAlloc(&RY_dev, NoC*sizeof(int));
if(res != CUDA_SUCCESS) {
printf("cuMemAlloc(RY) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemAlloc(&RX_dev, NoC*sizeof(int));
if(res != CUDA_SUCCESS) {
printf("cuMemAlloc(RX) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
gettimeofday(&tv_memcpy_start, nullptr);
/* upload date to GPU */
res = cuMemcpyHtoD(ac_score_dev, &ac_score[0], gpu_size_A_SCORE);
if(res != CUDA_SUCCESS) {
printf("cuMemcpyHtoD(ac_score) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemcpyHtoD(score_dev, &score[0][0], size_score);
if(res != CUDA_SUCCESS) {
printf("cuMemcpyHtoD(score) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemcpyHtoD(ssize_dev, &ssize_start[0], NoC*sizeof(int));
if(res != CUDA_SUCCESS) {
printf("cuMemcpyHtoD(ssize) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemcpyHtoD(size_score_dev, &size_score_array[0], NoC*sizeof(int));
if(res != CUDA_SUCCESS) {
printf("cuMemcpyHtoD(size_score) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemcpyHtoD(RY_dev, &RY_array[0], NoC*sizeof(int));
if(res != CUDA_SUCCESS) {
printf("cuMemcpyHtoD(RY) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemcpyHtoD(RX_dev, &RX_array[0], NoC*sizeof(int));
if(res != CUDA_SUCCESS) {
printf("cuMemcpyHtoD(RX) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
gettimeofday(&tv_memcpy_end, nullptr);
tvsub(&tv_memcpy_end, &tv_memcpy_start, &tv);
time_memcpy += tv.tv_sec * 1000.0 + (float)tv.tv_usec / 1000.0;
void* kernel_args[] = {
(void*)&IWID,
(void*)&IHEI,
(void*)&scale,
(void*)&padx_n,
(void*)&pady_n,
&RX_dev,
&RY_dev,
&ac_score_dev,
&score_dev,
&ssize_dev,
(void*)&NoC,
&size_score_dev
};
int sharedMemBytes = 0;
/* define CUDA block shape */
int max_threads_num = 0;
int thread_num_x, thread_num_y;
int block_num_x, block_num_y;
res = cuDeviceGetAttribute(&max_threads_num, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, dev[0]);
if(res != CUDA_SUCCESS){
printf("\ncuDeviceGetAttribute() failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
NR_MAXTHREADS_X[0] = (int)sqrt((double)max_threads_num/NoC);
NR_MAXTHREADS_Y[0] = (int)sqrt((double)max_threads_num/NoC);
thread_num_x = (IWID < NR_MAXTHREADS_X[0]) ? IWID : NR_MAXTHREADS_X[0];
thread_num_y = (IHEI < NR_MAXTHREADS_Y[0]) ? IHEI : NR_MAXTHREADS_Y[0];
block_num_x = IWID / thread_num_x;
block_num_y = IHEI / thread_num_y;
if(IWID % thread_num_x != 0) block_num_x++;
if(IHEI % thread_num_y != 0) block_num_y++;
gettimeofday(&tv_kernel_start, nullptr);
/* launch GPU kernel */
res = cuLaunchKernel(
func_calc_a_score[0], // call function
block_num_x, // gridDimX
block_num_y, // gridDimY
1, // gridDimZ
thread_num_x, // blockDimX
thread_num_y, // blockDimY
NoC, // blockDimZ
sharedMemBytes, // sharedMemBytes
nullptr, // hStream
kernel_args, // kernelParams
nullptr // extra
);
if(res != CUDA_SUCCESS) {
printf("cuLaunchKernel(calc_a_score) failed : res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuCtxSynchronize();
if(res != CUDA_SUCCESS) {
printf("cuCtxSynchronize(calc_a_score) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
gettimeofday(&tv_kernel_end, nullptr);
tvsub(&tv_kernel_end, &tv_kernel_start, &tv);
time_kernel += tv.tv_sec * 1000.0 + (float)tv.tv_usec / 1000.0;
gettimeofday(&tv_memcpy_start, nullptr);
/* download data from GPU */
res = cuMemcpyDtoH(ac_score, ac_score_dev, gpu_size_A_SCORE);
if(res != CUDA_SUCCESS) {
printf("cuMemcpyDtoH(ac_score) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
gettimeofday(&tv_memcpy_end, nullptr);
tvsub(&tv_memcpy_end, &tv_memcpy_start, &tv);
time_memcpy += tv.tv_sec * 1000.0 + (float)tv.tv_usec / 1000.0;
/* free GPU memory */
res = cuMemFree(ac_score_dev);
if(res != CUDA_SUCCESS) {
printf("cuMemFree(ac_score_dev) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemFree(score_dev);
if(res != CUDA_SUCCESS) {
printf("cuMemFree(score_dev) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemFree(ssize_dev);
if(res != CUDA_SUCCESS) {
printf("cuMemFree(ssize_dev) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemFree(size_score_dev);
if(res != CUDA_SUCCESS) {
printf("cuMemFree(size_score_dev) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemFree(RY_dev);
if(res != CUDA_SUCCESS) {
printf("cuMemFree(RY_dev) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemFree(RX_dev);
if(res != CUDA_SUCCESS) {
printf("cuMemFree(RX_dev) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
/* free CPU memory */
res = cuMemFreeHost(RY_array);
if(res != CUDA_SUCCESS) {
printf("cuMemFreeHost(RY_array) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
res = cuMemFreeHost(RX_array);
if(res != CUDA_SUCCESS) {
printf("cuMemFreeHost(RX_array) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
}
//detect boundary box
FLOAT *dpm_ttic_gpu_get_boxes(FLOAT **features,FLOAT *scales,int *feature_size, GPUModel *MO,
int *detected_count, FLOAT *acc_score, FLOAT thresh)
{
//constant parameters
const int max_scale = MO->MI->max_scale;
const int interval = MO->MI->interval;
const int sbin = MO->MI->sbin;
const int padx = MO->MI->padx;
const int pady = MO->MI->pady;
const int NoR = MO->RF->NoR;
const int NoP = MO->PF->NoP;
const int NoC = MO->MI->numcomponent;
const int *numpart = MO->MI->numpart;
const int LofFeat=(max_scale+interval)*NoC;
const int L_MAX = max_scale+interval;
/* for measurement */
struct timeval tv;
struct timeval tv_make_c_start, tv_make_c_end;
struct timeval tv_nucom_start, tv_nucom_end;
struct timeval tv_box_start, tv_box_end;
float time_box=0;
struct timeval tv_root_score_start, tv_root_score_end;
float time_root_score = 0;
struct timeval tv_part_score_start, tv_part_score_end;
float time_part_score = 0;
struct timeval tv_dt_start, tv_dt_end;
float time_dt = 0;
struct timeval tv_calc_a_score_start, tv_calc_a_score_end;
float time_calc_a_score = 0;
gettimeofday(&tv_make_c_start, nullptr);
int **RF_size = MO->RF->root_size;
int *rootsym = MO->RF->rootsym;
int *part_sym = MO->PF->part_sym;
int **part_size = MO->PF->part_size;
FLOAT **rootfilter = MO->RF->rootfilter;
FLOAT **partfilter=MO->PF->partfilter;
int **psize = MO->MI->psize;
int **rm_size_array = (int **)malloc(sizeof(int *)*L_MAX);
int **pm_size_array = (int **)malloc(sizeof(int *)*L_MAX);
pm_size_array = (int **)malloc(sizeof(int *)*L_MAX);
FLOAT **Tboxes=(FLOAT**)calloc(LofFeat,sizeof(FLOAT*)); //box coordinate information(Temp)
int *b_nums =(int*)calloc(LofFeat,sizeof(int)); //length of Tboxes
int count = 0;
int detected_boxes=0;
CUresult res;
/* matched score (root and part) */
FLOAT ***rootmatch,***partmatch = nullptr;
int *new_PADsize; // need new_PADsize[L_MAX*3]
size_t SUM_SIZE_feat = 0;
FLOAT **featp2 = (FLOAT **)malloc(L_MAX*sizeof(FLOAT *));
if(featp2 == nullptr) { // error semantics
printf("allocate featp2 failed\n");
exit(1);
}
/* allocate required memory for new_PADsize */
new_PADsize = (int *)malloc(L_MAX*3*sizeof(int));
if(new_PADsize == nullptr) { // error semantics
printf("allocate new_PADsize failed\n");
exit(1);
}
/* do padarray once and reuse it at calculating root and part time */
/* calculate sum of size of padded feature */
for(int tmpL=0; tmpL<L_MAX; tmpL++) {
int PADsize[3] = { feature_size[tmpL*2], feature_size[tmpL*2+1], 31 };
int NEW_Y = PADsize[0] + pady*2;
int NEW_X = PADsize[1] + padx*2;
SUM_SIZE_feat += (NEW_X*NEW_Y*PADsize[2])*sizeof(FLOAT);
}
/* allocate region for padded feat in a lump */
FLOAT *dst_feat;
res = cuMemHostAlloc((void **)&dst_feat, SUM_SIZE_feat, CU_MEMHOSTALLOC_DEVICEMAP);
if(res != CUDA_SUCCESS) {
printf("cuMemHostAlloc(dst_feat) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
memset(dst_feat, 0, SUM_SIZE_feat); // zero clear
/* distribute allocated region */
uintptr_t pointer_feat = (uintptr_t)dst_feat;
for(int tmpL=0; tmpL<L_MAX; tmpL++) {
featp2[tmpL] = (FLOAT *)pointer_feat;
int PADsize[3] = { feature_size[tmpL*2], feature_size[tmpL*2+1], 31 };
int NEW_Y = PADsize[0] + pady*2;
int NEW_X = PADsize[1] + padx*2;
pointer_feat += (uintptr_t)(NEW_X*NEW_Y*PADsize[2]*sizeof(FLOAT));
}
/* copy feat to feat2 */
for(int tmpL=0; tmpL<L_MAX; tmpL++) {
int PADsize[3] = { feature_size[tmpL*2], feature_size[tmpL*2+1], 31 };
int NEW_Y = PADsize[0] + pady*2;
int NEW_X = PADsize[1] + padx*2;
int L = NEW_Y*padx;
int SPL = PADsize[0] + pady;
int M_S = sizeof(FLOAT)*PADsize[0];
FLOAT *P = featp2[tmpL];
FLOAT *S = features[tmpL];
for(int i=0; i<PADsize[2]; i++)
{
P += L;
for(int j=0; j<PADsize[1]; j++)
{
P += pady;
memcpy(P, S, M_S);
S += PADsize[0];
P += SPL;
}
P += L;
}
new_PADsize[tmpL*3] = NEW_Y;
new_PADsize[tmpL*3 + 1] = NEW_X;
new_PADsize[tmpL*3 + 2] = PADsize[2];
}
/* do padarray once and reuse it at calculating root and part time */
/* allocation in a lump */
int *dst_rm_size = (int *)malloc(sizeof(int)*NoC*2*L_MAX);
if(dst_rm_size == nullptr) {
printf("allocate dst_rm_size failed\n");
exit(1);
}
/* distribution to rm_size_array[L_MAX] */
uintptr_t ptr = (uintptr_t)dst_rm_size;
for(int i=0; i<L_MAX; i++) {
rm_size_array[i] = (int *)ptr;
ptr += (uintptr_t)(NoC*2*sizeof(int));
}
/* allocation in a lump */
int *dst_pm_size = (int *)malloc(sizeof(int)*NoP*2*L_MAX);
if(dst_pm_size == nullptr) {
printf("allocate dst_pm_size failed\n");
exit(1);
}
/* distribution to pm_size_array[L_MAX] */
ptr = (uintptr_t)dst_pm_size;
for(int i=0; i<L_MAX; i++) {
pm_size_array[i] = (int *)ptr;
ptr += (uintptr_t)(NoP*2*sizeof(int));
}
///////level
for (int level=interval; level<L_MAX; level++) // feature's loop(A's loop) 1level 1picture
{
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
Tboxes[count]=nullptr;
count++;
continue;
}
} //for (level) // feature's loop(A's loop) 1level 1picture
///////root calculation/////////
/* calculate model score (only root) */
gettimeofday(&tv_root_score_start, nullptr);
rootmatch = fconvsMT_GPU(
featp2,
SUM_SIZE_feat,
rootfilter,
rootsym,
1,
NoR,
new_PADsize,
RF_size, rm_size_array,
L_MAX,
interval,
feature_size,
padx,
pady,
MO->MI->max_X,
MO->MI->max_Y,
ROOT
);
gettimeofday(&tv_root_score_end, nullptr);
tvsub(&tv_root_score_end, &tv_root_score_start, &tv);
time_root_score += tv.tv_sec * 1000.0 + (float)tv.tv_usec / 1000.0;
///////part calculation/////////
if(NoP>0)
{
/* calculate model score (only part) */
gettimeofday(&tv_part_score_start, nullptr);
partmatch = fconvsMT_GPU(
featp2,
SUM_SIZE_feat,
partfilter,
part_sym,
1,
NoP,
new_PADsize,
part_size,
pm_size_array,
L_MAX,
interval,
feature_size,
padx,
pady,
MO->MI->max_X,
MO->MI->max_Y,
PART
);
gettimeofday(&tv_part_score_end, nullptr);
tvsub(&tv_part_score_end, &tv_part_score_start, &tv);
time_part_score += tv.tv_sec * 1000.0 + (float)tv.tv_usec / 1000.0;
}
res = cuCtxSetCurrent(ctx[0]);
if(res != CUDA_SUCCESS) {
printf("cuCtxSetCurrent(ctx[0]) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
gettimeofday(&tv_make_c_end, nullptr);
gettimeofday(&tv_nucom_start, nullptr);
count = 0;
detected_boxes = 0;
int **RL_array = (int **)malloc((L_MAX-interval)*sizeof(int*));
int *dst_RL = (int *) malloc(NoC*(L_MAX-interval)*sizeof(int));
int **RI_array = (int **)malloc((L_MAX-interval)*sizeof(int*));
int *dst_RI = (int *)malloc(NoC*(L_MAX-interval)*sizeof(int));
int **OI_array = (int **)malloc((L_MAX-interval)*sizeof(int*));
int *dst_OI = (int *)malloc((NoC)*(L_MAX-interval)*sizeof(int));
int **RL_S_array = (int **)malloc((L_MAX-interval)*sizeof(int*));
int *dst_RL_S = (int *)malloc(NoC*(L_MAX-interval)*sizeof(int));
FLOAT **OFF_array = (FLOAT **)malloc((L_MAX-interval)*sizeof(FLOAT*));
FLOAT *dst_OFF = (FLOAT *)malloc(NoC*(L_MAX-interval)*sizeof(FLOAT));
FLOAT ***SCORE_array = (FLOAT ***)malloc((L_MAX-interval)*sizeof(FLOAT **));
FLOAT **sub_dst_SCORE = (FLOAT **)malloc(NoC*(L_MAX-interval)*sizeof(FLOAT*));
uintptr_t pointer_RL = (uintptr_t)dst_RL;
uintptr_t pointer_RI = (uintptr_t)dst_RI;
uintptr_t pointer_OI = (uintptr_t)dst_OI;
uintptr_t pointer_RL_S = (uintptr_t)dst_RL_S;
uintptr_t pointer_OFF = (uintptr_t)dst_OFF;
uintptr_t pointer_SCORE = (uintptr_t)sub_dst_SCORE;
for (int level=interval; level<L_MAX; level++) {
int L=level-interval;
RL_array[L] = (int *)pointer_RL;
pointer_RL += (uintptr_t)NoC*sizeof(int);
RI_array[L] = (int *)pointer_RI;
pointer_RI += (uintptr_t)NoC*sizeof(int);
OI_array[L] = (int *)pointer_OI;
pointer_OI += (uintptr_t)NoC*sizeof(int);
RL_S_array[L] = (int *)pointer_RL_S;
pointer_RL_S += (uintptr_t)NoC*sizeof(int);
OFF_array[L] = (FLOAT *)pointer_OFF;
pointer_OFF += (uintptr_t)NoC*sizeof(FLOAT);
SCORE_array[L] = (FLOAT **)pointer_SCORE;
pointer_SCORE += (uintptr_t)NoC*sizeof(FLOAT*);
}
int sum_RL_S = 0;
int sum_SNJ = 0;
/* prepare for parallel execution */
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
continue;
}
for(int j=0; j<NoC; j++) {
/* root score + offset */
RL_array[L][j] = rm_size_array[level][j*2]*rm_size_array[level][j*2+1]; //length of root-matching
RI_array[L][j] = MO->MI->ridx[j]; //root-index
OI_array[L][j] = MO->MI->oidx[j]; //offset-index
RL_S_array[L][j] =sizeof(FLOAT)*RL_array[L][j];
OFF_array[L][j] = MO->MI->offw[RI_array[L][j]]; //offset information
/* search max values */
max_RL_S = (max_RL_S < RL_S_array[L][j]) ? RL_S_array[L][j] : max_RL_S;
max_numpart = (max_numpart < numpart[j]) ? numpart[j] : max_numpart;
}
}
sum_RL_S = max_RL_S*NoC*(L_MAX-interval);
/* root matching size */
sum_SNJ = sizeof(int*)*max_numpart*NoC*(L_MAX-interval);
/* consolidated allocation for SCORE_array and distribute region */
FLOAT *dst_SCORE = (FLOAT *)malloc(sum_RL_S);
pointer_SCORE = (uintptr_t)dst_SCORE;
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
continue;
}
for(int j=0; j<NoC; j++) {
SCORE_array[L][j] = (FLOAT *)pointer_SCORE;
pointer_SCORE += (uintptr_t)max_RL_S;
}
}
/* add offset */
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
continue;
}
for(int j=0; j<NoC; j++) {
memcpy(SCORE_array[L][j], rootmatch[level][j], RL_S_array[L][j]);
FLOAT *SC_S = SCORE_array[L][j];
FLOAT *SC_E = SCORE_array[L][j]+RL_array[L][j];
while(SC_S<SC_E) *(SC_S++)+=OFF_array[L][j];
}
}
/* anchor matrix */ // consolidated allocation
int ***ax_array = (int ***)malloc((L_MAX-interval)*sizeof(int **));
int **sub_dst_ax = (int **)malloc(NoC*(L_MAX-interval)*sizeof(int *));
int *dst_ax = (int *)malloc(sum_SNJ);
int ***ay_array = (int ***)malloc((L_MAX-interval)*sizeof(int **));
int **sub_dst_ay = (int **)malloc(NoC*(L_MAX-interval)*sizeof(int *));
int *dst_ay = (int *)malloc(sum_SNJ);
/* boudary index */ // consolidated allocation
int ****Ix_array =(int ****)malloc((L_MAX-interval)*sizeof(int ***));
int ***sub_dst_Ix = (int ***)malloc(NoC*(L_MAX-interval)*sizeof(int **));
int **dst_Ix = (int **)malloc(sum_SNJ);
int ****Iy_array = (int ****)malloc((L_MAX-interval)*sizeof(int ***));
int ***sub_dst_Iy = (int ***)malloc(NoC*(L_MAX-interval)*sizeof(int **));
int **dst_Iy = (int **)malloc(sum_SNJ);
/* distribute region */
uintptr_t pointer_ax = (uintptr_t)sub_dst_ax;
uintptr_t pointer_ay = (uintptr_t)sub_dst_ay;
uintptr_t pointer_Ix = (uintptr_t)sub_dst_Ix;
uintptr_t pointer_Iy = (uintptr_t)sub_dst_Iy;
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
continue;
}
ax_array[L] = (int **)pointer_ax;
pointer_ax += (uintptr_t)(NoC*sizeof(int*));
ay_array[L] = (int **)pointer_ay;
pointer_ay += (uintptr_t)(NoC*sizeof(int*));
Ix_array[L] = (int ***)pointer_Ix;
pointer_Ix += (uintptr_t)(NoC*sizeof(int**));
Iy_array[L] = (int ***)pointer_Iy;
pointer_Iy += (uintptr_t)(NoC*sizeof(int**));
}
pointer_ax = (uintptr_t)dst_ax;
pointer_ay = (uintptr_t)dst_ay;
pointer_Ix = (uintptr_t)dst_Ix;
pointer_Iy = (uintptr_t)dst_Iy;
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
continue;
}
for(int j=0; j<NoC; j++) {
uintptr_t pointer_offset = sizeof(int*)*max_numpart;
ax_array[L][j] = (int *)pointer_ax;
pointer_ax += pointer_offset;
ay_array[L][j] = (int *)pointer_ay;
pointer_ay += pointer_offset;
Ix_array[L][j] = (int **)pointer_Ix;
pointer_Ix += pointer_offset;
Iy_array[L][j] = (int **)pointer_Iy;
pointer_Iy += pointer_offset;
}
}
/* add parts */
if(NoP>0)
{
/* arrays to store temporary loop variables */
int tmp_array_size = 0;
for(int level=interval; level<L_MAX; level++) {
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
continue;
}
for(int j=0; j<NoC; j++) {
tmp_array_size += max_numpart*sizeof(int);
}
}
int ***DIDX_array = (int ***)malloc((L_MAX-interval)*sizeof(int**));
int **sub_dst_DIDX = (int **)malloc(NoC*(L_MAX-interval)*sizeof(int*));
int *dst_DIDX = (int *)malloc(tmp_array_size);
int ***DID_4_array = (int ***)malloc((L_MAX-interval)*sizeof(int **));
int **sub_dst_DID_4 = (int **)malloc(NoC*(L_MAX-interval)*sizeof(int*));
int *dst_DID_4;
res = cuMemHostAlloc((void **)&dst_DID_4, tmp_array_size, CU_MEMHOSTALLOC_DEVICEMAP);
if(res != CUDA_SUCCESS) {
printf("cuMemHostAlloc(dst_DID_4) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
int ***PIDX_array = (int ***)malloc((L_MAX-interval)*sizeof(int **));
int **sub_dst_PIDX = (int **)malloc(NoC*(L_MAX-interval)*sizeof(int*));
int *dst_PIDX;
res = cuMemHostAlloc((void **)&dst_PIDX, tmp_array_size, CU_MEMHOSTALLOC_DEVICEMAP);
if(res != CUDA_SUCCESS) {
printf("cuMemHostAlloc(dst_PIDX) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
/* distribute consolidated region */
uintptr_t pointer_DIDX = (uintptr_t)sub_dst_DIDX;
uintptr_t pointer_DID_4 = (uintptr_t)sub_dst_DID_4;
uintptr_t pointer_PIDX = (uintptr_t)sub_dst_PIDX;
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
continue;
}
DIDX_array[L] = (int **)pointer_DIDX;
pointer_DIDX += (uintptr_t)(NoC*sizeof(int*));
DID_4_array[L] = (int **)pointer_DID_4;
pointer_DID_4 += (uintptr_t)(NoC*sizeof(int*));
PIDX_array[L] = (int **)pointer_PIDX;
pointer_PIDX += (uintptr_t)(NoC*sizeof(int*));
}
pointer_DIDX = (uintptr_t)dst_DIDX;
pointer_DID_4 = (uintptr_t)dst_DID_4;
pointer_PIDX = (uintptr_t)dst_PIDX;
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X)) {
continue;
}
for(int j=0; j<NoC; j++) {
uintptr_t pointer_offset = (uintptr_t)(max_numpart*sizeof(int));
DIDX_array[L][j] = (int *)pointer_DIDX;
pointer_DIDX += pointer_offset;
DID_4_array[L][j] = (int *)pointer_DID_4;
pointer_DID_4 += pointer_offset;
PIDX_array[L][j] = (int *)pointer_PIDX;
pointer_PIDX += pointer_offset;
}
}
/* prepare for parallel execution */
int sum_size_index_matrix = 0;
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X)) {
continue;
}
for(int j=0; j<NoC; j++) {
for (int k=0;k<numpart[j];k++) {
/* assign values to each element */
DIDX_array[L][j][k] = MO->MI->didx[j][k];
DID_4_array[L][j][k] = DIDX_array[L][j][k]*4;
PIDX_array[L][j][k] = MO->MI->pidx[j][k];
/* anchor */
ax_array[L][j][k] = MO->MI->anchor[DIDX_array[L][j][k]*2]+1;
ay_array[L][j][k] = MO->MI->anchor[DIDX_array[L][j][k]*2+1]+1;
int PSSIZE[2] ={pm_size_array[L][PIDX_array[L][j][k]*2], pm_size_array[L][PIDX_array[L][j][k]*2+1]}; // size of C
/* index matrix */
sum_size_index_matrix += sizeof(int)*PSSIZE[0]*PSSIZE[1];
}
}
}
int *dst_Ix_kk = (int *)malloc(sum_size_index_matrix);
int *dst_Iy_kk = (int *)malloc(sum_size_index_matrix);
uintptr_t pointer_Ix_kk = (uintptr_t)dst_Ix_kk;
uintptr_t pointer_Iy_kk = (uintptr_t)dst_Iy_kk;
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
continue;
}
for(int j=0; j<NoC; j++) {
for (int k=0;k<numpart[j];k++) {
int PSSIZE[2] ={pm_size_array[L][PIDX_array[L][j][k]*2], pm_size_array[L][PIDX_array[L][j][k]*2+1]}; // size of C
Ix_array[L][j][k] = (int *)pointer_Ix_kk;
Iy_array[L][j][k] = (int *)pointer_Iy_kk;
pointer_Ix_kk += (uintptr_t)(sizeof(int)*PSSIZE[0]*PSSIZE[1]);
pointer_Iy_kk += (uintptr_t)(sizeof(int)*PSSIZE[0]*PSSIZE[1]);
}
}
}
gettimeofday(&tv_dt_start, nullptr);
FLOAT ****M_array = dt_GPU(
Ix_array, // int ****Ix_array
Iy_array, // int ****Iy_array
PIDX_array, // int ***PIDX_array
pm_size_array, // int **size_array
NoP, // int NoP
numpart, // int *numpart
NoC, // int NoC
interval, // int interval
L_MAX, // int L_MAX
feature_size, // int *feature_size,
padx, // int padx,
pady, // int pady,
MO->MI->max_X, // int max_X
MO->MI->max_Y, // int max_Y
MO->MI->def, // FLOAT *def
tmp_array_size, // int tmp_array_size
dst_PIDX, // int *dst_PIDX
dst_DID_4 // int *DID_4
);
gettimeofday(&tv_dt_end, nullptr);
tvsub(&tv_dt_end, &tv_dt_start, &tv);
time_dt += tv.tv_sec * 1000.0 + (float)tv.tv_usec / 1000.0;
/* add part score */
for(int level=interval; level<L_MAX; level++){
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
continue;
}
for(int j=0; j<NoC; j++) {
for(int k=0; k<numpart[j]; k++) {
int PSSIZE[2] ={pm_size_array[L][PIDX_array[L][j][k]*2],
pm_size_array[L][PIDX_array[L][j][k]*2+1]}; // Size of C
int R_S[2]={rm_size_array[level][j*2], rm_size_array[level][j*2+1]};
dpm_ttic_add_part_calculation(SCORE_array[L][j], M_array[L][j][k], R_S,
PSSIZE, ax_array[L][j][k], ay_array[L][j][k]);
}
}
}
s_free(M_array[0][0][0]);
s_free(M_array[0][0]);
s_free(M_array[0]);
s_free(M_array);
/* free temporary arrays */
free(dst_DIDX);
free(sub_dst_DIDX);
free(DIDX_array);
res = cuMemFreeHost(dst_DID_4);
if(res != CUDA_SUCCESS) {
printf("cuMemFreeHost(dst_DID_4) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
free(sub_dst_DID_4);
free(DID_4_array);
res = cuMemFreeHost(dst_PIDX);
if(res != CUDA_SUCCESS) {
printf("cuMemFreeHost(dst_PIDX) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
free(sub_dst_PIDX);
free(PIDX_array);
res = cuCtxSetCurrent(ctx[0]);
if(res != CUDA_SUCCESS) {
printf("cuCtxSetCurrent(ctx[0]) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
} // start from if(NoP>0)
/* combine root and part score and detect boundary box for each-component */
FLOAT *scale_array = (FLOAT *)malloc((L_MAX-interval)*sizeof(FLOAT));
for(int level=interval; level<L_MAX; level++) {
int L = level - interval;
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X))
{
Tboxes[count]=nullptr;
count++;
continue;
}
scale_array[L] = (FLOAT)sbin/scales[level];
}
for (int level=interval; level<L_MAX; level++) // feature's loop(A's loop) 1level 1picture
{
/* parameters (related for level) */
int L=level-interval;
/* matched score size matrix */
FLOAT scale=(FLOAT)sbin/scales[level];
/* loop conditon */
if(feature_size[level*2]+2*pady<MO->MI->max_Y ||(feature_size[level*2+1]+2*padx<MO->MI->max_X)) {
Tboxes[count]=nullptr;
count++;
continue;
}
/* calculate accumulated score */
gettimeofday(&tv_calc_a_score_start, nullptr);
calc_a_score_GPU(
acc_score, // FLOAT *ac_score
SCORE_array[L], // FLOAT **score
rm_size_array[level], // int *ssize_start
MO->MI, // Model_info *MI
scale, // FLOAT scale
RL_S_array[L], // int *size_score_array
NoC // int NoC
);
gettimeofday(&tv_calc_a_score_end, nullptr);
tvsub(&tv_calc_a_score_end, &tv_calc_a_score_start, &tv);
time_calc_a_score += tv.tv_sec * 1000.0 + (float)tv.tv_usec / 1000.0;
for(int j = 0; j <NoC; j++) {
int R_S[2]={rm_size_array[level][j*2], rm_size_array[level][j*2+1]};
/* get all good matches */
int GMN;
int *GMPC = get_gmpc(SCORE_array[L][j],thresh,R_S,&GMN);
int RSIZE[2]={MO->MI->rsize[j*2], MO->MI->rsize[j*2+1]};
int GL = (numpart[j]+1)*4+3; //31
/* detected box coordinate(current level) */
FLOAT *t_boxes = (FLOAT*)calloc(GMN*GL,sizeof(FLOAT));
gettimeofday(&tv_box_start, nullptr);
// NO NEED TO USE GPU
for(int k = 0;k < GMN;k++) {
FLOAT *P_temp = t_boxes+GL*k;
int y = GMPC[2*k];
int x = GMPC[2*k+1];
/* calculate root box coordinate */
FLOAT *RB =rootbox(x,y,scale,padx,pady,RSIZE);
memcpy(P_temp, RB,sizeof(FLOAT)*4);
s_free(RB);
P_temp+=4;
for(int pp=0;pp<numpart[j];pp++) {
int PBSIZE[2]={psize[j][pp*2], psize[j][pp*2+1]};
int Isize[2]={pm_size_array[L][MO->MI->pidx[j][pp]*2], pm_size_array[L][MO->MI->pidx[j][pp]*2+1]};
/* calculate part box coordinate */
FLOAT *PB = partbox(x,y,ax_array[L][j][pp],ay_array[L][j][pp],scale,padx,pady,PBSIZE,Ix_array[L][j][pp],Iy_array[L][j][pp],Isize);
memcpy(P_temp, PB,sizeof(FLOAT)*4);
P_temp+=4;
s_free(PB);
}
/* component number and score */
*(P_temp++)=(FLOAT)j; //component number
*(P_temp++)=SCORE_array[L][j][x*R_S[0]+y]; //score of good match
*P_temp = scale;
}
// NO NEED TO USE GPU
gettimeofday(&tv_box_end, nullptr);
tvsub(&tv_box_end, &tv_box_start, &tv);
time_box += tv.tv_sec * 1000.0 + (float)tv.tv_usec / 1000.0;
/* save box information */
if (GMN > 0)
Tboxes[count] = t_boxes;
else
Tboxes[count] = nullptr;
b_nums[count]=GMN;
count++;
detected_boxes+=GMN; //number of detected box
/* release */
s_free(GMPC);
}
////numcom
}
////level
/* free temporary arrays */
free(dst_RL);
free(RL_array);
free(dst_RI);
free(RI_array);
free(dst_OI);
free(OI_array);
free(dst_RL_S);
free(RL_S_array);
free(dst_OFF);
free(OFF_array);
free(dst_SCORE);
free(sub_dst_SCORE);
free(SCORE_array);
free(dst_ax);
free(sub_dst_ax);
free(ax_array);
free(dst_ay);
free(sub_dst_ay);
free(ay_array);
free(Ix_array[0][0][0]);
free(dst_Ix);
free(sub_dst_Ix);
free(Ix_array);
free(Iy_array[0][0][0]);
free(dst_Iy);
free(sub_dst_Iy);
free(Iy_array);
free(scale_array);
gettimeofday(&tv_nucom_end, nullptr);
#ifdef PRINT_INFO
printf("root SCORE : %f\n", time_root_score);
printf("part SCORE : %f\n", time_part_score);
printf("dt : %f\n", time_dt);
printf("calc_a_score : %f\n", time_calc_a_score);
#endif
res = cuCtxSetCurrent(ctx[0]);
if(res != CUDA_SUCCESS) {
printf("cuCtxSetCurrent(ctx[0]) failed: res = %s\n",cuda_response_to_string(res));
exit(1);
}
/* free memory regions */
res = cuMemFreeHost((void *)featp2[0]);
if(res != CUDA_SUCCESS) {
printf("cuMemFreeHost(featp2[0]) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
s_free(featp2);
res = cuMemFreeHost((void *)rootmatch[interval][0]);
if(res != CUDA_SUCCESS) {
printf("cuMemFreeHost(rootmatch[0][0]) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
s_free(rootmatch[0]);
s_free(rootmatch);
if (partmatch != nullptr) {
res = cuMemFreeHost((void *)partmatch[0][0]);
if(res != CUDA_SUCCESS) {
printf("cuMemFreeHost(partmatch[0][0]) failed: res = %s\n", cuda_response_to_string(res));
exit(1);
}
s_free(partmatch[0]);
s_free(partmatch);
s_free(new_PADsize);
}
/* release */
s_free(rm_size_array[0]);
s_free(rm_size_array);
s_free(pm_size_array[0]);
s_free(pm_size_array);
/* Output boundary-box coorinate information */
int GL=(numpart[0]+1)*4+3;
FLOAT *boxes=(FLOAT*)calloc(detected_boxes*GL,sizeof(FLOAT)); //box coordinate information(Temp)
FLOAT *T1 = boxes;
for(int i = 0; i < LofFeat; i++) {
int num_t = b_nums[i]*GL;
if(num_t > 0) {
FLOAT *T2 = Tboxes[i];
//memcpy_s(T1,sizeof(FLOAT)*num_t,T2,sizeof(FLOAT)*num_t);
memcpy(T1, T2,sizeof(FLOAT)*num_t);
T1 += num_t;
}
}
FLOAT abs_threshold = abs(thresh);
/* accumulated score calculation */
FLOAT max_score = 0.0;
/* add offset to accumulated score */
for(int i = 0; i < MO->MI->IM_HEIGHT*MO->MI->IM_WIDTH; i++) {
if (acc_score[i] < thresh) {
acc_score[i] = 0.0;
} else {
acc_score[i] += abs_threshold;
if (acc_score[i] > max_score)
max_score = acc_score[i];
}
}
/* normalization */
if (max_score > 0.0) {
FLOAT ac_ratio = 1.0 / max_score;
for (int i = 0; i < MO->MI->IM_HEIGHT*MO->MI->IM_WIDTH; i++) {
acc_score[i] *= ac_ratio;
}
}
/* release */
free_boxes(Tboxes,LofFeat);
s_free(b_nums);
/* output result */
*detected_count = detected_boxes;
return boxes;
}
cl_int
pocl_cuda_alloc_mem_obj (cl_device_id device, cl_mem mem_obj, void *host_ptr)
{
cuCtxSetCurrent (((pocl_cuda_device_data_t *)device->data)->context);
CUresult result;
void *b = NULL;
/* if memory for this global memory is not yet allocated -> do it */
if (mem_obj->device_ptrs[device->global_mem_id].mem_ptr == NULL)
{
cl_mem_flags flags = mem_obj->flags;
if (flags & CL_MEM_USE_HOST_PTR)
{
#if defined __arm__
// cuMemHostRegister is not supported on ARN
// Allocate device memory and perform explicit copies
// before and after running a kernel
result = cuMemAlloc ((CUdeviceptr *)&b, mem_obj->size);
CUDA_CHECK (result, "cuMemAlloc");
#else
result = cuMemHostRegister (host_ptr, mem_obj->size,
CU_MEMHOSTREGISTER_DEVICEMAP);
if (result != CUDA_SUCCESS
&& result != CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED)
CUDA_CHECK (result, "cuMemHostRegister");
result = cuMemHostGetDevicePointer ((CUdeviceptr *)&b, host_ptr, 0);
CUDA_CHECK (result, "cuMemHostGetDevicePointer");
#endif
}
else if (flags & CL_MEM_ALLOC_HOST_PTR)
{
result = cuMemHostAlloc (&mem_obj->mem_host_ptr, mem_obj->size,
CU_MEMHOSTREGISTER_DEVICEMAP);
CUDA_CHECK (result, "cuMemHostAlloc");
result = cuMemHostGetDevicePointer ((CUdeviceptr *)&b,
mem_obj->mem_host_ptr, 0);
CUDA_CHECK (result, "cuMemHostGetDevicePointer");
}
else
{
result = cuMemAlloc ((CUdeviceptr *)&b, mem_obj->size);
if (result != CUDA_SUCCESS)
{
const char *err;
cuGetErrorName (result, &err);
POCL_MSG_PRINT2 (__FUNCTION__, __LINE__,
"-> Failed to allocate memory: %s\n", err);
return CL_MEM_OBJECT_ALLOCATION_FAILURE;
}
}
if (flags & CL_MEM_COPY_HOST_PTR)
{
result = cuMemcpyHtoD ((CUdeviceptr)b, host_ptr, mem_obj->size);
CUDA_CHECK (result, "cuMemcpyHtoD");
}
mem_obj->device_ptrs[device->global_mem_id].mem_ptr = b;
mem_obj->device_ptrs[device->global_mem_id].global_mem_id
= device->global_mem_id;
}
/* copy already allocated global mem info to devices own slot */
mem_obj->device_ptrs[device->dev_id]
= mem_obj->device_ptrs[device->global_mem_id];
return CL_SUCCESS;
}
/*
* Class: edu_syr_pcpratts_rootbeer_runtime2_cuda_CudaRuntime2
* Method: findReserveMem
* Signature: ()I
*/
JNIEXPORT jlong JNICALL Java_edu_syr_pcpratts_rootbeer_runtime2_cuda_CudaRuntime2_findReserveMem
(JNIEnv * env, jobject this_ref, jint max_blocks_per_proc, jint max_threads_per_block)
{
size_t to_space_size;
size_t temp_size;
int status;
int deviceCount = 0;
jlong prev_i;
jlong i;
size_t f_mem;
size_t t_mem;
jint num_blocks;
status = cuInit(0);
CHECK_STATUS(env,"error in cuInit",status)
printf("automatically determining CUDA reserve space...\n");
to_space_size = initContext(env, max_blocks_per_proc, max_threads_per_block);
//space for 100 types in the scene
classMemSize = sizeof(jint)*100;
num_blocks = numMultiProcessors * max_threads_per_block * max_blocks_per_proc;
gc_space_size = 1024;
to_space_size -= (num_blocks * sizeof(jlong));
to_space_size -= (num_blocks * sizeof(jlong));
to_space_size -= gc_space_size;
to_space_size -= classMemSize;
for(i = 1024L*1024L; i < to_space_size; i += 100L*1024L*1024L){
temp_size = to_space_size - i;
printf("attempting allocation with temp_size: %lu to_space_size: %lu i: %ld\n", temp_size, to_space_size, i);
status = cuMemHostAlloc(&toSpace, temp_size, 0);
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
status = cuMemAlloc(&gpuToSpace, temp_size);
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
status = cuMemAlloc(&gpuClassMemory, classMemSize);
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
status = cuMemHostAlloc(&handlesMemory, num_blocks * sizeof(jlong), CU_MEMHOSTALLOC_WRITECOMBINED);
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
status = cuMemAlloc(&gpuHandlesMemory, num_blocks * sizeof(jlong));
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
status = cuMemHostAlloc(&exceptionsMemory, num_blocks * sizeof(jlong), 0);
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
status = cuMemAlloc(&gpuExceptionsMemory, num_blocks * sizeof(jlong));
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
status = cuMemAlloc(&gcInfoSpace, gc_space_size);
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
status = cuMemAlloc(&gpuHeapEndPtr, 8);
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
status = cuMemAlloc(&gpuBufferSize, 8);
if(status != CUDA_SUCCESS){
cuCtxDestroy(cuContext);
initContext(env, max_blocks_per_proc, max_threads_per_block);
continue;
}
//done, free everything
cuMemFree(gpuToSpace);
cuMemFree(gpuClassMemory);
cuMemFree(gpuHandlesMemory);
cuMemFree(gpuExceptionsMemory);
cuMemFree(gcInfoSpace);
cuMemFree(gpuHeapEndPtr);
cuMemFree(gpuBufferSize);
cuMemFreeHost(toSpace);
cuMemFreeHost(handlesMemory);
cuMemFreeHost(exceptionsMemory);
return i;
}
throw_cuda_errror_exception(env, "unable to find enough space using CUDA", 0);
return 0;
}
void initDevice(JNIEnv * env, jobject this_ref, jint max_blocks_per_proc, jint max_threads_per_block, jlong free_space)
{
int status;
jint num_blocks;
int deviceCount = 0;
size_t f_mem;
size_t t_mem;
size_t to_space_size;
textureMemSize = 1;
status = cuDeviceGetCount(&deviceCount);
CHECK_STATUS(env,"error in cuDeviceGetCount",status)
getBestDevice(env);
status = cuCtxCreate(&cuContext, CU_CTX_MAP_HOST, cuDevice);
CHECK_STATUS(env,"error in cuCtxCreate",status)
status = cuMemGetInfo (&f_mem, &t_mem);
CHECK_STATUS(env,"error in cuMemGetInfo",status)
to_space_size = f_mem;
num_blocks = numMultiProcessors * max_threads_per_block * max_blocks_per_proc;
#if DEBUG
printf("Memory: %i(MB)/%i(MB) (Free/Total)\n",f_mem/1024/1024, t_mem/1024/1024);
printf("num_blocks = %i\n",num_blocks);
printf("numMultiProcessors = %i\n",numMultiProcessors);
printf("max_threads_per_block = %i\n",max_threads_per_block);
printf("max_blocks_per_proc = %i\n",max_blocks_per_proc);
fflush(stdout);
#endif
//space for 100 types in the scene
classMemSize = sizeof(jint)*100;
gc_space_size = 1024;
to_space_size -= (num_blocks * sizeof(jlong));
to_space_size -= (num_blocks * sizeof(jlong));
to_space_size -= gc_space_size;
to_space_size -= free_space;
to_space_size -= classMemSize;
//to_space_size -= textureMemSize;
bufferSize = to_space_size;
status = cuMemHostAlloc(&toSpace, to_space_size, 0);
CHECK_STATUS(env,"toSpace memory allocation failed",status)
status = cuMemAlloc(&gpuToSpace, to_space_size);
CHECK_STATUS(env,"gpuToSpace memory allocation failed",status)
status = cuMemAlloc(&gpuClassMemory, classMemSize);
CHECK_STATUS(env,"gpuClassMemory memory allocation failed",status)
/*
status = cuMemHostAlloc(&textureMemory, textureMemSize, 0);
if (CUDA_SUCCESS != status)
{
printf("error in cuMemHostAlloc textureMemory %d\n", status);
}
status = cuMemAlloc(&gpuTexture, textureMemSize);
if (CUDA_SUCCESS != status)
{
printf("error in cuMemAlloc gpuTexture %d\n", status);
}
*/
status = cuMemHostAlloc(&handlesMemory, num_blocks * sizeof(jlong), CU_MEMHOSTALLOC_WRITECOMBINED);
CHECK_STATUS(env,"handlesMemory memory allocation failed",status)
status = cuMemAlloc(&gpuHandlesMemory, num_blocks * sizeof(jlong));
CHECK_STATUS(env,"gpuHandlesMemory memory allocation failed",status)
status = cuMemHostAlloc(&exceptionsMemory, num_blocks * sizeof(jlong), 0);
CHECK_STATUS(env,"exceptionsMemory memory allocation failed",status)
status = cuMemAlloc(&gpuExceptionsMemory, num_blocks * sizeof(jlong));
CHECK_STATUS(env,"gpuExceptionsMemory memory allocation failed",status)
status = cuMemAlloc(&gcInfoSpace, gc_space_size);
CHECK_STATUS(env,"gcInfoSpace memory allocation failed",status)
status = cuMemAlloc(&gpuHeapEndPtr, 8);
CHECK_STATUS(env,"gpuHeapEndPtr memory allocation failed",status)
status = cuMemAlloc(&gpuBufferSize, 8);
CHECK_STATUS(env,"gpuBufferSize memory allocation failed",status)
thisRefClass = (*env)->GetObjectClass(env, this_ref);
setLongField(env, this_ref, "m_ToSpaceAddr", (jlong) toSpace);
setLongField(env, this_ref, "m_GpuToSpaceAddr", (jlong) gpuToSpace);
setLongField(env, this_ref, "m_TextureAddr", (jlong) textureMemory);
setLongField(env, this_ref, "m_GpuTextureAddr", (jlong) gpuTexture);
setLongField(env, this_ref, "m_HandlesAddr", (jlong) handlesMemory);
setLongField(env, this_ref, "m_GpuHandlesAddr", (jlong) gpuHandlesMemory);
setLongField(env, this_ref, "m_ExceptionsHandlesAddr", (jlong) exceptionsMemory);
setLongField(env, this_ref, "m_GpuExceptionsHandlesAddr", (jlong) gpuExceptionsMemory);
setLongField(env, this_ref, "m_ToSpaceSize", (jlong) bufferSize);
setLongField(env, this_ref, "m_MaxGridDim", (jlong) maxGridDim);
setLongField(env, this_ref, "m_NumMultiProcessors", (jlong) numMultiProcessors);
}
/*
* Class: edu_syr_pcpratts_rootbeer_runtime2_cuda_CudaRuntime2
* Method: setup
* Signature: ()V
*/
JNIEXPORT void JNICALL Java_edu_syr_pcpratts_rootbeer_runtime2_cuda_CudaRuntime2_setup
(JNIEnv *env, jobject this_ref, jint max_blocks_per_proc, jint max_threads_per_block, jint free_space)
{
int status;
jint num_blocks;
int deviceCount = 0;
size_t f_mem;
size_t t_mem;
size_t to_space_size;
//size_t free_space = 1530L*1024L*1024L;
textureMemSize = 1;
status = cuInit(0);
if (CUDA_SUCCESS != status)
{
printf("error in cuInit\n");
}
status = cuDeviceGetCount(&deviceCount);
if (CUDA_SUCCESS != status)
{
printf("error in cuDeviceGet\n");
}
getBestDevice();
status = cuCtxCreate(&cuContext, CU_CTX_MAP_HOST, cuDevice);
if (CUDA_SUCCESS != status)
{
printf("error in cuCtxCreate %d\n", status);
}
// ddb - not using this as this returns the total memory not the free memory
//to_space_size = memSize();
cuMemGetInfo(&f_mem, &t_mem);
to_space_size = f_mem;
num_blocks = numMultiProcessors * max_threads_per_block * max_blocks_per_proc;
#if DEBUG
printf("Memory: %i(MB)/%i(MB) (Free/Total)\n",f_mem/1024/1024, t_mem/1024/1024);
printf("num_blocks = %i\n",num_blocks);
printf("numMultiProcessors = %i\n",numMultiProcessors);
printf("max_threads_per_block = %i\n",max_threads_per_block);
printf("max_blocks_per_proc = %i\n",max_blocks_per_proc);
fflush(stdout);
#endif
gc_space_size = 1024;
to_space_size -= (num_blocks * sizeof(jlong));
to_space_size -= (num_blocks * sizeof(jlong));
to_space_size -= gc_space_size;
to_space_size -= free_space;
//to_space_size -= textureMemSize;
bufferSize = to_space_size;
status = cuMemHostAlloc(&toSpace, to_space_size, 0);
if (CUDA_SUCCESS != status) {
throw_cuda_errror_exception(env, "toSpace memory allocation failed", status);
return;
}
status = cuMemAlloc(&gpuToSpace, to_space_size);
if (CUDA_SUCCESS != status) {
throw_cuda_errror_exception(env, "gpuToSpace memory allocation failed", status);
return;
}
/*
status = cuMemHostAlloc(&textureMemory, textureMemSize, 0);
if (CUDA_SUCCESS != status)
{
printf("error in cuMemHostAlloc textureMemory %d\n", status);
}
status = cuMemAlloc(&gpuTexture, textureMemSize);
if (CUDA_SUCCESS != status)
{
printf("error in cuMemAlloc gpuTexture %d\n", status);
}
*/
status = cuMemHostAlloc(&handlesMemory, num_blocks * sizeof(jlong), CU_MEMHOSTALLOC_WRITECOMBINED);
if (CUDA_SUCCESS != status) {
throw_cuda_errror_exception(env, "handlesMemory memory allocation failed", status);
return;
}
status = cuMemAlloc(&gpuHandlesMemory, num_blocks * sizeof(jlong));
if (CUDA_SUCCESS != status) {
throw_cuda_errror_exception(env, "gpuHandlesMemory memory allocation failed", status);
return;
}
status = cuMemHostAlloc(&exceptionsMemory, num_blocks * sizeof(jlong), 0);
if (CUDA_SUCCESS != status) {
throw_cuda_errror_exception(env, "exceptionsMemory memory allocation failed", status);
return;
}
status = cuMemAlloc(&gpuExceptionsMemory, num_blocks * sizeof(jlong));
if (CUDA_SUCCESS != status) {
throw_cuda_errror_exception(env, "gpuExceptionsMemory memory allocation failed", status);
return;
}
status = cuMemAlloc(&gcInfoSpace, gc_space_size);
if (CUDA_SUCCESS != status) {
throw_cuda_errror_exception(env, "gcInfoSpace memory allocation failed", status);
return;
}
status = cuMemAlloc(&gpuHeapEndPtr, 8);
if (CUDA_SUCCESS != status) {
throw_cuda_errror_exception(env, "gpuHeapEndPtr memory allocation failed", status);
return;
}
status = cuMemAlloc(&gpuBufferSize, 8);
if (CUDA_SUCCESS != status) {
throw_cuda_errror_exception(env, "gpuBufferSize memory allocation failed", status);
return;
}
thisRefClass = (*env)->GetObjectClass(env, this_ref);
setLongField(env, this_ref, "m_ToSpaceAddr", (jlong) toSpace);
setLongField(env, this_ref, "m_GpuToSpaceAddr", (jlong) gpuToSpace);
setLongField(env, this_ref, "m_TextureAddr", (jlong) textureMemory);
setLongField(env, this_ref, "m_GpuTextureAddr", (jlong) gpuTexture);
setLongField(env, this_ref, "m_HandlesAddr", (jlong) handlesMemory);
setLongField(env, this_ref, "m_GpuHandlesAddr", (jlong) gpuHandlesMemory);
setLongField(env, this_ref, "m_ExceptionsHandlesAddr", (jlong) exceptionsMemory);
setLongField(env, this_ref, "m_GpuExceptionsHandlesAddr", (jlong) gpuExceptionsMemory);
setLongField(env, this_ref, "m_ToSpaceSize", (jlong) bufferSize);
setLongField(env, this_ref, "m_MaxGridDim", (jlong) maxGridDim);
setLongField(env, this_ref, "m_NumMultiProcessors", (jlong) numMultiProcessors);
}
//load model basic information
Model_info * load_modelinfo(char *filename)
{
CUresult res;
FILE *file; //File
errno_t err; //err ( for fopen)
Model_info *MI=(Model_info*)malloc(sizeof(Model_info)); //Model information
//fopen
//if( (err = fopen_s( &file,filename, "r"))!=0 )
if( (file=fopen(filename, "r"))==NULL )
{
printf("Model information file not found \n");
exit(-1);
}
FLOAT t1,t2,t3,t4;
int b;
//load basic information
if( sizeof(FLOAT) == sizeof(double) ){
b =fscanf(file,"%lf,",&t1);
MI->numcomponent=(int)t1; //number of components
b =fscanf(file,"%lf,",&t1);
MI->sbin=(int)t1; //sbin
b =fscanf(file,"%lf,",&t1);
MI->interval=(int)t1; //interval
b =fscanf(file,"%lf,",&t1);
MI->max_Y=(int)t1; //max_Y
b =fscanf(file,"%lf,",&t1);
MI->max_X=(int)t1; //max_X
}else{
b =fscanf(file,"%f,",&t1);
MI->numcomponent=(int)t1; //number of components
b =fscanf(file,"%f,",&t1);
MI->sbin=(int)t1; //sbin
b =fscanf(file,"%f,",&t1);
MI->interval=(int)t1; //interval
b =fscanf(file,"%f,",&t1);
MI->max_Y=(int)t1; //max_Y
b =fscanf(file,"%f,",&t1);
MI->max_X=(int)t1; //max_X
}
//root filter information
MI->ridx = (int*)malloc(sizeof(int)*MI->numcomponent);
MI->oidx = (int*)malloc(sizeof(int)*MI->numcomponent);
MI->offw = (FLOAT*)malloc(sizeof(FLOAT)*MI->numcomponent);
MI->rsize = (int*)malloc(sizeof(int)*MI->numcomponent*2);
MI->numpart = (int*)malloc(sizeof(int)*MI->numcomponent);
//part filter information
MI->pidx = (int**)malloc(sizeof(int*)*MI->numcomponent);
MI->didx = (int**)malloc(sizeof(int*)*MI->numcomponent);
MI->psize = (int**)malloc(sizeof(int*)*MI->numcomponent);
for(int ii=0;ii<MI->numcomponent;ii++) //LOOP (component)
{
if(sizeof(FLOAT)==sizeof(double)) {
b =fscanf(file,"%lf,",&t1);
MI->ridx[ii]=(int)t1-1; //root index
b =fscanf(file,"%lf,",&t1);
MI->oidx[ii]=(int)t1-1; //offset index
b =fscanf(file,"%lf,",&t1);
MI->offw[ii]=t1; //offset weight (FLOAT)
b =fscanf(file,"%lf,%lf,",&t1,&t2);
MI->rsize[ii*2]=(int)t1; //rsize (Y)
MI->rsize[ii*2+1]=(int)t2; //rsize (X)
b =fscanf(file,"%lf,",&t1);
MI->numpart[ii]=(int)t1; //number of part filter
}else{
b =fscanf(file,"%f,",&t1);
MI->ridx[ii]=(int)t1-1; //root index
b =fscanf(file,"%f,",&t1);
MI->oidx[ii]=(int)t1-1; //offset index
b =fscanf(file,"%f,",&t1);
MI->offw[ii]=t1; //offset weight (FLOAT)
b =fscanf(file,"%f,%f,",&t1,&t2);
MI->rsize[ii*2]=(int)t1; //rsize (Y)
MI->rsize[ii*2+1]=(int)t2; //rsize (X)
b =fscanf(file,"%f,",&t1);
MI->numpart[ii]=(int)t1; //number of part filter
}
MI->pidx[ii]=(int*)malloc(sizeof(int)*MI->numpart[ii]);
MI->didx[ii]=(int*)malloc(sizeof(int)*MI->numpart[ii]);
MI->psize[ii]=(int*)malloc(sizeof(int)*MI->numpart[ii]*2);
for(int jj=0;jj<MI->numpart[ii];jj++) //LOOP (part-filter)
{
if(sizeof(FLOAT)==sizeof(double)) {
b =fscanf(file,"%lf,",&t1);
MI->pidx[ii][jj]=(int)t1-1; //part index
b =fscanf(file,"%lf,",&t1);
MI->didx[ii][jj]=(int)t1-1; //define-index of part
b =fscanf(file,"%lf,%lf,",&t1,&t2);
MI->psize[ii][jj*2]=(int)t1;
MI->psize[ii][jj*2+1]=(int)t2;
}else{
b =fscanf(file,"%f,",&t1);
MI->pidx[ii][jj]=(int)t1-1; //part index
b =fscanf(file,"%f,",&t1);
MI->didx[ii][jj]=(int)t1-1; //define-index of part
b =fscanf(file,"%f,%f,",&t1,&t2);
MI->psize[ii][jj*2]=(int)t1;
MI->psize[ii][jj*2+1]=(int)t2;
}
}
}
//get defs information
if(sizeof(FLOAT)==sizeof(double)) {
b =fscanf(file,"%lf,",&t1);
}else{
b =fscanf(file,"%f,",&t1);
}
int DefL = int(t1);
//MI->def = (FLOAT*)malloc(sizeof(FLOAT)*DefL*4);
res = cuMemHostAlloc((void **)&(MI->def), sizeof(FLOAT)*DefL*4, CU_MEMHOSTALLOC_DEVICEMAP);
if(res != CUDA_SUCCESS) {
printf("cuMemHostAlloc(MI->def) failed: res = %s\n", conv(res));
exit(1);
}
sum_size_def_array = sizeof(FLOAT)*DefL*4;
MI->anchor = (int*)malloc(sizeof(int)*DefL*2);
for (int kk=0;kk<DefL;kk++)
{
if(sizeof(FLOAT)==sizeof(double)) {
b =fscanf(file,"%lf,%lf,%lf,%lf,",&t1,&t2,&t3,&t4);
MI->def[kk*4]=t1;
MI->def[kk*4+1]=t2;
MI->def[kk*4+2]=t3;
MI->def[kk*4+3]=t4;
b =fscanf(file,"%lf,%lf,",&t1,&t2);
MI->anchor[kk*2]=(int)t1;
MI->anchor[kk*2+1]=(int)t2;
}else{
b =fscanf(file,"%f,%f,%f,%f,",&t1,&t2,&t3,&t4);
MI->def[kk*4]=t1;
MI->def[kk*4+1]=t2;
MI->def[kk*4+2]=t3;
MI->def[kk*4+3]=t4;
b =fscanf(file,"%f,%f,",&t1,&t2);
MI->anchor[kk*2]=(int)t1;
MI->anchor[kk*2+1]=(int)t2;
}
}
//get least_square information
MI->x1 = (FLOAT **)malloc(sizeof(FLOAT*)*MI->numcomponent);
MI->x2 = (FLOAT **)malloc(sizeof(FLOAT*)*MI->numcomponent);
MI->y1 = (FLOAT **)malloc(sizeof(FLOAT*)*MI->numcomponent);
MI->y2 = (FLOAT **)malloc(sizeof(FLOAT*)*MI->numcomponent);
for(int ii=0;ii<MI->numcomponent;ii++)
{
int GL = 1+2*(1+MI->numpart[ii]);
MI->x1[ii] =(FLOAT *)malloc(sizeof(FLOAT)*GL);
MI->y1[ii] =(FLOAT *)malloc(sizeof(FLOAT)*GL);
MI->x2[ii] =(FLOAT *)malloc(sizeof(FLOAT)*GL);
MI->y2[ii] =(FLOAT *)malloc(sizeof(FLOAT)*GL);
if(sizeof(FLOAT)==sizeof(double)) {
for (int jj=0;jj<GL;jj++){b =fscanf(file,"%lf,",&t1); MI->x1[ii][jj]=t1;}
for (int jj=0;jj<GL;jj++){b =fscanf(file,"%lf,",&t1); MI->y1[ii][jj]=t1;}
for (int jj=0;jj<GL;jj++){b =fscanf(file,"%lf,",&t1); MI->x2[ii][jj]=t1;}
for (int jj=0;jj<GL;jj++){b =fscanf(file,"%lf,",&t1); MI->y2[ii][jj]=t1;}
}else{
for (int jj=0;jj<GL;jj++){b =fscanf(file,"%f,",&t1); MI->x1[ii][jj]=t1;}
for (int jj=0;jj<GL;jj++){b =fscanf(file,"%f,",&t1); MI->y1[ii][jj]=t1;}
for (int jj=0;jj<GL;jj++){b =fscanf(file,"%f,",&t1); MI->x2[ii][jj]=t1;}
for (int jj=0;jj<GL;jj++){b =fscanf(file,"%f,",&t1); MI->y2[ii][jj]=t1;}
}
}
MI->padx=(int)ceil((double)MI->max_X/2.0+1.0); //padx
MI->pady=(int)ceil((double)MI->max_Y/2.0+1.0); //padY
MI->ini=true;
//fclose
fclose(file);
return(MI);
}
Partfilters *load_partfilter(char *filename)
{
FILE *file; //File
errno_t err; //err ( for fopen)
CUresult res;
Partfilters *PF=(Partfilters*)malloc(sizeof(Partfilters)); //Part filter
//fopen
//if( (err = fopen_s( &file,filename, "r"))!=0 )
if( (file=fopen(filename, "r"))==NULL )
{
printf("Part-filter file not found \n");
exit(-1);
}
FLOAT t1,t2,t3;
FLOAT dummy_t1, dummy_t2, dummy_t3; // variable for dummy scan in order to adjust the location of file-pointer
int b;
if(sizeof(FLOAT)==sizeof(double)) {
b =fscanf(file,"%lf,",&t1);
}else{
b =fscanf(file,"%f,",&t1);
}
PF->NoP=(int)t1; //number of part filter
PF->part_size=(int**)malloc(sizeof(int*)*PF->NoP); //size of part filter
PF->partfilter=(FLOAT**)malloc(sizeof(FLOAT*)*PF->NoP); //weight of part filter
PF->part_partner=(int*)malloc(sizeof(int)*PF->NoP); //symmetric information of part
PF->part_sym=(int*)malloc(sizeof(int)*PF->NoP); //symmetric information of part
/* keep file pointer location */
long before_loop_location = ftell(file);
int SUM_SIZE_PART = 0;
for (int ii=0;ii<PF->NoP;ii++)
{
if(sizeof(FLOAT)==sizeof(double)) {
b =fscanf(file,"%lf,%lf,%lf,",&t1,&t2,&t3); //number of components
}else{
b =fscanf(file,"%f,%f,%f,",&t1,&t2,&t3); //number of components
}
PF->part_size[ii]=(int*)malloc(sizeof(int)*3);
PF->part_size[ii][0]=(int)t1;
PF->part_size[ii][1]=(int)t2;
PF->part_size[ii][2]=(int)t3;
//printf("***************%f %f %f\n",t1,t2,t3);
int NUMB=PF->part_size[ii][0]*PF->part_size[ii][1]*PF->part_size[ii][2];
#ifdef ORIGINAL
PF->partfilter[ii]=(FLOAT*)malloc(sizeof(FLOAT)*NUMB); //weight of root filter
#else
#ifdef SEPARETE_MEM
res = cuMemHostAlloc((void **)&(PF->partfilter[ii]), sizeof(FLOAT)*NUMB, CU_MEMHOSTALLOC_DEVICEMAP);
if(res != CUDA_SUCCESS){
printf("cuMemHostAlloc(PF->partfilter) failed: res = %s\n", conv(res));
exit(1);
}
#else
SUM_SIZE_PART += NUMB*sizeof(FLOAT);
#endif
#endif
/* adjust the location of file-pointer */
for(int jj=0; jj<NUMB; jj++) {
if(sizeof(FLOAT)==sizeof(double)) {
fscanf(file,"%lf,",&dummy_t1); // this is dummy scan
}else{
fscanf(file,"%f,",&dummy_t1); // this is dummy scan
}
}
if(sizeof(FLOAT)==sizeof(double)) {
fscanf(file,"%lf,",&dummy_t1); // this is dummy scan
}else{
fscanf(file,"%f,",&dummy_t1); // this is dummy scan
}
}
#ifndef ORIGINAL
#ifndef SEPARETE_MEM
/* allocate memory region for part in a lump */
FLOAT *dst_part;
res = cuMemHostAlloc((void **)&dst_part, SUM_SIZE_PART, CU_MEMHOSTALLOC_DEVICEMAP);
if(res != CUDA_SUCCESS){
printf("cuMemHostAlloc(dst_part) failed: res = %s\n", conv(res));
exit(1);
}
/* distribution */
unsigned long long int pointer = (unsigned long long int)dst_part;
for(int ii=0; ii<PF->NoP; ii++) {
PF->partfilter[ii] = (FLOAT *)pointer;
int NUMB=PF->part_size[ii][0]*PF->part_size[ii][1]*PF->part_size[ii][2];
pointer += NUMB*sizeof(FLOAT);
}
#endif
#endif
/* reset the location of file-pointer */
fseek(file, before_loop_location, SEEK_SET);
for(int ii=0; ii<PF->NoP; ii++) {
int NUMB=PF->part_size[ii][0]*PF->part_size[ii][1]*PF->part_size[ii][2];
/* adjust the location of file-pointer */
if(sizeof(FLOAT)==sizeof(double)) {
fscanf(file,"%lf,%lf,%lf,",&dummy_t1,&dummy_t2,&dummy_t3); // this is dummy scan
}else{
fscanf(file,"%f,%f,%f,",&dummy_t1,&dummy_t2,&dummy_t3); // this is dummy scan
}
for (int jj=0;jj<NUMB;jj++)
{
if(sizeof(FLOAT)==sizeof(double)) {
b =fscanf(file,"%lf,",&t1);
}else{
b =fscanf(file,"%f,",&t1);
}
PF->partfilter[ii][jj]=t1;
}
if(sizeof(FLOAT)==sizeof(double)) {
b =fscanf(file,"%lf,",&t1);
}else{
b =fscanf(file,"%f,",&t1);
}
PF->part_partner[ii]=(int)t1; //symmetric information of part
if(PF->part_partner[ii]==0) PF->part_sym[ii]=1;
else PF->part_sym[ii]=0;
}
//fclose
fclose(file);
return(PF);
}