void bli_hemv_blk_var1( conj_t conjh,
obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y,
cntx_t* cntx,
hemv_t* cntl )
{
obj_t a11, a11_pack;
obj_t a10;
obj_t x1, x1_pack;
obj_t x0;
obj_t y1, y1_pack;
obj_t y0;
dim_t mn;
dim_t ij;
dim_t b_alg;
// Even though this blocked algorithm is expressed only in terms of the
// lower triangular case, the upper triangular case is still supported:
// when bli_acquire_mpart_tl2br() is passed a matrix that is stored in
// in the upper triangle, and the requested subpartition resides in the
// lower triangle (as is the case for this algorithm), the routine fills
// the request as if the caller had actually requested the corresponding
// "mirror" subpartition in the upper triangle, except that it marks the
// subpartition for transposition (and conjugation).
// Initialize objects for packing.
bli_obj_init_pack( &a11_pack );
bli_obj_init_pack( &x1_pack );
bli_obj_init_pack( &y1_pack );
// Query dimension.
mn = bli_obj_length( a );
// y = beta * y;
bli_scalv_int( beta,
y,
cntx, bli_cntl_sub_scalv( cntl ) );
// Partition diagonally.
for ( ij = 0; ij < mn; ij += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( ij, mn, a,
bli_cntl_bszid( cntl ), cntx );
// Acquire partitions for A11, A10, x1, x0, y1, and y0.
bli_acquire_mpart_tl2br( BLIS_SUBPART11,
ij, b_alg, a, &a11 );
bli_acquire_mpart_tl2br( BLIS_SUBPART10,
ij, b_alg, a, &a10 );
bli_acquire_vpart_f2b( BLIS_SUBPART1,
ij, b_alg, x, &x1 );
bli_acquire_vpart_f2b( BLIS_SUBPART0,
ij, b_alg, x, &x0 );
bli_acquire_vpart_f2b( BLIS_SUBPART1,
ij, b_alg, y, &y1 );
bli_acquire_vpart_f2b( BLIS_SUBPART0,
ij, b_alg, y, &y0 );
// Initialize objects for packing A11, x1, and y1 (if needed).
bli_packm_init( &a11, &a11_pack,
cntx, bli_cntl_sub_packm_a11( cntl ) );
bli_packv_init( &x1, &x1_pack,
cntx, bli_cntl_sub_packv_x1( cntl ) );
bli_packv_init( &y1, &y1_pack,
cntx, bli_cntl_sub_packv_y1( cntl ) );
// Copy/pack A11, x1, y1 (if needed).
bli_packm_int( &a11, &a11_pack,
cntx, bli_cntl_sub_packm_a11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
bli_packv_int( &x1, &x1_pack,
cntx, bli_cntl_sub_packv_x1( cntl ) );
bli_packv_int( &y1, &y1_pack,
cntx, bli_cntl_sub_packv_y1( cntl ) );
// y0 = y0 + alpha * A10' * x1;
bli_gemv_int( bli_apply_conj( conjh, BLIS_TRANSPOSE ),
BLIS_NO_CONJUGATE,
alpha,
&a10,
&x1_pack,
&BLIS_ONE,
&y0,
cntx,
bli_cntl_sub_gemv_t_rp( cntl ) );
// y1 = y1 + alpha * A11 * x1;
bli_hemv_int( conjh,
alpha,
&a11_pack,
&x1_pack,
&BLIS_ONE,
&y1_pack,
cntx,
bli_cntl_sub_hemv( cntl ) );
// y1 = y1 + alpha * A10 * x0;
bli_gemv_int( BLIS_NO_TRANSPOSE,
BLIS_NO_CONJUGATE,
alpha,
&a10,
&x0,
&BLIS_ONE,
&y1_pack,
cntx,
bli_cntl_sub_gemv_n_rp( cntl ) );
// Copy/unpack y1 (if y1 was packed).
bli_unpackv_int( &y1_pack, &y1,
cntx, bli_cntl_sub_unpackv_y1( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_packm_release( &a11_pack, bli_cntl_sub_packm_a11( cntl ) );
bli_packv_release( &x1_pack, bli_cntl_sub_packv_x1( cntl ) );
bli_packv_release( &y1_pack, bli_cntl_sub_packv_y1( cntl ) );
}
void libblis_test_dotaxpyv_experiment( test_params_t* params,
test_op_t* op,
iface_t iface,
num_t datatype,
char* pc_str,
char* sc_str,
unsigned int p_cur,
double* perf,
double* resid )
{
unsigned int n_repeats = params->n_repeats;
unsigned int i;
double time_min = 1e9;
double time;
dim_t m;
conj_t conjxt, conjx, conjy;
conj_t conjconjxty;
obj_t alpha, xt, x, y, rho, z;
obj_t z_save;
// Map the dimension specifier to an actual dimension.
m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_conj( pc_str[0], &conjxt );
bli_param_map_char_to_blis_conj( pc_str[1], &conjx );
bli_param_map_char_to_blis_conj( pc_str[2], &conjy );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &alpha );
bli_obj_scalar_init_detached( datatype, &rho );
// Create test operands (vectors and/or matrices).
libblis_test_vobj_create( params, datatype, sc_str[0], m, &x );
libblis_test_vobj_create( params, datatype, sc_str[1], m, &y );
libblis_test_vobj_create( params, datatype, sc_str[2], m, &z );
libblis_test_vobj_create( params, datatype, sc_str[2], m, &z_save );
// Set alpha.
if ( bli_obj_is_real( z ) )
{
bli_setsc( -0.8, 0.0, &alpha );
}
else
{
bli_setsc( 0.0, -0.8, &alpha );
}
// Randomize x and z, and save z.
bli_randv( &x );
bli_randv( &z );
bli_copyv( &z, &z_save );
// Create an alias to x for xt. (Note that it doesn't actually need to be
// transposed.)
bli_obj_alias_to( x, xt );
// Determine whether to make a copy of x with or without conjugation.
//
// conjx conjy ~conjx^conjy y is initialized as
// n n c y = conj(x)
// n c n y = x
// c n n y = x
// c c c y = conj(x)
//
conjconjxty = bli_apply_conj( conjxt, conjy );
conjconjxty = bli_conj_toggled( conjconjxty );
bli_obj_set_conj( conjconjxty, xt );
bli_copyv( &xt, &y );
// Apply the parameters.
bli_obj_set_conj( conjxt, xt );
bli_obj_set_conj( conjx, x );
bli_obj_set_conj( conjy, y );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copysc( &BLIS_MINUS_ONE, &rho );
bli_copyv( &z_save, &z );
time = bli_clock();
libblis_test_dotaxpyv_impl( iface, &alpha, &xt, &x, &y, &rho, &z );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m + 2.0 * m ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( z ) ) *perf *= 4.0;
// Perform checks.
libblis_test_dotaxpyv_check( &alpha, &xt, &x, &y, &rho, &z, &z_save, resid );
// Zero out performance and residual if output vector is empty.
libblis_test_check_empty_problem( &z, perf, resid );
// Free the test objects.
bli_obj_free( &x );
bli_obj_free( &y );
bli_obj_free( &z );
bli_obj_free( &z_save );
}
