int main (int argc, char *argv[]) {
MKLVersion ver;
// Print information on CPU optimization in effect
MKLGetVersion(&ver);
printf("Processor optimization: %s\n",ver.Processor);
omp_set_num_threads(8);
mkl_set_dynamic(true);
unsigned short A = ((unsigned short) std::strtoul(argv[1],(char **)0, 10));
unsigned short K_min = ((unsigned short) std::strtoul(argv[2],(char **)0, 10));
unsigned short K_max = ((unsigned short) std::strtoul(argv[3],(char **)0, 10));
unsigned short P = ((unsigned short) std::strtoul(argv[4],(char **)0, 10));
unsigned short Parita_Orbitale = ((unsigned short) std::strtoul(argv[5],(char **)0, 10));; //0 = pari ; 1 = dispari ; 2 = tutte
unsigned short modo_cluster = 0;
if (A==4) {
if (argc == 7) {
modo_cluster = ((unsigned short) std::strtoul(argv[6],(char **)0, 10)); //mettere a 1 per avere modo H sui 4 corpi
}
}
init_main_0(A,K_min,K_max,P,modo_cluster,Parita_Orbitale);
return true;
}
double* ttm (double *X, const int Dim, long int *XDimSize,
double *U, long int *UDimSize, char tflag,
double *Y, long int *YDimSize,
int ModeComn, int *ModeCom, int ModeComLen)
{
//Decide transa and transb according to tflag and ModeComn and ModeCom's relationship.
CBLAS_TRANSPOSE transa = CblasNoTrans;
CBLAS_TRANSPOSE transb = CblasNoTrans;
long int m,n,k;
double alpha, beta;
long int rsu,csu,rsx,csx,rsy,csy;
long int XLoopStride = 0, YLoopStride = 0;
alpha = 1.0;
beta = 1.0;
/****Generated TM algo according to different modes and dimensions ***/
m = 10;
n = 1000;
k = 1000;
rsu = 1000;
csu = 1;
rsx = 1000;
csx = 1;
rsy = 1000;
csy = 1;
#pragma omp parallel for default(shared) schedule(static) num_threads(1)
for (long int i0=0;i0<1000;i0++)
{
mkl_set_dynamic(0);
mkl_set_num_threads_local(8);
XLoopStride = i0*1000000;
YLoopStride = i0*10000;
cblas_dgemm(CblasRowMajor, transa, transb,
m, n, k,
alpha,
U, rsu,
X+XLoopStride, rsx,
beta,
Y+YLoopStride, rsy);
}
return Y;
}
int main (int argc, char *argv[]) {
if (argc == 1){
std::cerr << "Sintassi: A m_max K_min K_max nucleoni S_tot masse_uguali [modo_H]" << std::endl;
return true;
}
MKLVersion ver;
// Print information on CPU optimization in effect
MKLGetVersion(&ver);
printf("Processor optimization: %s\n",ver.Processor);
omp_set_num_threads(8);
mkl_set_dynamic(true);
unsigned short A = ((unsigned short) std::strtoul(argv[1],(char **)0, 10));
unsigned short m_max = ((unsigned short) std::strtoul(argv[2],(char **)0, 10));
unsigned short K_min = ((unsigned short) std::strtoul(argv[3],(char **)0, 10));
unsigned short K_max = ((unsigned short) std::strtoul(argv[4],(char **)0, 10));
std::vector <char> nucleoni;
for (short i = 5; i<5+A; ++i) {
nucleoni.push_back(*argv[i]);
}
unsigned short S_tot = ((unsigned short) std::strtoul(argv[5+A],(char **)0, 10));
bool controllo = ((bool) std::strtoul(argv[6+A],(char **)0, 10));
bool modo_cluster = false;
if (A==4) {
beta_riferimento = 2.;
if (argc == 8+A) {
modo_cluster = ((bool) std::strtoul(argv[7+A],(char **)0, 10)); //mettere a 1 per avere modo H sui 4 corpi
}
}
Init_Corpi_LS Corp(nucleoni,controllo);
std::string Nome_dir_vec = "lista_vettori";
int crea_dir = mkdir(Nome_dir_vec.c_str(), 0777);
init_main_0(A,m_max,K_min,K_max,S_tot,Corp,modo_cluster);
return true;
}
void matmul_init() {
#ifdef WITH_MKL
mkl_set_dynamic(1);
#endif
// omp_set_dynamic(1);
}
void StVKReducedStiffnessMatrix::Evaluate(double * q, double * Rq)
{
// this is same as EvaluateSubset with start=0, end=quadraticSize
/*
int i,j,k;
int output;
// reset to free terms
int index = 0;
int indexEntry = 0;
for(output=0; output<r; output++)
{
for(i=output; i<r; i++)
{
Rq[indexEntry] = freeCoef_[index];
index++;
indexEntry++;
}
indexEntry += output + 1;
}
// add linear terms
index = 0;
indexEntry = 0;
for(output=0; output<r; output++)
{
for(i=output; i<r; i++)
{
for(j=0; j<r; j++)
{
Rq[indexEntry] += linearCoef_[index] * q[j];
index++;
}
indexEntry++;
}
indexEntry += output + 1;
}
// add quadratic terms
index = 0;
indexEntry = 0;
for(output=0; output<r; output++)
{
for(i=output; i<r; i++)
{
for(j=0; j<r; j++)
for(k=j; k<r; k++)
{
Rq[indexEntry] += quadraticCoef_[index] * q[j] * q[k];
index++;
}
indexEntry++;
}
indexEntry += output + 1;
}
// make symetric
for(output=0; output<r; output++)
for(i=0; i<output; i++)
Rq[ELT(r,i,output)] = Rq[ELT(r,output,i)];
*/
if (useSingleThread)
{
#if defined(WIN32) || defined(linux)
mkl_max_threads = mkl_get_max_threads();
mkl_dynamic = mkl_get_dynamic();
mkl_set_num_threads(1);
mkl_set_dynamic(0);
#elif defined(__APPLE__)
//setenv("VECLIB_MAXIMUM_THREADS", "1", true);
#endif
}
// reset to free terms
memcpy(buffer1,freeCoef_,sizeof(double)*quadraticSize);
// add linear terms
// multiply linearCoef_ and q
// linearCoef_ is r x quadraticSize array
cblas_dgemv(CblasColMajor, CblasTrans,
r, quadraticSize,
1.0,
linearCoef_, r,
q, 1,
1.0,
buffer1, 1);
// compute qiqj
int index = 0;
for(int output=0; output<r; output++)
for(int i=output; i<r; i++)
{
qiqj[index] = q[output] * q[i];
index++;
}
// update Rq
// quadraticCoef_ is quadraticSize x quadraticSize matrix
// each column gives quadratic coef for one matrix entry
cblas_dgemv(CblasColMajor, CblasTrans,
quadraticSize, quadraticSize,
1.0,
quadraticCoef_, quadraticSize,
qiqj, 1,
1.0,
buffer1, 1);
// unpack into a symmetric matrix
int i1=0,j1=0;
for(int i=0; i< quadraticSize; i++)
{
Rq[ELT(r,i1,j1)] = buffer1[i];
Rq[ELT(r,j1,i1)] = buffer1[i];
j1++;
if(j1 == r)
{
i1++;
j1 = i1;
}
}
if (useSingleThread)
{
#if defined(WIN32) || defined(linux)
mkl_set_num_threads(mkl_max_threads);
mkl_set_dynamic(mkl_dynamic);
#elif defined(__APPLE__)
//unsetenv("VECLIB_MAXIMUM_THREADS");
#endif
}
}
void StVKReducedInternalForces::Evaluate(double * q, double * fq)
{
/* // unoptimized version
// reset to zero
int i,j,k,l;
for(l=0; l<r; l++)
fq[l] = 0;
// add linear terms
int index = 0;
for(l=0; l<r; l++)
for(i=0; i<r; i++)
{
fq[l] += linearCoef_[index] * q[i];
index++;
}
// add quadratic terms
index = 0;
for(l=0; l<r; l++)
for(i=0; i<r; i++)
for(j=i; j<r; j++)
{
fq[l] += quadraticCoef_[index] * q[i] * q[j];
index++;
}
// add cubic terms
index = 0;
for(l=0; l<r; l++)
for(i=0; i<r; i++)
for(j=i; j<r; j++)
for(k=j; k<r; k++)
{
fq[l] += cubicCoef_[index] * q[i] * q[j] * q[k];
index++;
}
*/
if (useSingleThread)
{
#if defined(_WIN32) || defined(WIN32) || defined(linux)
mkl_max_threads = mkl_get_max_threads();
mkl_dynamic = mkl_get_dynamic();
mkl_set_num_threads(1);
mkl_set_dynamic(0);
#elif defined(__APPLE__)
//setenv("VECLIB_MAXIMUM_THREADS", "1", true);
#endif
}
// add linear terms
// multiply linearCoef_ and q
// linearCoef_ is r x r array
cblas_dgemv(CblasColMajor, CblasTrans,
r, r,
1.0,
linearCoef_, r,
q, 1,
0.0,
fq, 1);
// compute qiqj
int index = 0;
for(int output=0; output<r; output++)
for(int i=output; i<r; i++)
{
qiqj[index] = q[output] * q[i];
index++;
}
// add quadratic terms
// quadraticCoef_ is quadraticSize x r matrix
// each column gives quadratic coef for one force vector component
cblas_dgemv(CblasColMajor, CblasTrans,
quadraticSize, r,
1.0,
quadraticCoef_, quadraticSize,
qiqj, 1,
1.0,
fq, 1);
// add cubic terms
// cubicCoef_ is cubicSize x r matrix
// each column gives cubicSize coef for one force vector component
int size = quadraticSize;
double * qiqjPos = qiqj;
double * cubicCoefPos = cubicCoef_;
for(int i=0; i<r; i++)
{
cblas_dgemv(CblasColMajor, CblasTrans,
size, r,
q[i],
cubicCoefPos, cubicSize,
qiqjPos, 1,
1.0,
fq, 1);
int param = r-i;
size -= param;
qiqjPos += param;
cubicCoefPos += param * (param+1) / 2;
}
if (addGravity)
{
for(int i=0; i<r; i++)
fq[i] -= reducedGravityForce[i];
}
if (useSingleThread)
{
#if defined(_WIN32) || defined(WIN32) || defined(linux)
mkl_set_num_threads(mkl_max_threads);
mkl_set_dynamic(mkl_dynamic);
#elif defined(__APPLE__)
//unsetenv("VECLIB_MAXIMUM_THREADS");
#endif
}
}
double FastMatmul(Matrix<Scalar>& A, Matrix<Scalar>& B, Matrix<Scalar>& C,
int num_steps, double x=1e-8, Scalar alpha=Scalar(1.0), Scalar beta=Scalar(0.0)) {
MemoryManager<Scalar> mem_mngr;
#ifdef _PARALLEL_
mem_mngr.Allocate(2, 2, 2, 8, num_steps, A.m(), A.n(), B.n());
#endif
A.set_multiplier(alpha);
int num_multiplies_per_step = 8;
int total_multiplies = pow(num_multiplies_per_step, num_steps);
// Set parameters needed for all types of parallelism.
int num_threads = 0;
#ifdef _PARALLEL_
# pragma omp parallel
{
if (omp_get_thread_num() == 0) {
num_threads = omp_get_num_threads();
}
}
omp_set_nested(1);
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_)
# pragma omp parallel
{
mkl_set_num_threads_local(1);
mkl_set_dynamic(0);
}
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _DFS_PAR_)
mkl_set_dynamic(0);
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _HYBRID_PAR_)
if (num_threads > total_multiplies) {
mkl_set_dynamic(0);
} else {
# pragma omp parallel
{
mkl_set_num_threads_local(1);
mkl_set_dynamic(0);
}
}
#endif
LockAndCounter locker(total_multiplies - (total_multiplies % num_threads));
using FpMilliseconds = std::chrono::duration<float, std::chrono::milliseconds::period>;
auto t1 = std::chrono::high_resolution_clock::now();
#ifdef _PARALLEL_
# pragma omp parallel
{
# pragma omp single
#endif
FastMatmulRecursive(locker, mem_mngr, A, B, C, num_steps, num_steps, 0, x, num_threads, beta);
#ifdef _PARALLEL_
}
#endif
auto t2 = std::chrono::high_resolution_clock::now();
return FpMilliseconds(t2 - t1).count();
}
void FastMatmulRecursive(LockAndCounter& locker, MemoryManager<Scalar>& mem_mngr, Matrix<Scalar>& A, Matrix<Scalar>& B, Matrix<Scalar>& C, int total_steps, int steps_left, int start_index, double x, int num_threads, Scalar beta) {
// Update multipliers
C.UpdateMultiplier(A.multiplier());
C.UpdateMultiplier(B.multiplier());
A.set_multiplier(Scalar(1.0));
B.set_multiplier(Scalar(1.0));
// Base case for recursion
if (steps_left == 0) {
MatMul(A, B, C);
return;
}
Matrix<Scalar> A11 = A.Subblock(2, 2, 1, 1);
Matrix<Scalar> A12 = A.Subblock(2, 2, 1, 2);
Matrix<Scalar> A21 = A.Subblock(2, 2, 2, 1);
Matrix<Scalar> A22 = A.Subblock(2, 2, 2, 2);
Matrix<Scalar> B11 = B.Subblock(2, 2, 1, 1);
Matrix<Scalar> B12 = B.Subblock(2, 2, 1, 2);
Matrix<Scalar> B21 = B.Subblock(2, 2, 2, 1);
Matrix<Scalar> B22 = B.Subblock(2, 2, 2, 2);
Matrix<Scalar> C11 = C.Subblock(2, 2, 1, 1);
Matrix<Scalar> C12 = C.Subblock(2, 2, 1, 2);
Matrix<Scalar> C21 = C.Subblock(2, 2, 2, 1);
Matrix<Scalar> C22 = C.Subblock(2, 2, 2, 2);
// Matrices to store the results of multiplications.
#ifdef _PARALLEL_
Matrix<Scalar> M1(mem_mngr.GetMem(start_index, 1, total_steps - steps_left, M), C11.m(), C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M2(mem_mngr.GetMem(start_index, 2, total_steps - steps_left, M), C11.m(), C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M3(mem_mngr.GetMem(start_index, 3, total_steps - steps_left, M), C11.m(), C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M4(mem_mngr.GetMem(start_index, 4, total_steps - steps_left, M), C11.m(), C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M5(mem_mngr.GetMem(start_index, 5, total_steps - steps_left, M), C11.m(), C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M6(mem_mngr.GetMem(start_index, 6, total_steps - steps_left, M), C11.m(), C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M7(mem_mngr.GetMem(start_index, 7, total_steps - steps_left, M), C11.m(), C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M8(mem_mngr.GetMem(start_index, 8, total_steps - steps_left, M), C11.m(), C11.m(), C11.n(), C.multiplier());
#else
Matrix<Scalar> M1(C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M2(C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M3(C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M4(C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M5(C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M6(C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M7(C11.m(), C11.n(), C.multiplier());
Matrix<Scalar> M8(C11.m(), C11.n(), C.multiplier());
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
bool sequential1 = should_launch_task(8, total_steps, steps_left, start_index, 1, num_threads);
bool sequential2 = should_launch_task(8, total_steps, steps_left, start_index, 2, num_threads);
bool sequential3 = should_launch_task(8, total_steps, steps_left, start_index, 3, num_threads);
bool sequential4 = should_launch_task(8, total_steps, steps_left, start_index, 4, num_threads);
bool sequential5 = should_launch_task(8, total_steps, steps_left, start_index, 5, num_threads);
bool sequential6 = should_launch_task(8, total_steps, steps_left, start_index, 6, num_threads);
bool sequential7 = should_launch_task(8, total_steps, steps_left, start_index, 7, num_threads);
bool sequential8 = should_launch_task(8, total_steps, steps_left, start_index, 8, num_threads);
#else
bool sequential1 = false;
bool sequential2 = false;
bool sequential3 = false;
bool sequential4 = false;
bool sequential5 = false;
bool sequential6 = false;
bool sequential7 = false;
bool sequential8 = false;
#endif
// M1 = (1 * A11) * (1 * B11)
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
# pragma omp task if(sequential1) shared(mem_mngr, locker) untied
{
#endif
M1.UpdateMultiplier(Scalar(1));
M1.UpdateMultiplier(Scalar(1));
FastMatmulRecursive(locker, mem_mngr, A11, B11, M1, total_steps, steps_left - 1, (start_index + 1 - 1) * 8, x, num_threads, Scalar(0.0));
#ifndef _PARALLEL_
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
locker.Decrement();
}
if (should_task_wait(8, total_steps, steps_left, start_index, 1, num_threads)) {
# pragma omp taskwait
# if defined(_PARALLEL_) && (_PARALLEL_ == _HYBRID_PAR_)
SwitchToDFS(locker, num_threads);
# endif
}
#endif
// M2 = (1 * A12) * (1 * B21)
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
# pragma omp task if(sequential2) shared(mem_mngr, locker) untied
{
#endif
M2.UpdateMultiplier(Scalar(1));
M2.UpdateMultiplier(Scalar(1));
FastMatmulRecursive(locker, mem_mngr, A12, B21, M2, total_steps, steps_left - 1, (start_index + 2 - 1) * 8, x, num_threads, Scalar(0.0));
#ifndef _PARALLEL_
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
locker.Decrement();
}
if (should_task_wait(8, total_steps, steps_left, start_index, 2, num_threads)) {
# pragma omp taskwait
# if defined(_PARALLEL_) && (_PARALLEL_ == _HYBRID_PAR_)
SwitchToDFS(locker, num_threads);
# endif
}
#endif
// M3 = (1 * A11) * (1 * B12)
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
# pragma omp task if(sequential3) shared(mem_mngr, locker) untied
{
#endif
M3.UpdateMultiplier(Scalar(1));
M3.UpdateMultiplier(Scalar(1));
FastMatmulRecursive(locker, mem_mngr, A11, B12, M3, total_steps, steps_left - 1, (start_index + 3 - 1) * 8, x, num_threads, Scalar(0.0));
#ifndef _PARALLEL_
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
locker.Decrement();
}
if (should_task_wait(8, total_steps, steps_left, start_index, 3, num_threads)) {
# pragma omp taskwait
# if defined(_PARALLEL_) && (_PARALLEL_ == _HYBRID_PAR_)
SwitchToDFS(locker, num_threads);
# endif
}
#endif
// M4 = (1 * A12) * (1 * B22)
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
# pragma omp task if(sequential4) shared(mem_mngr, locker) untied
{
#endif
M4.UpdateMultiplier(Scalar(1));
M4.UpdateMultiplier(Scalar(1));
FastMatmulRecursive(locker, mem_mngr, A12, B22, M4, total_steps, steps_left - 1, (start_index + 4 - 1) * 8, x, num_threads, Scalar(0.0));
#ifndef _PARALLEL_
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
locker.Decrement();
}
if (should_task_wait(8, total_steps, steps_left, start_index, 4, num_threads)) {
# pragma omp taskwait
# if defined(_PARALLEL_) && (_PARALLEL_ == _HYBRID_PAR_)
SwitchToDFS(locker, num_threads);
# endif
}
#endif
// M5 = (1 * A21) * (1 * B11)
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
# pragma omp task if(sequential5) shared(mem_mngr, locker) untied
{
#endif
M5.UpdateMultiplier(Scalar(1));
M5.UpdateMultiplier(Scalar(1));
FastMatmulRecursive(locker, mem_mngr, A21, B11, M5, total_steps, steps_left - 1, (start_index + 5 - 1) * 8, x, num_threads, Scalar(0.0));
#ifndef _PARALLEL_
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
locker.Decrement();
}
if (should_task_wait(8, total_steps, steps_left, start_index, 5, num_threads)) {
# pragma omp taskwait
# if defined(_PARALLEL_) && (_PARALLEL_ == _HYBRID_PAR_)
SwitchToDFS(locker, num_threads);
# endif
}
#endif
// M6 = (1 * A22) * (1 * B21)
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
# pragma omp task if(sequential6) shared(mem_mngr, locker) untied
{
#endif
M6.UpdateMultiplier(Scalar(1));
M6.UpdateMultiplier(Scalar(1));
FastMatmulRecursive(locker, mem_mngr, A22, B21, M6, total_steps, steps_left - 1, (start_index + 6 - 1) * 8, x, num_threads, Scalar(0.0));
#ifndef _PARALLEL_
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
locker.Decrement();
}
if (should_task_wait(8, total_steps, steps_left, start_index, 6, num_threads)) {
# pragma omp taskwait
# if defined(_PARALLEL_) && (_PARALLEL_ == _HYBRID_PAR_)
SwitchToDFS(locker, num_threads);
# endif
}
#endif
// M7 = (1 * A21) * (1 * B12)
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
# pragma omp task if(sequential7) shared(mem_mngr, locker) untied
{
#endif
M7.UpdateMultiplier(Scalar(1));
M7.UpdateMultiplier(Scalar(1));
FastMatmulRecursive(locker, mem_mngr, A21, B12, M7, total_steps, steps_left - 1, (start_index + 7 - 1) * 8, x, num_threads, Scalar(0.0));
#ifndef _PARALLEL_
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
locker.Decrement();
}
if (should_task_wait(8, total_steps, steps_left, start_index, 7, num_threads)) {
# pragma omp taskwait
# if defined(_PARALLEL_) && (_PARALLEL_ == _HYBRID_PAR_)
SwitchToDFS(locker, num_threads);
# endif
}
#endif
// M8 = (1 * A22) * (1 * B22)
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
# pragma omp task if(sequential8) shared(mem_mngr, locker) untied
{
#endif
M8.UpdateMultiplier(Scalar(1));
M8.UpdateMultiplier(Scalar(1));
FastMatmulRecursive(locker, mem_mngr, A22, B22, M8, total_steps, steps_left - 1, (start_index + 8 - 1) * 8, x, num_threads, Scalar(0.0));
#ifndef _PARALLEL_
#endif
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
locker.Decrement();
}
if (should_task_wait(8, total_steps, steps_left, start_index, 8, num_threads)) {
# pragma omp taskwait
}
#endif
M_Add1(M1, M2, C11, x, false, beta);
M_Add2(M3, M4, C12, x, false, beta);
M_Add3(M5, M6, C21, x, false, beta);
M_Add4(M7, M8, C22, x, false, beta);
// Handle edge cases with dynamic peeling
#if defined(_PARALLEL_) && (_PARALLEL_ == _BFS_PAR_ || _PARALLEL_ == _HYBRID_PAR_)
if (total_steps == steps_left) {
mkl_set_num_threads_local(num_threads);
mkl_set_dynamic(0);
}
#endif
DynamicPeeling(A, B, C, 2, 2, 2, beta);
}