Files
blis/frame/base/bli_pool.c
Field G. Van Zee 075143dfd9 Added support for IC loop parallelism to trsm.
Details:
- Parallelism within the IC loop (3rd loop around the microkernel) is
  now supported within the trsm operation. This is done via a new branch
  on each of the control and thread trees, which guide execution of a
  new trsm-only subproblem from within bli_trsm_blk_var1(). This trsm
  subproblem corresponds to the macrokernel computation on only the
  block of A that contains the diagonal (labeled as A11 in algorithms
  with FLAME-like partitioning), and the corresponding row panel of C.
  During the trsm subproblem, all threads within the JC communicator
  participate and parallelize along the JR loop, including any
  parallelism that was specified for the IC loop. (IR loop parallelism
  is not supported for trsm due to inter-iteration dependencies.) After
  this trsm subproblem is complete, a barrier synchronizes all
  participating threads and then they proceed to apply the prescribed
  BLIS_IC_NT (or equivalent) ways of parallelism (and any BLIS_JR_NT
  parallelism specified within) to the remaining gemm subproblem (the
  rank-k update that is performed using the newly updated row-panel of
  B). Thus, trsm now supports JC, IC, and JR loop parallelism.
- Modified bli_trsm_l_cntl_create() to create the new "prenode" branch
  of the trsm_l cntl_t tree. The trsm_r tree was left unchanged, for
  now, since it is not currently used. (All trsm problems are cast in
  terms of left-side trsm.)
- Updated bli_cntl_free_w_thrinfo() to be able to free the newly shaped
  trsm cntl_t trees. Fixed a potentially latent bug whereby a cntl_t
  subnode is only recursed upon if there existed a corresponding
  thrinfo_t node, which may not always exist (for problems too small
  to employ full parallelization due to the minimum granularity imposed
  by micropanels).
- Updated other functions in frame/base/bli_cntl.c, such as
  bli_cntl_copy() and bli_cntl_mark_family(), to recurse on sub-prenodes
  if they exist.
- Updated bli_thrinfo_free() to recurse into sub-nodes and prenodes
  when they exist, and added support for growing a prenode branch to
  bli_thrinfo_grow() via a corresponding set of help functions named
  with the _prenode() suffix.
- Added a bszid_t field thrinfo_t nodes. This field comes in handy when
  debugging the allocation/release of thrinfo_t nodes, as it helps trace
  the "identity" of each nodes as it is created/destroyed.
- Renamed
    bli_l3_thrinfo_print_paths() -> bli_l3_thrinfo_print_gemm_paths()
  and created a separate bli_l3_thrinfo_print_trsm_paths() function to
  print out the newly reconfigured thrinfo_t trees for the trsm
  operation.
- Trival changes to bli_gemm_blk_var?.c and bli_trsm_blk_var?.c
  regarding variable declarations.
- Removed subpart_t enum values BLIS_SUBPART1T, BLIS_SUBPART1B,
  BLIS_SUBPART1L, BLIS_SUBPART1R. Then added support for two new labels
  (semantically speaking): BLIS_SUBPART1A and BLIS_SUBPART1B, which
  represent the subpartition ahead of and behind, respectively,
  BLIS_SUBPART1. Updated check functions in bli_check.c accordingly.
- Shuffled layering/APIs for bli_acquire_mpart_[mn]dim() and
  bli_acquire_mpart_t2b/b2t(), _l2r/r2l().
- Deprecated old functions in frame/3/bli_l3_thrinfo.c.
2019-02-14 18:52:45 -06:00

585 lines
19 KiB
C

/*
BLIS
An object-based framework for developing high-performance BLAS-like
libraries.
Copyright (C) 2014, The University of Texas at Austin
Copyright (C) 2018, Advanced Micro Devices, Inc.
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"
//#define BLIS_ENABLE_MEM_TRACING
void bli_pool_init
(
siz_t num_blocks,
siz_t block_ptrs_len,
siz_t block_size,
siz_t align_size,
malloc_ft malloc_fp,
free_ft free_fp,
pool_t* restrict pool
)
{
// Make sure that block_ptrs_len is at least num_blocks.
block_ptrs_len = bli_max( block_ptrs_len, num_blocks );
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_init(): allocating block_ptrs (length %d): ",
( int )block_ptrs_len );
#endif
// Allocate the block_ptrs array.
// FGVZ: Do we want to call malloc_fp() for internal data structures as
// well as pool blocks? If so, don't forget to s/bli_free_intl/free_fp/g.
pblk_t* restrict block_ptrs
=
bli_malloc_intl( block_ptrs_len * sizeof( pblk_t ) );
// Allocate and initialize each entry in the block_ptrs array.
for ( dim_t i = 0; i < num_blocks; ++i )
{
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_init(): allocating block %d of size %d (align %d).\n",
( int )i, ( int )block_size, ( int )align_size );
fflush( stdout );
#endif
bli_pool_alloc_block
(
block_size,
align_size,
malloc_fp,
&(block_ptrs[i])
);
}
// NOTE: The semantics of top_index approximate a stack, where a "full"
// stack (no blocks checked out) is one where top_index == 0 and an empty
// stack (all blocks checked out) one where top_index == num_blocks.
// (Here, num_blocks tracks the number of blocks currently allocated as
// part of the pool.) This "orientation" of the stack was chosen
// intentionally, in contrast to one where top_index == -1 means the
// stack is empty and top_index = num_blocks - 1 means the stack is
// full. The chosen scheme allows one to conceptualize the stack as a
// number line in which blocks are checked out from lowest to highest,
// and additional blocks are added at the higher end.
// Initialize the pool_t structure.
bli_pool_set_block_ptrs( block_ptrs, pool );
bli_pool_set_block_ptrs_len( block_ptrs_len, pool );
bli_pool_set_top_index( 0, pool );
bli_pool_set_num_blocks( num_blocks, pool );
bli_pool_set_block_size( block_size, pool );
bli_pool_set_align_size( align_size, pool );
bli_pool_set_malloc_fp( malloc_fp, pool );
bli_pool_set_free_fp( free_fp, pool );
}
void bli_pool_finalize
(
pool_t* restrict pool
)
{
// NOTE: This implementation assumes that either:
// - all blocks have been checked in by all threads, or
// - some subset of blocks have been checked in and the caller
// is bli_pool_reinit().
// Query the block_ptrs array.
pblk_t* restrict block_ptrs = bli_pool_block_ptrs( pool );
// Query the total number of blocks currently allocated.
const siz_t num_blocks = bli_pool_num_blocks( pool );
// Query the top_index of the pool.
const siz_t top_index = bli_pool_top_index( pool );
// Sanity check: The top_index should be zero.
if ( top_index != 0 )
{
printf( "bli_pool_finalize(): final top_index == %d (expected 0); block_size: %d.\n",
( int )top_index, ( int )bli_pool_block_size( pool ) );
printf( "bli_pool_finalize(): Implication: not all blocks were checked back in!\n" );
bli_abort();
}
// Query the free() function pointer for the pool.
free_ft free_fp = bli_pool_free_fp( pool );
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_finalize(): freeing %d blocks of size %d (align %d).\n",
( int )num_blocks, ( int )bli_pool_block_size( pool ),
( int )bli_pool_align_size( pool ) );
fflush( stdout );
#endif
// Free the individual blocks currently in the pool.
for ( dim_t i = 0; i < num_blocks; ++i )
{
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_finalize(): block %d: ", ( int )i );
#endif
bli_pool_free_block( free_fp, &(block_ptrs[i]) );
}
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_finalize(): freeing block_ptrs (length %d): ",
( int )( bli_pool_block_ptrs_len( pool ) ) );
#endif
// Free the block_ptrs array.
bli_free_intl( block_ptrs );
// This explicit clearing of the pool_t struct is not strictly
// necessary and so it has been commented out.
#if 0
// Clear the contents of the pool_t struct.
bli_pool_set_block_ptrs( NULL, pool );
bli_pool_set_block_ptrs_len( 0, pool );
bli_pool_set_num_blocks( 0, pool );
bli_pool_set_top_index( 0, pool );
bli_pool_set_block_size( 0, pool );
bli_pool_set_align_size( 0, pool );
#endif
}
void bli_pool_reinit
(
siz_t num_blocks_new,
siz_t block_ptrs_len_new,
siz_t block_size_new,
siz_t align_size_new,
pool_t* restrict pool
)
{
// Preserve the pointers to malloc() and free() provided when the pool
// was first initialized.
malloc_ft malloc_fp = bli_pool_malloc_fp( pool );
free_ft free_fp = bli_pool_free_fp( pool );
// Finalize the pool as it is currently configured. If some blocks
// are still checked out to threads, those blocks are not freed
// here, and instead will be freed when the threads attempt to check
// those blocks back into the pool. (This condition can be detected
// since the block size is encoded into each pblk, which is copied
// upon checkout.)
bli_pool_finalize( pool );
// Reinitialize the pool with the new parameters, in particular,
// the new block size.
bli_pool_init
(
num_blocks_new,
block_ptrs_len_new,
block_size_new,
align_size_new,
malloc_fp,
free_fp,
pool
);
}
void bli_pool_checkout_block
(
siz_t req_size,
pblk_t* restrict block,
pool_t* restrict pool
)
{
// If the requested block size is smaller than what the pool was
// initialized with, reinitialize the pool to contain blocks of the
// requested size.
if ( bli_pool_block_size( pool ) < req_size )
{
const siz_t num_blocks_new = bli_pool_num_blocks( pool );
const siz_t block_ptrs_len_new = bli_pool_block_ptrs_len( pool );
const siz_t align_size_new = bli_pool_align_size( pool );
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_checkout_block(): old block size %d < req size %d; "
"reiniting.\n",
( int )bli_pool_block_size( pool ), ( int )req_size );
fflush( stdout );
#endif
bli_pool_reinit
(
num_blocks_new,
block_ptrs_len_new,
req_size,
align_size_new,
pool
);
}
// If the pool is exhausted, add a block.
if ( bli_pool_is_exhausted( pool ) )
{
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_checkout_block(): pool is exhausted (block size %d); "
"growing by 1.\n", ( int )bli_pool_block_size( pool ) );
fflush( stdout );
#endif
bli_pool_grow( 1, pool );
}
// At this point, at least one block is guaranteed to be available.
// Query the block_ptrs array.
pblk_t* restrict block_ptrs = bli_pool_block_ptrs( pool );
// Query the top_index of the pool.
const siz_t top_index = bli_pool_top_index( pool );
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_checkout_block(): checking out block %d of size %d "
"(align %d).\n",
( int )top_index, ( int )bli_pool_block_size( pool ),
( int )bli_pool_align_size( pool ) );
fflush( stdout );
#endif
// Copy the pblk_t at top_index to the caller's pblk_t struct.
*block = block_ptrs[ top_index ];
// Notice that we don't actually need to clear the contents of
// block_ptrs[top_index]. It will get overwritten eventually when
// the block is checked back in.
bli_pblk_clear( &block_ptrs[top_index] );
// Increment the pool's top_index.
bli_pool_set_top_index( top_index + 1, pool );
}
void bli_pool_checkin_block
(
pblk_t* restrict block,
pool_t* restrict pool
)
{
// If the pblk_t being checked in was allocated with a different block
// size than is currently in use in the pool, we simply free it and
// return. These "orphaned" blocks are no longer of use because the pool
// has since been reinitialized to a different (larger) block size.
if ( bli_pblk_block_size( block ) != bli_pool_block_size( pool ) )
{
// Query the free() function pointer for the pool.
free_ft free_fp = bli_pool_free_fp( pool );
bli_pool_free_block( free_fp, block );
return;
}
// Query the block_ptrs array.
pblk_t* restrict block_ptrs = bli_pool_block_ptrs( pool );
// Query the top_index of the pool.
const siz_t top_index = bli_pool_top_index( pool );
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_checkin_block(): checking in block %d of size %d "
"(align %d).\n",
( int )top_index - 1, ( int )bli_pool_block_size( pool ),
( int )bli_pool_align_size( pool ) );
fflush( stdout );
#endif
// Copy the caller's pblk_t struct to the block at top_index - 1.
block_ptrs[ top_index - 1 ] = *block;
// Decrement the pool's top_index.
bli_pool_set_top_index( top_index - 1, pool );
}
void bli_pool_grow
(
siz_t num_blocks_add,
pool_t* restrict pool
)
{
// If the requested increase is zero, return early.
if ( num_blocks_add == 0 ) return;
// Query the allocated length of the block_ptrs array and also the
// total number of blocks currently allocated.
const siz_t block_ptrs_len_cur = bli_pool_block_ptrs_len( pool );
const siz_t num_blocks_cur = bli_pool_num_blocks( pool );
// Compute the total number of allocated blocks that will exist
// after we grow the pool.
const siz_t num_blocks_new = num_blocks_cur + num_blocks_add;
// If adding num_blocks_add new blocks will exceed the current capacity
// of the block_ptrs array, we need to first put in place a new (larger)
// array.
if ( block_ptrs_len_cur < num_blocks_new )
{
// To prevent this from happening often, we double the current
// length of the block_ptrs array.
const siz_t block_ptrs_len_new = 2 * block_ptrs_len_cur;
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_grow(): growing block_ptrs_len (%d -> %d): ",
( int )block_ptrs_len_cur, ( int )block_ptrs_len_new );
#endif
// Query the current block_ptrs array.
pblk_t* restrict block_ptrs_cur = bli_pool_block_ptrs( pool );
// Allocate a new block_ptrs array.
// FGVZ: Do we want to call malloc_fp() for internal data structures as
// well as pool blocks? If so, don't forget to s/bli_free_intl/free_fp/g.
pblk_t* restrict block_ptrs_new
=
bli_malloc_intl( block_ptrs_len_new * sizeof( pblk_t ) );
// Query the top_index of the pool.
const siz_t top_index = bli_pool_top_index( pool );
// Copy the contents of the old block_ptrs array to the new/resized
// array. Notice that we can begin with top_index since all entries
// from 0 to top_index-1 have been (and are currently) checked out
// to threads.
for ( dim_t i = top_index; i < num_blocks_cur; ++i )
{
block_ptrs_new[i] = block_ptrs_cur[i];
}
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_grow(): freeing prev block_ptrs: " );
#endif
// Free the old block_ptrs array.
bli_free_intl( block_ptrs_cur );
// Update the pool_t struct with the new block_ptrs array and
// record its allocated length.
bli_pool_set_block_ptrs( block_ptrs_new, pool );
bli_pool_set_block_ptrs_len( block_ptrs_len_new, pool );
}
// At this point, we are guaranteed to have enough unused elements
// in the block_ptrs array to accommodate an additional num_blocks_add
// blocks.
// Query the current block_ptrs array (which was mabye just resized).
pblk_t* restrict block_ptrs = bli_pool_block_ptrs( pool );
// Query the block size and alignment size of the pool.
const siz_t block_size = bli_pool_block_size( pool );
const siz_t align_size = bli_pool_align_size( pool );
// Query the malloc() function pointer for the pool.
malloc_ft malloc_fp = bli_pool_malloc_fp( pool );
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_grow(): growing pool from (%d -> %d).\n",
( int )num_blocks_cur, ( int )num_blocks_new );
fflush( stdout );
#endif
// Allocate the requested additional blocks in the resized array.
for ( dim_t i = num_blocks_cur; i < num_blocks_new; ++i )
{
bli_pool_alloc_block
(
block_size,
align_size,
malloc_fp,
&(block_ptrs[i])
);
}
// Update the pool_t struct with the new number of allocated blocks.
// Notice that top_index remains unchanged, as do the block_size and
// align_size fields.
bli_pool_set_num_blocks( num_blocks_new, pool );
}
void bli_pool_shrink
(
siz_t num_blocks_sub,
pool_t* restrict pool
)
{
// If the requested decrease is zero, return early.
if ( num_blocks_sub == 0 ) return;
// Query the total number of blocks currently allocated.
const siz_t num_blocks = bli_pool_num_blocks( pool );
// Query the top_index of the pool.
const siz_t top_index = bli_pool_top_index( pool );
// Compute the number of blocks available to be checked out
// (and thus available for removal).
const siz_t num_blocks_avail = num_blocks - top_index;
// If the requested decrease is more than the number of available
// blocks in the pool, only remove the number of blocks actually
// available.
num_blocks_sub = bli_min( num_blocks_sub, num_blocks_avail );
// Query the block_ptrs array.
pblk_t* restrict block_ptrs = bli_pool_block_ptrs( pool );
// Compute the new total number of blocks.
const siz_t num_blocks_new = num_blocks - num_blocks_sub;
// Query the free() function pointer for the pool.
free_ft free_fp = bli_pool_free_fp( pool );
// Free the individual blocks.
for ( dim_t i = num_blocks_new; i < num_blocks; ++i )
{
bli_pool_free_block( free_fp, &(block_ptrs[i]) );
}
// Update the pool_t struct.
bli_pool_set_num_blocks( num_blocks_new, pool );
// Note that after shrinking the pool, num_blocks < block_ptrs_len.
// This means the pool can grow again by num_blocks_sub before
// a re-allocation of block_ptrs is triggered.
}
void bli_pool_alloc_block
(
siz_t block_size,
siz_t align_size,
malloc_ft malloc_fp,
pblk_t* restrict block
)
{
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_alloc_block(): calling fmalloc_align(): size %d (align %d)\n",
( int )block_size, ( int )align_size );
fflush( stdout );
#endif
// Allocate the block via the bli_fmalloc_align() wrapper, which performs
// alignment logic and opaquely saves the original pointer so that it can
// be recovered when it's time to free the block.
void* restrict buf
=
bli_fmalloc_align( malloc_fp, block_size, align_size );
#if 0
// NOTE: This code is disabled because it is not needed, since
// bli_fmalloc_align() is guaranteed to return an aligned address.
// Advance the pointer to achieve the necessary alignment, if it is not
// already aligned.
if ( bli_is_unaligned_to( ( siz_t )buf_sys, ( siz_t )align_size ) )
{
// C99's stdint.h guarantees that a void* can be safely cast to a
// uintptr_t and then back to a void*, hence the casting of buf_sys
// and align_size to uintptr_t. buf_align is initially cast to char*
// to allow pointer arithmetic in units of bytes, and then advanced
// to the next nearest alignment boundary, and finally cast back to
// void* before being stored. Notice that the arithmetic works even
// if the alignment value is not a power of two.
buf_align = ( void* )( ( char* )buf_align +
( ( uintptr_t )align_size -
( uintptr_t )buf_sys %
( uintptr_t )align_size )
);
}
#endif
// Save the results in the pblk_t structure.
bli_pblk_set_buf( buf, block );
bli_pblk_set_block_size( block_size, block );
}
void bli_pool_free_block
(
free_ft free_fp,
pblk_t* restrict block
)
{
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_pool_free_block(): calling ffree_align(): size %d.\n",
( int )bli_pblk_block_size( block ) );
fflush( stdout );
#endif
// Extract the pblk_t buffer, which is the aligned address returned from
// bli_fmalloc_align() when the block was allocated.
void* restrict buf = bli_pblk_buf( block );
// Free the block via the bli_ffree_align() wrapper, which recovers the
// original pointer that was returned by the pool's malloc() function when
// the block was allocated.
bli_ffree_align( free_fp, buf );
}
void bli_pool_print
(
pool_t* restrict pool
)
{
pblk_t* block_ptrs = bli_pool_block_ptrs( pool );
siz_t block_ptrs_len = bli_pool_block_ptrs_len( pool );
siz_t top_index = bli_pool_top_index( pool );
siz_t num_blocks = bli_pool_num_blocks( pool );
siz_t block_size = bli_pool_block_size( pool );
siz_t align_size = bli_pool_align_size( pool );
dim_t i;
printf( "pool struct ---------------\n" );
printf( " block_ptrs: %p\n", block_ptrs );
printf( " block_ptrs_len: %d\n", ( int )block_ptrs_len );
printf( " top_index: %d\n", ( int )top_index );
printf( " num_blocks: %d\n", ( int )num_blocks );
printf( " block_size: %d\n", ( int )block_size );
printf( " align_size: %d\n", ( int )align_size );
printf( " pblks sys align\n" );
for ( i = 0; i < num_blocks; ++i )
{
printf( " %d: %p\n", ( int )i, bli_pblk_buf( &block_ptrs[i] ) );
}
}
void bli_pblk_print
(
pblk_t* restrict pblk
)
{
void* buf = bli_pblk_buf( pblk );
printf( "pblk struct ---------------\n" );
printf( " block address (aligned): %p\n", buf );
}