template<class T, paradigm P, const bool& b> inline T
Matrix<T,P,b>::dotc (const Matrix<T,P>& M) const {
return DOTC (*this, M);
}
CAMLprim value LFUN(gemm_trace_stub)(
value vTRANSA,
value vTRANSB,
value vN, value vK,
value vAR, value vAC, value vA,
value vBR, value vBC, value vB)
{
CAMLparam2(vA, vB);
integer GET_INT(N), GET_INT(K);
char GET_INT(TRANSA), GET_INT(TRANSB);
MAT_PARAMS(A);
MAT_PARAMS(B);
unsigned long iter_incr_A, iter_incr_B;
integer dot_incr_A, dot_incr_B;
NUMBER res = NUMBER_ZERO;
NUMBER *last_A;
caml_enter_blocking_section(); /* Allow other threads */
if (TRANSA == 'N') {
if (TRANSB == 'N') {
if (N < K) {
iter_incr_A = 1;
dot_incr_A = rows_A;
iter_incr_B = rows_B;
dot_incr_B = 1;
} else {
NUMBER *tmp_A_data = A_data;
integer tmp_N = N;
A_data = B_data;
B_data = tmp_A_data;
N = K;
K = tmp_N;
iter_incr_A = 1;
dot_incr_A = rows_B;
iter_incr_B = rows_A;
dot_incr_B = 1;
}
} else {
if (N == rows_A && N == rows_B) {
integer NK = N * K;
res =
#ifdef LACAML_COMPLEX
(TRANSB == 'C')
? DOTC(&NK, B_data, &integer_one, A_data, &integer_one)
:
#endif
DOTU(&NK, B_data, &integer_one, A_data, &integer_one);
goto end;
} else {
integer tmp_N = N;
NUMBER *tmp_A_data = A_data;
A_data = B_data;
B_data = tmp_A_data;
N = K;
K = tmp_N;
iter_incr_A = rows_B;
dot_incr_A = 1;
iter_incr_B = rows_A;
dot_incr_B = 1;
}
}
} else {
if (TRANSB == 'N') {
if (K == rows_A && K == rows_B) {
integer NK = N * K;
res =
#ifdef LACAML_COMPLEX
(TRANSA == 'C')
? DOTC(&NK, A_data, &integer_one, B_data, &integer_one)
:
#endif
DOTU(&NK, A_data, &integer_one, B_data, &integer_one);
goto end;
} else {
iter_incr_A = rows_A;
dot_incr_A = 1;
iter_incr_B = rows_B;
dot_incr_B = 1;
}
} else {
if (N < K) {
NUMBER *tmp_A_data = A_data;
integer tmp_N = N;
A_data = B_data;
B_data = tmp_A_data;
N = K;
K = tmp_N;
iter_incr_A = rows_B;
dot_incr_A = 1;
iter_incr_B = 1;
dot_incr_B = rows_A;
} else {
iter_incr_A = rows_A;
dot_incr_A = 1;
iter_incr_B = 1;
dot_incr_B = rows_B;
}
}
}
last_A = A_data + N * iter_incr_A;
#ifndef LACAML_COMPLEX /* Real number */
COMMON_GEMM_TRACE_LOOP
#else /* Complex number */
if (TRANSA == 'C')
if (TRANSB == 'C')
while (A_data != last_A) {
NUMBER cd = DOTU(&K, A_data, &dot_incr_A, B_data, &dot_incr_B);
NUMBER d = COMLEX_CONJ(cd);
res = ADD_NUMBER(res, d);
GEMM_TRACE_INCR;
}
else
while (A_data != last_A) {
NUMBER d = DOTC(&K, A_data, &dot_incr_A, B_data, &dot_incr_B);
res = ADD_NUMBER(res, d);
GEMM_TRACE_INCR;
}
else if (TRANSB == 'C')
while (A_data != last_A) {
NUMBER d = DOTC(&K, B_data, &dot_incr_B, A_data, &dot_incr_A);
res = ADD_NUMBER(res, d);
GEMM_TRACE_INCR;
}
else COMMON_GEMM_TRACE_LOOP
#endif
end:
caml_leave_blocking_section(); /* Disallow other threads */
CAMLreturn(COPY_NUMBER(res));
}