void bli_her2_int( conj_t conjh,
obj_t* alpha,
obj_t* alpha_conj,
obj_t* x,
obj_t* y,
obj_t* c,
cntx_t* cntx,
her2_t* cntl )
{
varnum_t n;
impl_t i;
FUNCPTR_T f;
obj_t alpha_local;
obj_t alpha_conj_local;
obj_t x_local;
obj_t y_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
{
if ( bli_is_conj( conjh ) ) bli_her2_check( alpha, x, y, c );
else bli_syr2_check( alpha, x, y, c );
}
// If C, x, or y has a zero dimension, return early.
if ( bli_obj_has_zero_dim( c ) ) return;
if ( bli_obj_has_zero_dim( x ) ) return;
if ( bli_obj_has_zero_dim( y ) ) return;
// Alias the operands in case we need to apply conjugations.
bli_obj_alias_to( x, &x_local );
bli_obj_alias_to( y, &y_local );
bli_obj_alias_to( c, &c_local );
// If matrix C is marked for conjugation, we interpret this as a request
// to apply a conjugation to the other operands.
if ( bli_obj_has_conj( &c_local ) )
{
bli_obj_toggle_conj( &c_local );
bli_obj_toggle_conj( &x_local );
bli_obj_toggle_conj( &y_local );
bli_obj_scalar_init_detached_copy_of( bli_obj_dt( alpha ),
BLIS_CONJUGATE,
alpha,
&alpha_local );
bli_obj_scalar_init_detached_copy_of( bli_obj_dt( alpha_conj ),
BLIS_CONJUGATE,
alpha_conj,
&alpha_conj_local );
}
else
{
bli_obj_alias_to( *alpha, alpha_local );
bli_obj_alias_to( *alpha_conj, alpha_conj_local );
}
// Extract the variant number and implementation type.
n = bli_cntl_var_num( cntl );
i = bli_cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[n][i];
// Invoke the variant.
f( conjh,
&alpha_local,
&alpha_conj_local,
&x_local,
&y_local,
&c_local,
cntx,
cntl );
}
void bli_trsv_int( obj_t* alpha,
obj_t* a,
obj_t* x,
cntx_t* cntx,
trsv_t* cntl )
{
varnum_t n;
impl_t i;
bool_t uplo;
FUNCPTR_T f;
obj_t a_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_trsv_check( alpha, a, x );
// If A or x has a zero dimension, return early.
if ( bli_obj_has_zero_dim( a ) ) return;
if ( bli_obj_has_zero_dim( x ) ) return;
// Alias A in case we need to induce a transformation (ie: transposition).
bli_obj_alias_to( a, &a_local );
// NOTE: to support cases where B is complex and A is real, we will
// need to have the default side case be BLIS_RIGHT and then express
// the left case in terms of it, rather than the other way around.
// Determine uplo (for indexing to the correct function pointer).
if ( bli_obj_is_lower( &a_local ) ) uplo = 0;
else uplo = 1;
// We do not explicitly implement the cases where A is transposed.
// However, we can still handle them. Specifically, if A is marked as
// needing a transposition, we simply toggle the uplo value to cause the
// correct algorithm to be induced. When that algorithm partitions into
// A, it will grab the correct subpartitions, which will inherit A's
// transposition bit and thus downstream subproblems will do the right
// thing. Alternatively, we could accomplish the same end goal by
// inducing a transposition, via bli_obj_induce_trans(), in the code
// block below. That macro function swaps dimensions, strides, and
// offsets. As an example, given a lower triangular, column-major matrix
// that needs a transpose, we would induce that transposition by recasting
// the object as an upper triangular, row-major matrix (with no transpose
// needed). Note that how we choose to handle transposition here does NOT
// affect the optimal choice of kernel (ie: a column-major column panel
// matrix with transpose times a vector would use the same kernel as a
// row-major row panel matrix with no transpose times a vector).
if ( bli_obj_has_trans( &a_local ) )
{
//bli_obj_induce_trans( &a_local );
//bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, &a_local );
if ( uplo == 1 ) uplo = 0;
else uplo = 1;
}
// Extract the variant number and implementation type.
n = bli_cntl_var_num( cntl );
i = bli_cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[uplo][n][i];
// Invoke the variant.
f( alpha,
&a_local,
x,
cntx,
cntl );
}
void bli_unpackv_int( obj_t* p,
obj_t* a,
cntx_t* cntx,
unpackv_t* cntl )
{
// The unpackv operation consists of an optional casting post-process.
// (This post-process is analogous to the cast pre-process in packv.)
// Here are the following possible ways unpackv can execute:
// 1. unpack and cast: Unpack to a temporary vector c and then cast
// c to a.
// 2. unpack only: Unpack directly to vector a since typecasting is
// not needed.
// 3. cast only: Not yet supported / not used.
// 4. no-op: The control tree directs us to skip the unpack operation
// entirely. No action is taken.
obj_t c;
varnum_t n;
impl_t i;
FUNCPTR_T f;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_unpackv_check( p, a, cntx );
// Sanity check; A should never have a zero dimension. If we must support
// it, then we should fold it into the next alias-and-early-exit block.
if ( bli_obj_has_zero_dim( *a ) ) bli_abort();
// First check if we are to skip this operation because the control tree
// is NULL, and if so, simply return.
if ( bli_cntl_is_noop( cntl ) )
{
return;
}
// If p was aliased to a during the pack stage (because it was already
// in an acceptable packed/contiguous format), then no unpack is actually
// necessary, so we return.
if ( bli_obj_is_alias_of( *p, *a ) )
{
return;
}
// Now, if we are not skipping the unpack operation, then the only
// question left is whether we are to typecast vector a after unpacking.
if ( bli_obj_datatype( *p ) != bli_obj_datatype( *a ) )
bli_abort();
/*
if ( bli_obj_datatype( *p ) != bli_obj_datatype( *a ) )
{
// Initialize an object c for the intermediate typecast vector.
bli_unpackv_init_cast( p,
a,
&c );
}
else
*/
{
// If no cast is needed, then aliasing object c to the original
// vector serves as a minor optimization. This causes the unpackv
// implementation to unpack directly into vector a.
bli_obj_alias_to( *a, c );
}
// Now we are ready to proceed with the unpacking.
// Extract the variant number and implementation type.
n = bli_cntl_var_num( cntl );
i = bli_cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[n][i];
// Invoke the variant.
f( p,
&c,
cntx,
cntl );
// Now, if necessary, we cast the contents of c to vector a. If casting
// was not necessary, then we are done because the call to the unpackv
// implementation would have unpacked directly to vector a.
/*
if ( bli_obj_datatype( *p ) != bli_obj_datatype( *a ) )
{
// Copy/typecast vector c to vector a.
// NOTE: Here, we use copynzv instead of copym because, in the cases
// where we are unpacking/typecasting a real vector c to a complex
// vector a, we want to touch only the real components of a, rather
// than also set the imaginary components to zero. This comes about
// because of the fact that, if we are unpacking real-to-complex,
// then it is because all of the computation occurred in the real
// domain, and so we would want to leave whatever imaginary values
// there are in vector a untouched. Notice that for unpackings that
// entail complex-to-complex data movements, the copynzv operation
// behaves exactly as copym, so no use cases are lost (at least none
// that I can think of).
bli_copynzv( &c,
a );
// NOTE: The above code/comment is outdated. What should happen is
// as follows:
// - If dt(a) is complex and dt(p) is real, then create an alias of
// a and then tweak it so that it looks like a real domain object.
// This will involve:
// - projecting the datatype to real domain
// - scaling both the row and column strides by 2
// ALL OF THIS should be done in the front-end, NOT here, as
// unpackv() won't even be needed in that case.
}
*/
}
void bli_hemv_int( conj_t conjh,
obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y,
cntx_t* cntx,
hemv_t* cntl )
{
varnum_t n;
impl_t i;
FUNCPTR_T f;
obj_t a_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
{
if ( bli_is_conj( conjh ) ) bli_hemv_check( alpha, a, x, beta, y );
else bli_symv_check( alpha, a, x, beta, y );
}
// If y has a zero dimension, return early.
if ( bli_obj_has_zero_dim( *y ) ) return;
// If x has a zero dimension, scale y by beta and return early.
if ( bli_obj_has_zero_dim( *x ) )
{
bli_scalm( beta, y );
return;
}
// Alias A in case we need to induce the upper triangular case.
bli_obj_alias_to( *a, a_local );
/*
// Our blocked algorithms only [explicitly] implement the lower triangular
// case, so if matrix A is stored as upper triangular, we must toggle the
// transposition (and conjugation) bits so that the diagonal partitioning
// routines grab the correct partitions corresponding to the upper
// triangular case. But we only need to do this for blocked algorithms,
// since unblocked algorithms are responsible for handling the upper case
// explicitly (and they should not be inspecting the transposition bit anyway).
if ( bli_cntl_is_blocked( cntl ) && bli_obj_is_upper( *a ) )
{
bli_obj_toggle_conj( a_local );
bli_obj_toggle_trans( a_local );
}
*/
// Extract the variant number and implementation type.
n = bli_cntl_var_num( cntl );
i = bli_cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[n][i];
// Invoke the variant.
f( conjh,
alpha,
&a_local,
x,
beta,
y,
cntx,
cntl );
}
void bli_ger_int( conj_t conjx,
conj_t conjy,
obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* a,
cntx_t* cntx,
ger_t* cntl )
{
varnum_t n;
impl_t i;
FUNCPTR_T f;
obj_t alpha_local;
obj_t x_local;
obj_t y_local;
obj_t a_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_ger_check( alpha, x, y, a );
// If A has a zero dimension, return early.
if ( bli_obj_has_zero_dim( a ) ) return;
// If x or y has a zero dimension, return early.
if ( bli_obj_has_zero_dim( x ) ||
bli_obj_has_zero_dim( y ) ) return;
// Alias the objects, applying conjx and conjy to x and y, respectively.
bli_obj_alias_with_conj( conjx, x, &x_local );
bli_obj_alias_with_conj( conjy, y, &y_local );
bli_obj_alias_to( a, &a_local );
// If matrix A is marked for conjugation, we interpret this as a request
// to apply a conjugation to the other operands.
if ( bli_obj_has_conj( &a_local ) )
{
bli_obj_toggle_conj( &a_local );
bli_obj_toggle_conj( &x_local );
bli_obj_toggle_conj( &y_local );
bli_obj_scalar_init_detached_copy_of( bli_obj_dt( alpha ),
BLIS_CONJUGATE,
alpha,
&alpha_local );
}
else
{
bli_obj_alias_to( *alpha, alpha_local );
}
// If we are about the call a leaf-level implementation, and matrix A
// still needs a transposition, then we must induce one by swapping the
// strides and dimensions.
if ( bli_cntl_is_leaf( cntl ) && bli_obj_has_trans( &a_local ) )
{
bli_obj_induce_trans( &a_local );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, &a_local );
}
// Extract the variant number and implementation type.
n = bli_cntl_var_num( cntl );
i = bli_cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[n][i];
// Invoke the variant.
f( &alpha_local,
&x_local,
&y_local,
&a_local,
cntx,
cntl );
}
void bli_gemm_int( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
cntx_t* cntx,
gemm_t* cntl,
thrinfo_t* thread )
{
obj_t a_local;
obj_t b_local;
obj_t c_local;
varnum_t n;
impl_t i;
FUNCPTR_T f;
ind_t im;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_gemm_basic_check( alpha, a, b, beta, c, cntx );
// If C has a zero dimension, return early.
if ( bli_obj_has_zero_dim( *c ) ) return;
// If A or B has a zero dimension, scale C by beta and return early.
if ( bli_obj_has_zero_dim( *a ) ||
bli_obj_has_zero_dim( *b ) )
{
if( bli_thread_am_ochief( thread ) )
bli_scalm( beta, c );
bli_thread_obarrier( thread );
return;
}
// If A or B is marked as being filled with zeros, scale C by beta and
// return early.
if ( bli_obj_is_zeros( *a ) ||
bli_obj_is_zeros( *b ) )
{
if( bli_thread_am_ochief( thread ) )
bli_scalm( beta, c );
bli_thread_obarrier( thread );
return;
}
// Alias A and B in case we need to update attached scalars.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
// Alias C in case we need to induce a transposition.
bli_obj_alias_to( *c, c_local );
// If we are about to call a leaf-level implementation, and matrix C
// still needs a transposition, then we must induce one by swapping the
// strides and dimensions. Note that this transposition would normally
// be handled explicitly in the packing of C, but if C is not being
// packed, this is our last chance to handle the transposition.
if ( bli_cntl_is_leaf( cntl ) && bli_obj_has_trans( *c ) )
{
//if( bli_thread_am_ochief( thread ) ) {
bli_obj_induce_trans( c_local );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, c_local );
// }
}
// If alpha is non-unit, typecast and apply it to the scalar attached
// to B.
if ( !bli_obj_equals( alpha, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( alpha, &b_local );
}
// If beta is non-unit, typecast and apply it to the scalar attached
// to C.
if ( !bli_obj_equals( beta, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( beta, &c_local );
}
// Extract the variant number and implementation type.
n = bli_cntl_var_num( cntl );
i = bli_cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[n][i];
// Somewhat hackish support for 3m3, 3m2, and 4m1b method implementations.
im = bli_cntx_get_ind_method( cntx );
if ( im != BLIS_NAT )
{
if ( im == BLIS_3M3 && f == bli_gemm_blk_var1f ) f = bli_gemm_blk_var4f;
else if ( im == BLIS_3M2 && f == bli_gemm_ker_var2 ) f = bli_gemm_ker_var4;
else if ( im == BLIS_4M1B && f == bli_gemm_ker_var2 ) f = bli_gemm_ker_var3;
}
// Invoke the variant.
f( &a_local,
&b_local,
&c_local,
cntx,
cntl,
thread );
}
void bli_packm_int( obj_t* a,
obj_t* p,
cntx_t* cntx,
packm_t* cntl,
thrinfo_t* thread )
{
varnum_t n;
impl_t i;
FUNCPTR_T f;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_packm_int_check( a, p, cntx );
// Sanity check; A should never have a zero dimension. If we must support
// it, then we should fold it into the next alias-and-early-exit block.
//if ( bli_obj_has_zero_dim( *a ) ) bli_abort();
// First check if we are to skip this operation because the control tree
// is NULL. We return without taking any action because a was already
// aliased to p in packm_init().
if ( bli_cntl_is_noop( cntl ) )
{
return;
}
// Let us now check to see if the object has already been packed. First
// we check if it has been packed to an unspecified (row or column)
// format, in which case we can return, since by now aliasing has already
// taken place in packm_init().
// NOTE: The reason we don't need to even look at the control tree in
// this case is as follows: an object's pack status is only set to
// BLIS_PACKED_UNSPEC for situations when the actual format used is
// not important, as long as its packed into contiguous rows or
// contiguous columns. A good example of this is packing for matrix
// operands in the level-2 operations.
if ( bli_obj_pack_schema( *a ) == BLIS_PACKED_UNSPEC )
{
return;
}
// At this point, we can be assured that cntl is not NULL. Now we check
// if the object has already been packed to the desired schema (as en-
// coded in the control tree). If so, we can return, as above.
// NOTE: In most cases, an object's pack status will be BLIS_NOT_PACKED
// and thus packing will be called for (but in some cases packing has
// already taken place, or does not need to take place, and so that will
// be indicated by the pack status). Also, not all combinations of
// current pack status and desired pack schema are valid.
if ( bli_obj_pack_schema( *a ) == cntl_pack_schema( cntl ) )
{
return;
}
// If the object is marked as being filled with zeros, then we can skip
// the packm operation entirely.
if ( bli_obj_is_zeros( *a ) )
{
return;
}
// Extract the variant number and implementation type.
n = bli_cntl_var_num( cntl );
i = bli_cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[n][i];
// Invoke the variant with kappa_use.
f( a,
p,
cntx,
thread );
// Barrier so that packing is done before computation
bli_thread_obarrier( thread );
}