FLA_Error FLA_Syrk_ln_omp1t_var5_fc( FLA_Obj A, FLA_Obj C, int nb_alg )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj MyC;
int b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
#pragma intel omp parallel taskq
{
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
nb_alg = FLA_Obj_width( A )/omp_get_num_threads() + 1;
b = min( FLA_Obj_width( AR ), nb_alg );
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
#pragma intel omp task captureprivate(A1) private(MyC)
{
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &MyC );
FLA_Obj_set_to_zero( MyC );
/* MyC := A1 * A1' */
FLA_Syrk( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A1, FLA_ZERO, MyC );
/* C := MyC */
FLA_Axpy_sync_circular( FLA_ONE, MyC, C );
FLA_Obj_free( &MyC );
}
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
}
}
return FLA_SUCCESS;
}
FLA_Error FLA_Syrk_ln_omp1t_var5( FLA_Obj A, FLA_Obj C )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj MyC;
int b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
#pragma intel omp parallel taskq
{
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
b = FLA_Task_compute_blocksize( 0, A, AL, FLA_LEFT );
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
#pragma intel omp task captureprivate(A1) private(MyC)
{
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &MyC );
FLA_Obj_set_to_zero( MyC );
/* MyC := A1 * A1' */
FLA_Syrk_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_ONE, A1, FLA_ZERO, MyC );
/* C := MyC */
FLA_Axpy_sync_pipeline( FLA_ONE, MyC, C );
FLA_Obj_free( &MyC );
}
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
}
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_nn_omp_var15( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj C, fla_gemm_t* cntl )
{
FLA_Obj AT, A0,
AB, A1,
A2;
FLA_Obj CT, C0,
CB, C1,
C2;
FLA_Obj AL, AR, A10, A11, A12;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj C1_local;
int i, j, lock_ldim, lock_i;
int b_m, b_k;
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_TOP );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_TOP );
#pragma intel omp parallel taskq
{
while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) )
{
b_m = FLA_Determine_blocksize( A, AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x1_to_3x1( AT, &A0,
/* ** */ /* ** */
&A1,
AB, &A2, b_m, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&C1,
CB, &C2, b_m, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C1 = alpha * A1 * B + C1; */
FLA_Part_1x2( A1, &AL, &AR, 0, FLA_LEFT );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) )
{
b_k = FLA_Determine_blocksize( A, AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
// Get the index of the current partition.
// FIX THIS: need + b_m - 1 or something like this
//j = FLA_Obj_length( CT ) / b_m;
//i = FLA_Obj_width( AL ) / b_k;
//lock_ldim = FLA_get_num_threads_in_m_dim(omp_get_num_threads());
lock_i = FLA_Obj_length( CT ) / b_m;
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A10, /**/ &A11, &A12,
b_k, FLA_RIGHT );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* ** */
&B1,
BB, &B2, b_k, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* C1 = alpha * A11 * B1 + C1; */
//// FLA_Gemm( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
//// alpha, A11, B1, FLA_ONE, C1 );
#pragma intel omp task captureprivate( lock_i, A11, B1, C1 ), private( C1_local )
{
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C1, &C1_local );
FLA_Obj_set_to_zero( C1_local );
/* C1_local = alpha * A1 * B11 + C1_local; */
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A11, B1, FLA_ONE, C1_local );
// Acquire lock[i] (the lock for C1).
omp_set_lock( &fla_omp_lock[lock_i] );
/* C1 += C1_local */
FLA_Axpy_external( FLA_ONE, C1_local, C1 );
//FLA_Axpy_sync_pipeline2( j*lock_ldim, FLA_ONE, C1_local, C1 );
//FLA_Axpy_sync_circular2( j*lock_ldim, i, FLA_ONE, C1_local, C1 );
//REF_Axpy_sync_circular2( j*lock_ldim, i, FLA_ONE, C1_local, C1 );
// Release lock[i] (the lock for C1).
omp_unset_lock( &fla_omp_lock[lock_i] );
FLA_Obj_free( &C1_local );
}
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A10, A11, /**/ A12,
FLA_LEFT );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
B1,
/* ** */ /* ** */
&BB, B2, FLA_TOP );
}
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
A1,
/* ** */ /* ** */
&AB, A2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
C1,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
}
}
return FLA_SUCCESS;
}
FLA_Error FLA_Gemm_nn_omp_var5( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj C, fla_gemm_t* cntl )
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj BT, B0,
BB, B1,
B2;
FLA_Obj C_local;
int b;
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_2x1( B, &BT,
&BB, 0, FLA_TOP );
#pragma intel omp parallel taskq
{
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
b = FLA_Determine_blocksize( A, AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
//b = min( FLA_Obj_width( AR ), nb_alg );
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
b, FLA_RIGHT );
FLA_Repart_2x1_to_3x1( BT, &B0,
/* ** */ /* ** */
&B1,
BB, &B2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
#pragma intel omp task captureprivate(A1,B1) private(C_local)
{
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_local );
FLA_Obj_set_to_zero( C_local );
/* C = alpha * A1 * B1 + C; */
FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
alpha, A1, B1, FLA_ONE, C_local );
REF_Axpy_sync_circular( FLA_ONE, C_local, C );
FLA_Obj_free( &C_local );
}
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
B1,
/* ** */ /* ** */
&BB, B2, FLA_TOP );
}
}
return FLA_SUCCESS;
}
