mirror of
https://github.com/amd/blis.git
synced 2026-04-19 23:28:52 +00:00
7a0ba4194f
Details: - Implemented a new feature called addons, which are similar to sandboxes except that there is no requirement to define gemm or any other particular operation. - Updated configure to accept --enable-addon=<name> or -a <name> syntax for requesting an addon be included within a BLIS build. configure now outputs the list of enabled addons into config.mk. It also outputs the corresponding #include directives for the addons' headers to a new companion to the bli_config.h header file named bli_addon.h. Because addons may wish to make use of existing BLIS types within their own definitions, the addons' headers must be included sometime after that of bli_config.h (which currently is #included before bli_type_defs.h). This is why the #include directives needed to go into a new top-level header file rather than the existing bli_config.h file. - Added a markdown document, docs/Addons.md, to explain addons, how to build with them, and what assumptions their authors should keep in mind as they create them. - Added a gemmlike-like implementation of sandwich gemm called 'gemmd' as an addon in addon/gemmd. The code uses a 'bao_' prefix for local functions, including the user-level object and typed APIs. - Updated .gitignore so that git ignores bli_addon.h files. Change-Id: Ie7efdea366481ce25075cb2459bdbcfd52309717
331 lines
10 KiB
C
331 lines
10 KiB
C
/*
|
|
|
|
BLIS
|
|
An object-based framework for developing high-performance BLAS-like
|
|
libraries.
|
|
|
|
Copyright (C) 2021, The University of Texas at Austin
|
|
|
|
Redistribution and use in source and binary forms, with or without
|
|
modification, are permitted provided that the following conditions are
|
|
met:
|
|
- Redistributions of source code must retain the above copyright
|
|
notice, this list of conditions and the following disclaimer.
|
|
- Redistributions in binary form must reproduce the above copyright
|
|
notice, this list of conditions and the following disclaimer in the
|
|
documentation and/or other materials provided with the distribution.
|
|
- Neither the name(s) of the copyright holder(s) nor the names of its
|
|
contributors may be used to endorse or promote products derived
|
|
from this software without specific prior written permission.
|
|
|
|
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
|
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
|
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
|
|
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
|
|
HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
|
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
|
|
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
|
|
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
|
|
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
|
|
*/
|
|
|
|
#include "blis.h"
|
|
|
|
#undef GENTFUNC
|
|
#define GENTFUNC( ctype, ch, opname ) \
|
|
\
|
|
void PASTECH2(bao_,ch,opname) \
|
|
( \
|
|
dim_t m, \
|
|
dim_t k, \
|
|
dim_t mr, \
|
|
cntx_t* restrict cntx, \
|
|
rntm_t* restrict rntm, \
|
|
mem_t* restrict mem, \
|
|
thrinfo_t* restrict thread \
|
|
) \
|
|
{ \
|
|
/* Set the pack buffer type so that we are obtaining memory blocks from
|
|
the pool dedicated to blocks of A. */ \
|
|
const packbuf_t pack_buf_type = BLIS_BUFFER_FOR_A_BLOCK; \
|
|
\
|
|
/* NOTE: This "rounding up" of the last upanel is absolutely necessary since
|
|
we NEED that last micropanel to have the same ldim (cs_p) as the other
|
|
micropanels. Why? Because the microkernel assumes that the register (MR,
|
|
NR) AND storage (PACKMR, PACKNR) blocksizes do not change. */ \
|
|
const dim_t m_pack = ( m / mr + ( m % mr ? 1 : 0 ) ) * mr; \
|
|
const dim_t k_pack = k; \
|
|
\
|
|
/* Barrier to make sure all threads are caught up and ready to begin the
|
|
packm stage. */ \
|
|
bli_thread_barrier( thread ); \
|
|
\
|
|
/* Compute the size of the memory block eneded. */ \
|
|
siz_t size_needed = sizeof( ctype ) * m_pack * k_pack; \
|
|
\
|
|
/* Check the mem_t entry provided by the caller. If it is unallocated,
|
|
then we need to acquire a block from the packed block allocator. */ \
|
|
if ( bli_mem_is_unalloc( mem ) ) \
|
|
{ \
|
|
if ( bli_thread_am_ochief( thread ) ) \
|
|
{ \
|
|
/* Acquire directly to the chief thread's mem_t that was passed in.
|
|
It needs to be that mem_t struct, and not a local (temporary)
|
|
mem_t, since there is no barrier until after packing is finished,
|
|
which could allow a race condition whereby the chief thread exits
|
|
the current function before the other threads have a chance to
|
|
copy from it. (A barrier would fix that race condition, but then
|
|
again, I prefer to keep barriers to a minimum.) */ \
|
|
bli_pba_acquire_m \
|
|
( \
|
|
rntm, \
|
|
size_needed, \
|
|
pack_buf_type, \
|
|
mem \
|
|
); \
|
|
} \
|
|
\
|
|
/* Broadcast the address of the chief thread's passed-in mem_t to all
|
|
threads. */ \
|
|
mem_t* mem_p = bli_thread_broadcast( thread, mem ); \
|
|
\
|
|
/* Non-chief threads: Copy the contents of the chief thread's
|
|
passed-in mem_t to the passed-in mem_t for this thread. (The
|
|
chief thread already has the mem_t, so it does not need to
|
|
perform any copy.) */ \
|
|
if ( !bli_thread_am_ochief( thread ) ) \
|
|
{ \
|
|
*mem = *mem_p; \
|
|
} \
|
|
} \
|
|
else /* if ( bli_mem_is_alloc( mem ) ) */ \
|
|
{ \
|
|
/* If the mem_t entry provided by the caller does NOT contain a NULL
|
|
buffer, then a block has already been acquired from the packed
|
|
block allocator and cached by the caller. */ \
|
|
\
|
|
/* As a sanity check, we should make sure that the mem_t object isn't
|
|
associated with a block that is too small compared to the size of
|
|
the packed matrix buffer that is needed, according to the value
|
|
computed above. */ \
|
|
siz_t mem_size = bli_mem_size( mem ); \
|
|
\
|
|
if ( mem_size < size_needed ) \
|
|
{ \
|
|
if ( bli_thread_am_ochief( thread ) ) \
|
|
{ \
|
|
/* The chief thread releases the existing block associated
|
|
with the mem_t, and then re-acquires a new block, saving
|
|
the associated mem_t to its passed-in mem_t. (See coment
|
|
above for why the acquisition needs to be directly to
|
|
the chief thread's passed-in mem_t and not a local
|
|
(temporary) mem_t. */ \
|
|
bli_pba_release \
|
|
( \
|
|
rntm, \
|
|
mem \
|
|
); \
|
|
bli_pba_acquire_m \
|
|
( \
|
|
rntm, \
|
|
size_needed, \
|
|
pack_buf_type, \
|
|
mem \
|
|
); \
|
|
} \
|
|
\
|
|
/* Broadcast the address of the chief thread's passed-in mem_t
|
|
to all threads. */ \
|
|
mem_t* mem_p = bli_thread_broadcast( thread, mem ); \
|
|
\
|
|
/* Non-chief threads: Copy the contents of the chief thread's
|
|
passed-in mem_t to the passed-in mem_t for this thread. (The
|
|
chief thread already has the mem_t, so it does not need to
|
|
perform any copy.) */ \
|
|
if ( !bli_thread_am_ochief( thread ) ) \
|
|
{ \
|
|
*mem = *mem_p; \
|
|
} \
|
|
} \
|
|
else \
|
|
{ \
|
|
/* If the mem_t entry is already allocated and sufficiently large,
|
|
then we use it as-is. No action is needed. */ \
|
|
} \
|
|
} \
|
|
}
|
|
|
|
//INSERT_GENTFUNC_BASIC0( packm_init_mem_a )
|
|
GENTFUNC( float, s, packm_init_mem_a )
|
|
GENTFUNC( double, d, packm_init_mem_a )
|
|
GENTFUNC( scomplex, c, packm_init_mem_a )
|
|
GENTFUNC( dcomplex, z, packm_init_mem_a )
|
|
|
|
|
|
#undef GENTFUNC
|
|
#define GENTFUNC( ctype, ch, opname ) \
|
|
\
|
|
void PASTECH2(bao_,ch,opname) \
|
|
( \
|
|
rntm_t* restrict rntm, \
|
|
mem_t* restrict mem, \
|
|
thrinfo_t* restrict thread \
|
|
) \
|
|
{ \
|
|
if ( thread != NULL ) \
|
|
if ( bli_thread_am_ochief( thread ) ) \
|
|
{ \
|
|
/* Check the mem_t entry provided by the caller. Only proceed if it
|
|
is allocated, which it should be. */ \
|
|
if ( bli_mem_is_alloc( mem ) ) \
|
|
{ \
|
|
bli_pba_release \
|
|
( \
|
|
rntm, \
|
|
mem \
|
|
); \
|
|
} \
|
|
} \
|
|
}
|
|
|
|
//INSERT_GENTFUNC_BASIC0( packm_finalize_mem_a )
|
|
GENTFUNC( float, s, packm_finalize_mem_a )
|
|
GENTFUNC( double, d, packm_finalize_mem_a )
|
|
GENTFUNC( scomplex, c, packm_finalize_mem_a )
|
|
GENTFUNC( dcomplex, z, packm_finalize_mem_a )
|
|
|
|
|
|
#undef GENTFUNC
|
|
#define GENTFUNC( ctype, ch, opname ) \
|
|
\
|
|
void PASTECH2(bao_,ch,opname) \
|
|
( \
|
|
pack_t* restrict schema, \
|
|
dim_t m, \
|
|
dim_t k, \
|
|
dim_t mr, \
|
|
dim_t* restrict m_max, \
|
|
dim_t* restrict k_max, \
|
|
ctype** p, inc_t* restrict rs_p, inc_t* restrict cs_p, \
|
|
dim_t* restrict pd_p, inc_t* restrict ps_p, \
|
|
mem_t* restrict mem \
|
|
) \
|
|
{ \
|
|
/* NOTE: This "rounding up" of the last upanel is absolutely necessary since
|
|
we NEED that last micropanel to have the same ldim (cs_p) as the other
|
|
micropanels. Why? Because the microkernel assumes that the register (MR,
|
|
NR) AND storage (PACKMR, PACKNR) blocksizes do not change. */ \
|
|
*m_max = ( m / mr + ( m % mr ? 1 : 0 ) ) * mr; \
|
|
*k_max = k; \
|
|
\
|
|
/* Determine the dimensions and strides for the packed matrix A. */ \
|
|
{ \
|
|
/* Pack A to column-stored row-panels. */ \
|
|
*rs_p = 1; \
|
|
*cs_p = mr; \
|
|
\
|
|
*pd_p = mr; \
|
|
*ps_p = mr * k; \
|
|
\
|
|
/* Set the schema to "packed row panels" to indicate packing to
|
|
conventional column-stored row panels. */ \
|
|
*schema = BLIS_PACKED_ROW_PANELS; \
|
|
} \
|
|
\
|
|
/* Set the buffer address provided by the caller to point to the memory
|
|
associated with the mem_t entry acquired from the memory pool. */ \
|
|
*p = bli_mem_buffer( mem ); \
|
|
}
|
|
|
|
//INSERT_GENTFUNC_BASIC0( packm_init_a )
|
|
GENTFUNC( float, s, packm_init_a )
|
|
GENTFUNC( double, d, packm_init_a )
|
|
GENTFUNC( scomplex, c, packm_init_a )
|
|
GENTFUNC( dcomplex, z, packm_init_a )
|
|
|
|
|
|
//
|
|
// Define BLAS-like interfaces to the variant chooser.
|
|
//
|
|
|
|
#undef GENTFUNC
|
|
#define GENTFUNC( ctype, ch, opname ) \
|
|
\
|
|
void PASTECH2(bao_,ch,opname) \
|
|
( \
|
|
conj_t conj, \
|
|
dim_t m_alloc, \
|
|
dim_t k_alloc, \
|
|
dim_t m, \
|
|
dim_t k, \
|
|
dim_t mr, \
|
|
ctype* restrict kappa, \
|
|
ctype* restrict d, inc_t incd, \
|
|
ctype* restrict a, inc_t rs_a, inc_t cs_a, \
|
|
ctype** restrict p, inc_t* restrict rs_p, inc_t* restrict cs_p, \
|
|
inc_t* restrict ps_p, \
|
|
cntx_t* restrict cntx, \
|
|
rntm_t* restrict rntm, \
|
|
mem_t* restrict mem, \
|
|
thrinfo_t* restrict thread \
|
|
) \
|
|
{ \
|
|
pack_t schema; \
|
|
dim_t m_max; \
|
|
dim_t k_max; \
|
|
dim_t pd_p; \
|
|
\
|
|
/* Prepare the packing destination buffer. */ \
|
|
PASTECH2(bao_,ch,packm_init_mem_a) \
|
|
( \
|
|
m_alloc, k_alloc, mr, \
|
|
cntx, \
|
|
rntm, \
|
|
mem, \
|
|
thread \
|
|
); \
|
|
\
|
|
/* Determine the packing buffer and related parameters for matrix A. */ \
|
|
PASTECH2(bao_,ch,packm_init_a) \
|
|
( \
|
|
&schema, \
|
|
m, k, mr, \
|
|
&m_max, &k_max, \
|
|
p, rs_p, cs_p, \
|
|
&pd_p, ps_p, \
|
|
mem \
|
|
); \
|
|
\
|
|
/* Pack matrix A to the destination buffer chosen above. Here, the packed
|
|
matrix is stored to column-stored MR x k micropanels. */ \
|
|
PASTECH2(bao_,ch,packm_var1) \
|
|
( \
|
|
conj, \
|
|
schema, \
|
|
m, \
|
|
k, \
|
|
m_max, \
|
|
k_max, \
|
|
kappa, \
|
|
d, incd, \
|
|
a, rs_a, cs_a, \
|
|
*p, *rs_p, *cs_p, \
|
|
pd_p, *ps_p, \
|
|
cntx, \
|
|
thread \
|
|
); \
|
|
\
|
|
/* Barrier so that packing is done before computation. */ \
|
|
bli_thread_barrier( thread ); \
|
|
}
|
|
|
|
//INSERT_GENTFUNC_BASIC0( packm_a )
|
|
GENTFUNC( float, s, packm_a )
|
|
GENTFUNC( double, d, packm_a )
|
|
GENTFUNC( scomplex, c, packm_a )
|
|
GENTFUNC( dcomplex, z, packm_a )
|
|
|