double CNAME(blasint n, float *x, blasint incx, float *y, blasint incy){
double ret = 0.0;
PRINT_DEBUG_CNAME;
if (n <= 0) return 0;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (n - 1) * incx;
if (incy < 0) y -= (n - 1) * incy;
ret=DSDOT_K(n, x, incx, y, incy);
FUNCTION_PROFILE_END(1, n, n);
IDEBUG_END;
return ret;
}
void NAME(char *TRANS, blasint *M, blasint *N,
FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX,
FLOAT *BETA, FLOAT *y, blasint *INCY){
char trans = *TRANS;
blasint m = *M;
blasint n = *N;
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
FLOAT alpha = *ALPHA;
FLOAT beta = *BETA;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T,
};
blasint info;
blasint lenx, leny;
blasint i;
PRINT_DEBUG_NAME;
TOUPPER(trans);
info = 0;
i = -1;
if (trans == 'N') i = 0;
if (trans == 'T') i = 1;
if (trans == 'R') i = 0;
if (trans == 'C') i = 1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (i < 0) info = 1;
trans = i;
if (info != 0){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_TRANSPOSE TransA,
blasint m, blasint n,
FLOAT alpha,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT beta,
FLOAT *y, blasint incy){
FLOAT *buffer;
blasint lenx, leny;
int trans;
blasint info, t;
#ifdef SMP
int nthreads;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T,
};
PRINT_DEBUG_CNAME;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 0;
if (TransA == CblasConjTrans) trans = 1;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (order == CblasRowMajor) {
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 1;
if (TransA == CblasConjTrans) trans = 0;
info = -1;
t = n;
n = m;
m = t;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if ((m==0) || (n==0)) return;
lenx = n;
leny = m;
if (trans) lenx = m;
if (trans) leny = n;
if (beta != ONE) SCAL_K(leny, 0, 0, beta, y, abs(incy), NULL, 0, NULL, 0);
if (alpha == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (lenx - 1) * incx;
if (incy < 0) y -= (leny - 1) * incy;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
int nthreads_max = num_cpu_avail(2);
int nthreads_avail = nthreads_max;
double MNK = (double) m * (double) n;
if ( MNK <= (24.0 * 24.0 * (double) (GEMM_MULTITHREAD_THRESHOLD*GEMM_MULTITHREAD_THRESHOLD) ) )
nthreads_max = 1;
if ( nthreads_max > nthreads_avail )
nthreads = nthreads_avail;
else
nthreads = nthreads_max;
if (nthreads == 1) {
#endif
(gemv[(int)trans])(m, n, 0, alpha, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(gemv_thread[(int)trans])(m, n, alpha, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA,
FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY, FLOAT *a, blasint *LDA){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (lda < MAX(1, n)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT *ALPHA, FLOAT *x, blasint incx, FLOAT *y, blasint incy, FLOAT *a, blasint lda) {
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
FLOAT *buffer;
int uplo;
blasint info;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
if (lda < MAX(1, n)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 3;
if (Uplo == CblasLower) uplo = 2;
info = -1;
if (lda < MAX(1, n)) info = 9;
if (incx == 0) info = 7;
if (incy == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx * 2;
if (incy < 0 ) y -= (n - 1) * incy * 2;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(her2[uplo])(n, alpha_r, alpha_i, x, incx, y, incy, a, lda, buffer);
#ifdef SMP
} else {
(her2_thread[uplo])(n, ALPHA, x, incx, y, incy, a, lda, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, n * n / 2 + 2 * n, 2 * n * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, blasint *N, blasint *K, FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX, FLOAT *BETA, FLOAT *y, blasint *INCY){
char uplo_arg = *UPLO;
blasint n = *N;
blasint k = *K;
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
blasint lda = *LDA;
blasint incx = *INCX;
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
blasint incy = *INCY;
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMPBUG
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
if (uplo_arg == 'V') uplo = 2;
if (uplo_arg == 'M') uplo = 3;
info = 0;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < k + 1) info = 6;
if (k < 0) info = 3;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_UPLO Uplo,
blasint n, blasint k,
FLOAT *ALPHA,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT *BETA,
FLOAT *y, blasint incy){
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
FLOAT *buffer;
int uplo;
blasint info;
#ifdef SMPBUG
int nthreads;
#endif
PRINT_DEBUG_CNAME;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < k + 1) info = 6;
if (k < 0) info = 3;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 3;
if (Uplo == CblasLower) uplo = 2;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < k + 1) info = 6;
if (k < 0) info = 3;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
if ((beta_r != ONE) || (beta_i != ZERO)) SCAL_K(n, 0, 0, beta_r, beta_i, y, abs(incy), NULL, 0, NULL, 0);
if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx * COMPSIZE;
if (incy < 0 ) y -= (n - 1) * incy * COMPSIZE;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMPBUG
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(hbmv[uplo])(n, k, alpha_r, alpha_i, a, lda, x, incx, y, incy, buffer);
#ifdef SMPBUG
} else {
(hbmv_thread[uplo])(n, k, ALPHA, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, n * k / 2 + n, n * k);
IDEBUG_END;
return;
}
void NAME(char *UPLO, char *TRANS, char *DIAG,
blasint *N, blasint *K,
FLOAT *a, blasint *LDA, FLOAT *x, blasint *INCX){
char uplo_arg = *UPLO;
char trans_arg = *TRANS;
char diag_arg = *DIAG;
blasint n = *N;
blasint k = *K;
blasint lda = *LDA;
blasint incx = *INCX;
blasint info;
int uplo;
int unit;
int trans;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
TOUPPER(trans_arg);
TOUPPER(diag_arg);
trans = -1;
unit = -1;
uplo = -1;
if (trans_arg == 'N') trans = 0;
if (trans_arg == 'T') trans = 1;
if (trans_arg == 'R') trans = 0;
if (trans_arg == 'C') trans = 1;
if (diag_arg == 'U') unit = 0;
if (diag_arg == 'N') unit = 1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 9;
if (lda < k + 1) info = 7;
if (k < 0) info = 5;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
blasint n, blasint k, FLOAT *a, blasint lda, FLOAT *x, blasint incx) {
int trans, uplo, unit;
blasint info;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
unit = -1;
uplo = -1;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 0;
if (TransA == CblasConjTrans) trans = 1;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 9;
if (lda < k + 1) info = 7;
if (k < 0) info = 5;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 1;
if (TransA == CblasConjTrans) trans = 0;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 9;
if (lda < k + 1) info = 7;
if (k < 0) info = 5;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(tbmv[(trans<<2) | (uplo<<1) | unit])(n, k, a, lda, x, incx, buffer);
#ifdef SMP
} else {
(tbmv_thread[(trans<<2) | (uplo<<1) | unit])(n, k, a, lda, x, incx, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, n * k / 2 + n, n * k);
IDEBUG_END;
return;
}
void NAME(blasint *N, FLOAT *ALPHA, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY){
BLASLONG n = *N;
BLASLONG incx = *INCX;
BLASLONG incy = *INCY;
FLOAT alpha = *ALPHA;
#else
void CNAME(blasint n, FLOAT alpha, FLOAT *x, blasint incx, FLOAT *y, blasint incy){
#endif
#ifdef SMP
int mode, nthreads;
#endif
#ifndef CBLAS
PRINT_DEBUG_NAME;
#else
PRINT_DEBUG_CNAME;
#endif
if (n <= 0) return;
if (alpha == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (n - 1) * incx;
if (incy < 0) y -= (n - 1) * incy;
#ifdef SMP
nthreads = num_cpu_avail(1);
//disable multi-thread when incx==0 or incy==0
//In that case, the threads would be dependent.
if (incx == 0 || incy == 0)
nthreads = 1;
//Temporarily work-around the low performance issue with small imput size &
//multithreads.
if (n <= 10000)
nthreads = 1;
if (nthreads == 1) {
#endif
AXPYU_K(n, 0, 0, alpha, x, incx, y, incy, NULL, 0);
#ifdef SMP
} else {
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_REAL;
#else
mode = BLAS_SINGLE | BLAS_REAL;
#endif
blas_level1_thread(mode, n, 0, 0, &alpha,
x, incx, y, incy, NULL, 0, (void *)AXPYU_K, nthreads);
}
#endif
FUNCTION_PROFILE_END(1, 2 * n, 2 * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX, FLOAT *BETA, FLOAT *y, blasint *INCY){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha = *ALPHA;
blasint lda = *LDA;
blasint incx = *INCX;
FLOAT beta = *BETA;
blasint incy = *INCY;
int (*symv[])(BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
SYMV_U, SYMV_L,
};
#ifdef SMP
int (*symv_thread[])(BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, int) = {
SYMV_THREAD_U, SYMV_THREAD_L,
};
#endif
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incy == 0) info = 10;
if (incx == 0) info = 7;
if (lda < MAX(1, n)) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha,
FLOAT *a, blasint lda, FLOAT *x, blasint incx, FLOAT beta, FLOAT *y, blasint incy) {
FLOAT *buffer;
int uplo;
blasint info;
#ifdef SMP
int nthreads;
#endif
int (*symv[])(BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
SYMV_U, SYMV_L,
};
#ifdef SMP
int (*symv_thread[])(BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, int) = {
SYMV_THREAD_U, SYMV_THREAD_L,
};
#endif
PRINT_DEBUG_CNAME;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
if (incy == 0) info = 10;
if (incx == 0) info = 7;
if (lda < MAX(1, n)) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
info = -1;
if (incy == 0) info = 10;
if (incx == 0) info = 7;
if (lda < MAX(1, n)) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
if (beta != ONE) SCAL_K(n, 0, 0, beta, y, abs(incy), NULL, 0, NULL, 0);
if (alpha == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
if (incy < 0 ) y -= (n - 1) * incy;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(symv[uplo])(n, n, alpha, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(symv_thread[uplo])(n, alpha, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, n * n / 2 + 2 * n, 2 * n * n);
IDEBUG_END;
return;
}
void NAME(char *TRANS, blasint *M, blasint *N,
blasint *KU, blasint *KL,
FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX,
FLOAT *BETA, FLOAT *y, blasint *INCY){
char trans = *TRANS;
blasint m = *M;
blasint n = *N;
blasint ku = *KU;
blasint kl = *KL;
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
blasint info;
blasint lenx, leny;
blasint i;
PRINT_DEBUG_NAME;
TOUPPER(trans);
info = 0;
i = -1;
if (trans == 'N') i = 0;
if (trans == 'T') i = 1;
if (trans == 'R') i = 2;
if (trans == 'C') i = 3;
if (trans == 'O') i = 4;
if (trans == 'U') i = 5;
if (trans == 'S') i = 6;
if (trans == 'D') i = 7;
if (incy == 0) info = 13;
if (incx == 0) info = 10;
if (lda < kl + ku + 1) info = 8;
if (kl < 0) info = 5;
if (ku < 0) info = 4;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (i < 0) info = 1;
trans = i;
if (info != 0){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_TRANSPOSE TransA,
blasint m, blasint n,
blasint ku, blasint kl,
FLOAT *ALPHA,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT *BETA,
FLOAT *y, blasint incy){
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
FLOAT *buffer;
blasint lenx, leny;
int trans;
blasint info, t;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 2;
if (TransA == CblasConjTrans) trans = 3;
info = -1;
if (incy == 0) info = 13;
if (incx == 0) info = 10;
if (lda < kl + ku + 1) info = 8;
if (kl < 0) info = 5;
if (ku < 0) info = 4;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (order == CblasRowMajor) {
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 3;
if (TransA == CblasConjTrans) trans = 2;
info = -1;
t = n;
n = m;
m = t;
t = ku;
ku = kl;
kl = t;
if (incy == 0) info = 13;
if (incx == 0) info = 10;
if (lda < kl + ku + 1) info = 8;
if (kl < 0) info = 5;
if (ku < 0) info = 4;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if ((m==0) || (n==0)) return;
lenx = n;
leny = m;
if (trans & 1) lenx = m;
if (trans & 1) leny = n;
if (beta_r != ONE || beta_i != ZERO) SCAL_K(leny, 0, 0, beta_r, beta_i, y, abs(incy), NULL, 0, NULL, 0);
if (alpha_r == ZERO && alpha_i == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (lenx - 1) * incx * 2;
if (incy < 0) y -= (leny - 1) * incy * 2;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(gbmv[(int)trans])(m, n, kl, ku, alpha_r, alpha_i, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(gbmv_thread[(int)trans])(m, n, kl, ku, ALPHA, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, m * n / 2 + n, m * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, char *TRANS, char *DIAG,
blasint *N, FLOAT *a, blasint *LDA, FLOAT *x, blasint *INCX){
char uplo_arg = *UPLO;
char trans_arg = *TRANS;
char diag_arg = *DIAG;
blasint n = *N;
blasint lda = *LDA;
blasint incx = *INCX;
blasint info;
int uplo;
int unit;
int trans, buffer_size;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
TOUPPER(trans_arg);
TOUPPER(diag_arg);
trans = -1;
unit = -1;
uplo = -1;
if (trans_arg == 'N') trans = 0;
if (trans_arg == 'T') trans = 1;
if (trans_arg == 'R') trans = 2;
if (trans_arg == 'C') trans = 3;
if (diag_arg == 'U') unit = 0;
if (diag_arg == 'N') unit = 1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 8;
if (lda < MAX(1, n)) info = 6;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
blasint n, FLOAT *a, blasint lda, FLOAT *x, blasint incx) {
int trans, uplo, unit, buffer_size;
blasint info;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
unit = -1;
uplo = -1;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 2;
if (TransA == CblasConjTrans) trans = 3;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 8;
if (lda < MAX(1, n)) info = 6;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 3;
if (TransA == CblasConjTrans) trans = 2;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 8;
if (lda < MAX(1, n)) info = 6;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx * 2;
#ifdef SMP
// Calibrated on a Xeon E5-2630
if(1L * n * n > 36L * sizeof(FLOAT) * sizeof(FLOAT) * GEMM_MULTITHREAD_THRESHOLD) {
nthreads = num_cpu_avail(2);
if(nthreads > 2 && 1L * n * n < 64L * sizeof(FLOAT) * sizeof(FLOAT) * GEMM_MULTITHREAD_THRESHOLD)
nthreads = 2;
} else
nthreads = 1;
if(nthreads > 1) {
buffer_size = n > 16 ? 0 : n * 4 + 40;
}
else
#endif
{
buffer_size = ((n - 1) / DTB_ENTRIES) * 2 * DTB_ENTRIES + 32 / sizeof(FLOAT);
if(incx != 1)
buffer_size += n * 2;
}
STACK_ALLOC(buffer_size, FLOAT, buffer);
#ifdef SMP
if (nthreads == 1) {
#endif
(trmv[(trans<<2) | (uplo<<1) | unit])(n, a, lda, x, incx, buffer);
#ifdef SMP
} else {
(trmv_thread[(trans<<2) | (uplo<<1) | unit])(n, a, lda, x, incx, buffer, nthreads);
}
#endif
STACK_FREE(buffer);
FUNCTION_PROFILE_END(4, n * n / 2 + n, n * n);
IDEBUG_END;
return;
}
void NAME(char *SIDE, char *UPLO,
blasint *M, blasint *N,
FLOAT *alpha, FLOAT *a, blasint *ldA,
FLOAT *b, blasint *ldB,
FLOAT *beta, FLOAT *c, blasint *ldC){
char side_arg = *SIDE;
char uplo_arg = *UPLO;
blas_arg_t args;
FLOAT *buffer;
FLOAT *sa, *sb;
#if defined(SMP) && !defined(NO_AFFINITY)
int nodes;
#endif
blasint info;
int side;
int uplo;
PRINT_DEBUG_NAME;
args.alpha = (void *)alpha;
args.beta = (void *)beta;
TOUPPER(side_arg);
TOUPPER(uplo_arg);
side = -1;
uplo = -1;
if (side_arg == 'L') side = 0;
if (side_arg == 'R') side = 1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
args.m = *M;
args.n = *N;
args.c = (void *)c;
args.ldc = *ldC;
info = 0;
if (args.ldc < MAX(1, args.m)) info = 12;
if (!side) {
args.a = (void *)a;
args.b = (void *)b;
args.lda = *ldA;
args.ldb = *ldB;
if (args.ldb < MAX(1, args.m)) info = 9;
if (args.lda < MAX(1, args.m)) info = 7;
} else {
args.a = (void *)b;
args.b = (void *)a;
args.lda = *ldB;
args.ldb = *ldA;
if (args.lda < MAX(1, args.m)) info = 9;
if (args.ldb < MAX(1, args.n)) info = 7;
}
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (uplo < 0) info = 2;
if (side < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo,
blasint m, blasint n,
#ifndef COMPLEX
FLOAT alpha,
#else
FLOAT *alpha,
#endif
FLOAT *a, blasint lda,
FLOAT *b, blasint ldb,
#ifndef COMPLEX
FLOAT beta,
#else
FLOAT *beta,
#endif
FLOAT *c, blasint ldc) {
blas_arg_t args;
int side, uplo;
blasint info;
FLOAT *buffer;
FLOAT *sa, *sb;
#if defined(SMP) && !defined(NO_AFFINITY)
int nodes;
#endif
PRINT_DEBUG_CNAME;
#ifndef COMPLEX
args.alpha = (void *)α
args.beta = (void *)β
#else
args.alpha = (void *)alpha;
args.beta = (void *)beta;
#endif
args.c = (void *)c;
args.ldc = ldc;
side = -1;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Side == CblasLeft) side = 0;
if (Side == CblasRight) side = 1;
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
args.m = m;
args.n = n;
if (args.ldc < MAX(1, args.m)) info = 12;
if (!side) {
args.a = (void *)a;
args.b = (void *)b;
args.lda = lda;
args.ldb = ldb;
if (args.ldb < MAX(1, args.m)) info = 9;
if (args.lda < MAX(1, args.m)) info = 7;
} else {
args.a = (void *)b;
args.b = (void *)a;
args.lda = ldb;
args.ldb = lda;
if (args.lda < MAX(1, args.m)) info = 9;
if (args.ldb < MAX(1, args.n)) info = 7;
}
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (uplo < 0) info = 2;
if (side < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Side == CblasLeft) side = 1;
if (Side == CblasRight) side = 0;
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
info = -1;
args.m = n;
args.n = m;
if (args.ldc < MAX(1, args.m)) info = 12;
if (!side) {
args.a = (void *)a;
args.b = (void *)b;
args.lda = lda;
args.ldb = ldb;
if (args.ldb < MAX(1, args.m)) info = 9;
if (args.lda < MAX(1, args.m)) info = 7;
} else {
args.a = (void *)b;
args.b = (void *)a;
args.lda = ldb;
args.ldb = lda;
if (args.lda < MAX(1, args.m)) info = 9;
if (args.ldb < MAX(1, args.n)) info = 7;
}
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (uplo < 0) info = 2;
if (side < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (args.m == 0 || args.n == 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
buffer = (FLOAT *)blas_memory_alloc(0);
sa = (FLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
sb = (FLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#ifdef SMP
args.common = NULL;
args.nthreads = num_cpu_avail(3);
if (args.nthreads == 1) {
#endif
(symm[(side << 1) | uplo ])(&args, NULL, NULL, sa, sb, 0);
#ifdef SMP
} else {
#ifndef NO_AFFINITY
nodes = get_num_nodes();
if (nodes > 1) {
args.nthreads /= nodes;
gemm_thread_mn(MODE, &args, NULL, NULL,
symm[4 | (side << 1) | uplo ], sa, sb, nodes);
} else {
#endif
#ifndef USE_SIMPLE_THREADED_LEVEL3
(symm[4 | (side << 1) | uplo ])(&args, NULL, NULL, sa, sb, 0);
#else
GEMM_THREAD(MODE, &args, NULL, NULL, symm[(side << 1) | uplo ], sa, sb, args.nthreads);
#endif
#ifndef NO_AFFINITY
}
#endif
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(COMPSIZE * COMPSIZE,
(!side)? args.m * (args.m / 2 + args.n) : args.n * (args.m + args.n / 2),
(!side)? 2 * args.m * args.m * args.n : 2 * args.m * args.n * args.n);
IDEBUG_END;
return;
}
void NAME(char *TRANS, blasint *M, blasint *N,
FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX,
FLOAT *BETA, FLOAT *y, blasint *INCY){
char trans = *TRANS;
blasint m = *M;
blasint n = *N;
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
FLOAT alpha = *ALPHA;
FLOAT beta = *BETA;
FLOAT *buffer;
#ifdef SMP
int nthreads;
int nthreads_max;
int nthreads_avail;
double MNK;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T,
};
blasint info;
blasint lenx, leny;
blasint i;
PRINT_DEBUG_NAME;
TOUPPER(trans);
info = 0;
i = -1;
if (trans == 'N') i = 0;
if (trans == 'T') i = 1;
if (trans == 'R') i = 0;
if (trans == 'C') i = 1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (i < 0) info = 1;
trans = i;
if (info != 0){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_TRANSPOSE TransA,
blasint m, blasint n,
FLOAT alpha,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT beta,
FLOAT *y, blasint incy){
FLOAT *buffer;
blasint lenx, leny;
int trans;
blasint info, t;
#ifdef SMP
int nthreads;
int nthreads_max;
int nthreads_avail;
double MNK;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T,
};
PRINT_DEBUG_CNAME;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 0;
if (TransA == CblasConjTrans) trans = 1;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (order == CblasRowMajor) {
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 1;
if (TransA == CblasConjTrans) trans = 0;
info = -1;
t = n;
n = m;
m = t;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
//printf("m=%d, n=%d, trans=%d, incx=%d, incy=%d, alpha=%f, beta=%f\n", m, n, trans, incx, incy, alpha, beta);
if ((m==0) || (n==0)) return;
lenx = n;
leny = m;
if (trans) lenx = m;
if (trans) leny = n;
if (beta != ONE) SCAL_K(leny, 0, 0, beta, y, abs(incy), NULL, 0, NULL, 0);
if (alpha == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (lenx - 1) * incx;
if (incy < 0) y -= (leny - 1) * incy;
#ifdef MAX_STACK_ALLOC
// make it volatile because some gemv implementation (ex: dgemv_n.S)
// do not restore all register
volatile int stack_alloc_size = 0;
//for gemv_n and gemv_t, try to allocate on stack
stack_alloc_size = m + n;
#ifdef ALIGNED_ACCESS
stack_alloc_size += 3;
#endif
// if(stack_alloc_size < 128)
//dgemv_n.S require a 128 bytes buffer
// increasing instead of capping 128
// ABI STACK for windows 288 bytes
stack_alloc_size += 288 / sizeof(FLOAT) ;
if(stack_alloc_size > MAX_STACK_ALLOC / sizeof(FLOAT))
stack_alloc_size = 0;
// stack overflow check
volatile double stack_check = 3.14159265358979323846;
FLOAT stack_buffer[stack_alloc_size];
buffer = stack_alloc_size ? stack_buffer : (FLOAT *)blas_memory_alloc(1);
// printf("stack_alloc_size=%d\n", stack_alloc_size);
#else
//Original OpenBLAS/GotoBLAS codes.
buffer = (FLOAT *)blas_memory_alloc(1);
#endif
#ifdef SMP
nthreads_max = num_cpu_avail(2);
nthreads_avail = nthreads_max;
MNK = (double) m * (double) n;
if ( MNK <= (24.0 * 24.0 * (double) (GEMM_MULTITHREAD_THRESHOLD*GEMM_MULTITHREAD_THRESHOLD) ) )
nthreads_max = 1;
if ( nthreads_max > nthreads_avail )
nthreads = nthreads_avail;
else
nthreads = nthreads_max;
if (nthreads == 1) {
#endif
(gemv[(int)trans])(m, n, 0, alpha, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(gemv_thread[(int)trans])(m, n, alpha, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
// stack overflow check
assert(stack_check==3.14159265358979323846);
#ifdef MAX_STACK_ALLOC
if(!stack_alloc_size){
blas_memory_free(buffer);
}
#else
blas_memory_free(buffer);
#endif
FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);
IDEBUG_END;
return;
}
void NAME(FLOAT *DA, FLOAT *DB, FLOAT *C, FLOAT *S){
#if defined(__i386__) || defined(__x86_64__) || defined(__ia64__) || defined(_M_X64) || defined(_M_IX86)
long double da_r = *(DA + 0);
long double da_i = *(DA + 1);
long double db_r = *(DB + 0);
long double db_i = *(DB + 1);
long double r;
long double ada = fabs(da_r) + fabs(da_i);
PRINT_DEBUG_NAME;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (ada == ZERO) {
*C = ZERO;
*(S + 0) = ONE;
*(S + 1) = ZERO;
*(DA + 0) = db_r;
*(DA + 1) = db_i;
} else {
long double alpha_r, alpha_i;
ada = sqrt(da_r * da_r + da_i * da_i);
r = sqrt(da_r * da_r + da_i * da_i + db_r * db_r + db_i * db_i);
alpha_r = da_r / ada;
alpha_i = da_i / ada;
*(C + 0) = ada / r;
*(S + 0) = (alpha_r * db_r + alpha_i *db_i) / r;
*(S + 1) = (alpha_i * db_r - alpha_r *db_i) / r;
*(DA + 0) = alpha_r * r;
*(DA + 1) = alpha_i * r;
}
#else
FLOAT da_r = *(DA + 0);
FLOAT da_i = *(DA + 1);
FLOAT db_r = *(DB + 0);
FLOAT db_i = *(DB + 1);
FLOAT r;
FLOAT ada = fabs(da_r) + fabs(da_i);
FLOAT adb;
PRINT_DEBUG_NAME;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (ada == ZERO) {
*C = ZERO;
*(S + 0) = ONE;
*(S + 1) = ZERO;
*(DA + 0) = db_r;
*(DA + 1) = db_i;
} else {
FLOAT scale;
FLOAT aa_r, aa_i, bb_r, bb_i;
FLOAT alpha_r, alpha_i;
aa_r = fabs(da_r);
aa_i = fabs(da_i);
if (aa_i > aa_r) {
aa_r = fabs(da_i);
aa_i = fabs(da_r);
}
scale = (aa_i / aa_r);
ada = aa_r * sqrt(ONE + scale * scale);
bb_r = fabs(db_r);
bb_i = fabs(db_i);
if (bb_i > bb_r) {
bb_r = fabs(bb_i);
bb_i = fabs(bb_r);
}
scale = (bb_i / bb_r);
adb = bb_r * sqrt(ONE + scale * scale);
scale = ada + adb;
aa_r = da_r / scale;
aa_i = da_i / scale;
bb_r = db_r / scale;
bb_i = db_i / scale;
r = scale * sqrt(aa_r * aa_r + aa_i * aa_i + bb_r * bb_r + bb_i * bb_i);
alpha_r = da_r / ada;
alpha_i = da_i / ada;
*(C + 0) = ada / r;
*(S + 0) = (alpha_r * db_r + alpha_i *db_i) / r;
*(S + 1) = (alpha_i * db_r - alpha_r *db_i) / r;
*(DA + 0) = alpha_r * r;
*(DA + 1) = alpha_i * r;
}
#endif
FUNCTION_PROFILE_END(4, 4, 4);
IDEBUG_END;
return;
}
void NAME(char *TRANS, blasint *M, blasint *N,
FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX,
FLOAT *BETA, FLOAT *y, blasint *INCY){
char trans = *TRANS;
blasint m = *M;
blasint n = *N;
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT, FLOAT *, BLASLONG,
FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T, GEMV_R, GEMV_C,
GEMV_O, GEMV_U, GEMV_S, GEMV_D,
};
blasint info;
blasint lenx, leny;
blasint i;
PRINT_DEBUG_NAME;
FLOAT alpha_r = *(ALPHA + 0);
FLOAT alpha_i = *(ALPHA + 1);
FLOAT beta_r = *(BETA + 0);
FLOAT beta_i = *(BETA + 1);
TOUPPER(trans);
info = 0;
i = -1;
if (trans == 'N') i = 0;
if (trans == 'T') i = 1;
if (trans == 'R') i = 2;
if (trans == 'C') i = 3;
if (trans == 'O') i = 4;
if (trans == 'U') i = 5;
if (trans == 'S') i = 6;
if (trans == 'D') i = 7;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1,m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (i < 0) info = 1;
trans = i;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_TRANSPOSE TransA,
blasint m, blasint n,
FLOAT *ALPHA,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT *BETA,
FLOAT *y, blasint incy){
FLOAT *buffer;
blasint lenx, leny;
int trans;
blasint info, t;
#ifdef SMP
int nthreads;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT, FLOAT *, BLASLONG,
FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T, GEMV_R, GEMV_C,
GEMV_O, GEMV_U, GEMV_S, GEMV_D,
};
PRINT_DEBUG_CNAME;
FLOAT alpha_r = *(ALPHA + 0);
FLOAT alpha_i = *(ALPHA + 1);
FLOAT beta_r = *(BETA + 0);
FLOAT beta_i = *(BETA + 1);
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 2;
if (TransA == CblasConjTrans) trans = 3;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (order == CblasRowMajor) {
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 3;
if (TransA == CblasConjTrans) trans = 2;
info = -1;
t = n;
n = m;
m = t;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
/* Quick return if possible. */
if (m == 0 || n == 0) return;
lenx = n;
leny = m;
if (trans & 1) lenx = m;
if (trans & 1) leny = n;
if (beta_r != ONE || beta_i != ZERO) SCAL_K(leny, 0, 0, beta_r, beta_i, y, abs(incy), NULL, 0, NULL, 0);
if (alpha_r == ZERO && alpha_i == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (lenx - 1) * incx * 2;
if (incy < 0) y -= (leny - 1) * incy * 2;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(gemv[(int)trans])(m, n, 0, alpha_r, alpha_i, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(gemv_thread[(int)trans])(m, n, ALPHA, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, m * n + m + n, 2 * m * n);
IDEBUG_END;
return;
}
void NAME(blasint *N, FLOAT *ALPHA, FLOAT *x, blasint *INCX){
blasint n = *N;
blasint incx = *INCX;
FLOAT alpha = *ALPHA;
#else
void CNAME(blasint n, FLOAT alpha, FLOAT *x, blasint incx){
#endif
#ifdef SMP
int mode, nthreads;
#endif
#ifndef CBLAS
PRINT_DEBUG_NAME;
#else
PRINT_DEBUG_CNAME;
#endif
if (incx <= 0 || n <= 0) return;
if (alpha == ONE) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
#ifdef SMP
nthreads = num_cpu_avail(1);
if (nthreads == 1) {
#endif
SCAL_K(n, 0, 0, alpha, x, incx, NULL, 0, NULL, 0);
#ifdef SMP
} else {
#ifdef DOUBLE
mode = BLAS_DOUBLE | BLAS_REAL;
#else
mode = BLAS_SINGLE | BLAS_REAL;
#endif
blas_level1_thread(mode, n, 0, 0,
#ifndef CBLAS
ALPHA,
#else
&alpha,
#endif
x, incx, NULL, 0, NULL, 0, (void *)SCAL_K, nthreads);
}
#endif
FUNCTION_PROFILE_END(1, n, n);
IDEBUG_END;
return;
}
void NAME(blasint *M, blasint *N, FLOAT *Alpha,
FLOAT *x, blasint *INCX,
FLOAT *y, blasint *INCY,
FLOAT *a, blasint *LDA){
blasint m = *M;
blasint n = *N;
FLOAT alpha = *Alpha;
blasint incx = *INCX;
blasint incy = *INCY;
blasint lda = *LDA;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
blasint info;
PRINT_DEBUG_NAME;
info = 0;
if (lda < MAX(1,m)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (m < 0) info = 1;
if (info){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
blasint m, blasint n,
FLOAT alpha,
FLOAT *x, blasint incx,
FLOAT *y, blasint incy,
FLOAT *a, blasint lda) {
FLOAT *buffer;
blasint info, t;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
info = 0;
if (order == CblasColMajor) {
info = -1;
if (lda < MAX(1,m)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (m < 0) info = 1;
}
if (order == CblasRowMajor) {
info = -1;
t = n;
n = m;
m = t;
t = incx;
incx = incy;
incy = t;
buffer = x;
x = y;
y = buffer;
if (lda < MAX(1,m)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (m < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
/* Quick return if possible. */
if (m == 0 || n == 0) return;
if (alpha == 0.) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incy < 0) y -= (n - 1) * incy;
if (incx < 0) x -= (m - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
GER(m, n, 0, alpha, x, incx, y, incy, a, lda, buffer);
#ifdef SMP
} else {
GER_THREAD(m, n, alpha, x, incx, y, incy, a, lda, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA,
FLOAT *x, blasint *INCX, FLOAT *a){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha = *ALPHA;
blasint incx = *INCX;
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_UPLO Uplo,
blasint n,
FLOAT alpha,
FLOAT *x, blasint incx,
FLOAT *a) {
FLOAT *buffer;
int uplo;
blasint info;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
info = -1;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
if (alpha == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(spr[uplo])(n, alpha, x, incx, a, buffer);
#ifdef SMP
} else {
(spr_thread[uplo])(n, alpha, x, incx, a, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, n * n / 2 + n, n * n);
IDEBUG_END;
return;
}
void NAME(blasint *N, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY){
blasint n = *N;
blasint incx = *INCX;
blasint incy = *INCY;
#else
void CNAME(blasint n, FLOAT *x, blasint incx, FLOAT *y, blasint incy){
#endif
#ifdef SMP
int mode;
FLOAT dummyalpha[2] = {ZERO, ZERO};
int nthreads;
#endif
#ifndef CBLAS
PRINT_DEBUG_NAME;
#else
PRINT_DEBUG_CNAME;
#endif
if (n <= 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (n - 1) * incx * 2;
if (incy < 0) y -= (n - 1) * incy * 2;
#ifdef SMP
nthreads = num_cpu_avail(1);
if (nthreads == 1) {
#endif
SWAP_K(n, 0, 0, ZERO, ZERO, x, incx, y, incy, NULL, 0);
#ifdef SMP
} else {
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
blas_level1_thread(mode, n, 0, 0, dummyalpha,
x, incx, y, incy, NULL, 0, (void *)SWAP_K, nthreads);
}
#endif
FUNCTION_PROFILE_END(2, 2 * n, 0);
IDEBUG_END;
return;
}
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA,
FLOAT *x, blasint *INCX, FLOAT *a){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
blasint incx = *INCX;
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
if (n == 0) return;
if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(spr[uplo])(n, alpha_r, alpha_i, x, incx, a, buffer);
#ifdef SMP
} else {
(spr_thread[uplo])(n, ALPHA, x, incx, a, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, n * n / 2 + n, n * n);
IDEBUG_END;
return;
}
void NAME(char *TRANSA, char *TRANSB,
blasint *M, blasint *N, blasint *K,
FLOAT *alpha,
FLOAT *a, blasint *ldA,
FLOAT *b, blasint *ldB,
FLOAT *beta,
FLOAT *c, blasint *ldC){
blas_arg_t args;
int transa, transb, nrowa, nrowb;
blasint info;
char transA, transB;
FLOAT *buffer;
FLOAT *sa, *sb;
#ifdef SMP
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
#if defined(SMP) && !defined(NO_AFFINITY) && !defined(USE_SIMPLE_THREADED_LEVEL3)
int nodes;
#endif
PRINT_DEBUG_NAME;
args.m = *M;
args.n = *N;
args.k = *K;
args.a = (void *)a;
args.b = (void *)b;
args.c = (void *)c;
args.lda = *ldA;
args.ldb = *ldB;
args.ldc = *ldC;
args.alpha = (void *)alpha;
args.beta = (void *)beta;
transA = *TRANSA;
transB = *TRANSB;
TOUPPER(transA);
TOUPPER(transB);
transa = -1;
transb = -1;
if (transA == 'N') transa = 0;
if (transA == 'T') transa = 1;
#ifndef COMPLEX
if (transA == 'R') transa = 0;
if (transA == 'C') transa = 1;
#else
if (transA == 'R') transa = 2;
if (transA == 'C') transa = 3;
#endif
if (transB == 'N') transb = 0;
if (transB == 'T') transb = 1;
#ifndef COMPLEX
if (transB == 'R') transb = 0;
if (transB == 'C') transb = 1;
#else
if (transB == 'R') transb = 2;
if (transB == 'C') transb = 3;
#endif
nrowa = args.m;
if (transa & 1) nrowa = args.k;
nrowb = args.k;
if (transb & 1) nrowb = args.n;
info = 0;
if (args.ldc < args.m) info = 13;
if (args.ldb < nrowb) info = 10;
if (args.lda < nrowa) info = 8;
if (args.k < 0) info = 5;
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (transb < 0) info = 2;
if (transa < 0) info = 1;
if (info){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB,
blasint m, blasint n, blasint k,
#ifndef COMPLEX
FLOAT alpha,
#else
FLOAT *alpha,
#endif
FLOAT *a, blasint lda,
FLOAT *b, blasint ldb,
#ifndef COMPLEX
FLOAT beta,
#else
FLOAT *beta,
#endif
FLOAT *c, blasint ldc) {
blas_arg_t args;
int transa, transb;
blasint nrowa, nrowb, info;
XFLOAT *buffer;
XFLOAT *sa, *sb;
#ifdef SMP
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
#if defined(SMP) && !defined(NO_AFFINITY) && !defined(USE_SIMPLE_THREADED_LEVEL3)
int nodes;
#endif
PRINT_DEBUG_CNAME;
#ifndef COMPLEX
args.alpha = (void *)α
args.beta = (void *)β
#else
args.alpha = (void *)alpha;
args.beta = (void *)beta;
#endif
transa = -1;
transb = -1;
info = 0;
if (order == CblasColMajor) {
args.m = m;
args.n = n;
args.k = k;
args.a = (void *)a;
args.b = (void *)b;
args.c = (void *)c;
args.lda = lda;
args.ldb = ldb;
args.ldc = ldc;
if (TransA == CblasNoTrans) transa = 0;
if (TransA == CblasTrans) transa = 1;
#ifndef COMPLEX
if (TransA == CblasConjNoTrans) transa = 0;
if (TransA == CblasConjTrans) transa = 1;
#else
if (TransA == CblasConjNoTrans) transa = 2;
if (TransA == CblasConjTrans) transa = 3;
#endif
if (TransB == CblasNoTrans) transb = 0;
if (TransB == CblasTrans) transb = 1;
#ifndef COMPLEX
if (TransB == CblasConjNoTrans) transb = 0;
if (TransB == CblasConjTrans) transb = 1;
#else
if (TransB == CblasConjNoTrans) transb = 2;
if (TransB == CblasConjTrans) transb = 3;
#endif
nrowa = args.m;
if (transa & 1) nrowa = args.k;
nrowb = args.k;
if (transb & 1) nrowb = args.n;
info = -1;
if (args.ldc < args.m) info = 13;
if (args.ldb < nrowb) info = 10;
if (args.lda < nrowa) info = 8;
if (args.k < 0) info = 5;
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (transb < 0) info = 2;
if (transa < 0) info = 1;
}
if (order == CblasRowMajor) {
args.m = n;
args.n = m;
args.k = k;
args.a = (void *)b;
args.b = (void *)a;
args.c = (void *)c;
args.lda = ldb;
args.ldb = lda;
args.ldc = ldc;
if (TransB == CblasNoTrans) transa = 0;
if (TransB == CblasTrans) transa = 1;
#ifndef COMPLEX
if (TransB == CblasConjNoTrans) transa = 0;
if (TransB == CblasConjTrans) transa = 1;
#else
if (TransB == CblasConjNoTrans) transa = 2;
if (TransB == CblasConjTrans) transa = 3;
#endif
if (TransA == CblasNoTrans) transb = 0;
if (TransA == CblasTrans) transb = 1;
#ifndef COMPLEX
if (TransA == CblasConjNoTrans) transb = 0;
if (TransA == CblasConjTrans) transb = 1;
#else
if (TransA == CblasConjNoTrans) transb = 2;
if (TransA == CblasConjTrans) transb = 3;
#endif
nrowa = args.m;
if (transa & 1) nrowa = args.k;
nrowb = args.k;
if (transb & 1) nrowb = args.n;
info = -1;
if (args.ldc < args.m) info = 13;
if (args.ldb < nrowb) info = 10;
if (args.lda < nrowa) info = 8;
if (args.k < 0) info = 5;
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (transb < 0) info = 2;
if (transa < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if ((args.m == 0) || (args.n == 0)) return;
#if 0
fprintf(stderr, "m = %4d n = %d k = %d lda = %4d ldb = %4d ldc = %4d\n",
args.m, args.n, args.k, args.lda, args.ldb, args.ldc);
#endif
IDEBUG_START;
FUNCTION_PROFILE_START();
buffer = (XFLOAT *)blas_memory_alloc(0);
sa = (XFLOAT *)((BLASLONG)buffer +GEMM_OFFSET_A);
sb = (XFLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#ifdef SMP
mode |= (transa << BLAS_TRANSA_SHIFT);
mode |= (transb << BLAS_TRANSB_SHIFT);
args.common = NULL;
args.nthreads = num_cpu_avail(3);
if (args.nthreads == 1) {
#endif
(gemm[(transb << 2) | transa])(&args, NULL, NULL, sa, sb, 0);
#ifdef SMP
} else {
#ifndef USE_SIMPLE_THREADED_LEVEL3
#ifndef NO_AFFINITY
nodes = get_num_nodes();
if ((nodes > 1) && get_node_equal()) {
args.nthreads /= nodes;
gemm_thread_mn(mode, &args, NULL, NULL, gemm[16 | (transb << 2) | transa], sa, sb, nodes);
} else {
#endif
(gemm[16 | (transb << 2) | transa])(&args, NULL, NULL, sa, sb, 0);
#else
GEMM_THREAD(mode, &args, NULL, NULL, gemm[(transb << 2) | transa], sa, sb, args.nthreads);
#endif
#ifndef USE_SIMPLE_THREADED_LEVEL3
#ifndef NO_AFFINITY
}
#endif
#endif
#endif
#ifdef SMP
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(COMPSIZE * COMPSIZE, args.m * args.k + args.k * args.n + args.m * args.n, 2 * args.m * args.n * args.k);
IDEBUG_END;
return;
}
void NAME(blasint *N, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY){
blasint n = *N;
blasint incx = *INCX;
blasint incy = *INCY;
#else
void CNAME(blasint n, FLOAT *x, blasint incx, FLOAT *y, blasint incy){
#endif
#ifdef SMP
int mode, nthreads;
FLOAT dummyalpha[2] = {ZERO, ZERO};
#endif
#ifndef CBLAS
PRINT_DEBUG_NAME;
#else
PRINT_DEBUG_CNAME;
#endif
if (n <= 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (n - 1) * incx;
if (incy < 0) y -= (n - 1) * incy;
#ifdef SMP
nthreads = num_cpu_avail(1);
//disable multi-thread when incx==0 or incy==0
//In that case, the threads would be dependent.
if (incx == 0 || incy == 0)
nthreads = 1;
if (nthreads == 1) {
#endif
SWAP_K(n, 0, 0, ZERO, x, incx, y, incy, NULL, 0);
#ifdef SMP
} else {
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_REAL;
#else
mode = BLAS_SINGLE | BLAS_REAL;
#endif
blas_level1_thread(mode, n, 0, 0, dummyalpha,
x, incx, y, incy, NULL, 0, (void *)SWAP_K, nthreads);
}
#endif
FUNCTION_PROFILE_END(1, 2 * n, 0);
IDEBUG_END;
return;
}
void NAME(char *UPLO, char *TRANS, char *DIAG,
blasint *N, FLOAT *a, FLOAT *x, blasint *INCX){
char uplo_arg = *UPLO;
char trans_arg = *TRANS;
char diag_arg = *DIAG;
blasint n = *N;
blasint incx = *INCX;
blasint info;
int uplo;
int unit;
int trans;
FLOAT *buffer;
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
TOUPPER(trans_arg);
TOUPPER(diag_arg);
trans = -1;
unit = -1;
uplo = -1;
if (trans_arg == 'N') trans = 0;
if (trans_arg == 'T') trans = 1;
if (trans_arg == 'R') trans = 0;
if (trans_arg == 'C') trans = 1;
if (diag_arg == 'U') unit = 0;
if (diag_arg == 'N') unit = 1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 7;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
blasint n, FLOAT *a, FLOAT *x, blasint incx) {
int trans, uplo, unit;
blasint info;
FLOAT *buffer;
PRINT_DEBUG_CNAME;
unit = -1;
uplo = -1;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 0;
if (TransA == CblasConjTrans) trans = 1;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 7;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 1;
if (TransA == CblasConjTrans) trans = 0;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 7;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
(tpsv[(trans<<2) | (uplo<<1) | unit])(n, a, x, incx, buffer);
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, n * n / 2 + n, n * n);
IDEBUG_END;
return;
}
void NAME(blasint *N, FLOAT *ALPHA, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY){
blasint n = *N;
blasint incx = *INCX;
blasint incy = *INCY;
#else
void CNAME(blasint n, FLOAT *ALPHA, FLOAT *x, blasint incx, FLOAT *y, blasint incy){
#endif
FLOAT alpha_r = *(ALPHA + 0);
FLOAT alpha_i = *(ALPHA + 1);
#ifdef SMP
int mode, nthreads;
#endif
#ifndef CBLAS
PRINT_DEBUG_CNAME;
#else
PRINT_DEBUG_CNAME;
#endif
if (n <= 0) return;
if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (n - 1) * incx * 2;
if (incy < 0) y -= (n - 1) * incy * 2;
#ifdef SMP
nthreads = num_cpu_avail(1);
if (nthreads == 1) {
#endif
#ifndef CONJ
AXPYU_K (n, 0, 0, alpha_r, alpha_i, x, incx, y, incy, NULL, 0);
#else
AXPYC_K(n, 0, 0, alpha_r, alpha_i, x, incx, y, incy, NULL, 0);
#endif
#ifdef SMP
} else {
#ifdef XDOUBLE
mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
blas_level1_thread(mode, n, 0, 0, ALPHA, x, incx, y, incy, NULL, 0,
#ifndef CONJ
(void *)AXPYU_K,
#else
(void *)AXPYC_K,
#endif
nthreads);
}
#endif
FUNCTION_PROFILE_END(4, 2 * n, 2 * n);
IDEBUG_END;
return;
}
int NAME(char *UPLO, char *DIAG, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){
blas_arg_t args;
blasint uplo_arg = *UPLO;
blasint diag_arg = *DIAG;
blasint uplo, diag;
blasint info;
FLOAT *buffer;
#ifdef PPC440
extern
#endif
FLOAT *sa, *sb;
PRINT_DEBUG_NAME;
args.n = *N;
args.a = (void *)a;
args.lda = *ldA;
TOUPPER(uplo_arg);
TOUPPER(diag_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
diag = -1;
if (diag_arg == 'U') diag = 0;
if (diag_arg == 'N') diag = 1;
info = 0;
if (args.lda < MAX(1,args.n)) info = 5;
if (args.n < 0) info = 3;
if (diag < 0) info = 2;
if (uplo < 0) info = 1;
if (info) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
*Info = - info;
return 0;
}
*Info = 0;
if (args.n <= 0) return 0;
IDEBUG_START;
FUNCTION_PROFILE_START();
#ifndef PPC440
buffer = (FLOAT *)blas_memory_alloc(1);
sa = (FLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
sb = (FLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#endif
info = (trti2[(uplo << 1) | diag])(&args, NULL, NULL, sa, sb, 0);
*Info = info;
#ifndef PPC440
blas_memory_free(buffer);
#endif
FUNCTION_PROFILE_END(1, .5 * args.n * args.n,
2. * args.n * (1./3. + args.n * ( 1./2. + args.n * 1./6.))
+ 6. * args.n * (1./3. + args.n * (-1./2. + args.n * 1./6.)));
IDEBUG_END;
return 0;
}
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *b, blasint *INCX, FLOAT *BETA, FLOAT *c, blasint *INCY){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
blasint lda = *LDA;
blasint incx = *INCX;
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
blasint incy = *INCY;
int (*symv[])(BLASLONG, BLASLONG, FLOAT, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
SYMV_U, SYMV_L,
};
#ifdef SMP
int (*symv_thread[])(BLASLONG, FLOAT *, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, int) = {
SYMV_THREAD_U, SYMV_THREAD_L,
};
#endif
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incy == 0) info = 10;
if (incx == 0) info = 7;
if (lda < MAX(1, n)) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
if (n == 0) return;
if ((beta_r != ONE) || (beta_i != ZERO)) SCAL_K(n, 0, 0, beta_r, beta_i, c, abs(incy), NULL, 0, NULL, 0);
if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) b -= (n - 1) * incx * COMPSIZE;
if (incy < 0 ) c -= (n - 1) * incy * COMPSIZE;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(symv[uplo])(n, n, alpha_r, alpha_i, a, lda, b, incx, c, incy, buffer);
#ifdef SMP
} else {
(symv_thread[uplo])(n, ALPHA, a, lda, b, incx, c, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, n * n / 2 + 2 * n, 2 * n * n);
IDEBUG_END;
return;
}
int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){
blas_arg_t args;
blasint uplo_arg = *UPLO;
blasint uplo;
blasint info;
FLOAT *buffer;
#ifdef PPC440
extern
#endif
FLOAT *sa, *sb;
PRINT_DEBUG_NAME;
args.n = *N;
args.a = (void *)a;
args.lda = *ldA;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (args.lda < MAX(1,args.n)) info = 4;
if (args.n < 0) info = 2;
if (uplo < 0) info = 1;
if (info) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
*Info = - info;
return 0;
}
*Info = 0;
if (args.n == 0) return 0;
IDEBUG_START;
FUNCTION_PROFILE_START();
#ifndef PPC440
buffer = (FLOAT *)blas_memory_alloc(1);
sa = (FLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
sb = (FLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#endif
#ifdef SMP
args.common = NULL;
args.nthreads = num_cpu_avail(4);
if (args.nthreads == 1) {
#endif
*Info = (lauum_single[uplo])(&args, NULL, NULL, sa, sb, 0);
#ifdef SMP
} else {
*Info = (lauum_parallel[uplo])(&args, NULL, NULL, sa, sb, 0);
}
#endif
#ifndef PPC440
blas_memory_free(buffer);
#endif
FUNCTION_PROFILE_END(1, .5 * args.n * args.n,
2. * args.n * (1./3. + args.n * ( 1./2. + args.n * 1./6.))
+ args.n * (args.n * args.n - 1));
IDEBUG_END;
return 0;
}
int NAME(blasint *N, blasint *NRHS, FLOAT *a, blasint *ldA, blasint *ipiv,
FLOAT *b, blasint *ldB, blasint *Info){
blas_arg_t args;
blasint info;
FLOAT *buffer;
#ifdef PPC440
extern
#endif
FLOAT *sa, *sb;
PRINT_DEBUG_NAME;
args.m = *N;
args.n = *NRHS;
args.a = (void *)a;
args.lda = *ldA;
args.b = (void *)b;
args.ldb = *ldB;
args.c = (void *)ipiv;
info = 0;
if (args.ldb < MAX(1,args.m)) info = 7;
if (args.lda < MAX(1,args.m)) info = 4;
if (args.n < 0) info = 2;
if (args.m < 0) info = 1;
if (info) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
*Info = - info;
return 0;
}
args.alpha = NULL;
args.beta = NULL;
*Info = 0;
if (args.m == 0 || args.n == 0) return 0;
IDEBUG_START;
FUNCTION_PROFILE_START();
#ifndef PPC440
buffer = (FLOAT *)blas_memory_alloc(1);
sa = (FLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
sb = (FLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#endif
#ifdef SMP
args.common = NULL;
args.nthreads = num_cpu_avail(4);
if (args.nthreads == 1) {
#endif
args.n = *N;
info = GETRF_SINGLE(&args, NULL, NULL, sa, sb, 0);
if (info == 0){
args.n = *NRHS;
GETRS_N_SINGLE(&args, NULL, NULL, sa, sb, 0);
}
#ifdef SMP
} else {
args.n = *N;
info = GETRF_PARALLEL(&args, NULL, NULL, sa, sb, 0);
if (info == 0){
args.n = *NRHS;
GETRS_N_PARALLEL(&args, NULL, NULL, sa, sb, 0);
}
}
#endif
#ifndef PPC440
blas_memory_free(buffer);
#endif
*Info = info;
FUNCTION_PROFILE_END(COMPSIZE * COMPSIZE, *N * *N, 2. / 3. * *N * *N * *N + *N * *N);
IDEBUG_END;
return 0;
}