amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-05-11 17:50:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

Bunch of minor fixes

Browse Source 
Removed barrier after unpackm in all level3 blocked variants
Now there is an implicit barrier inside unpackm that only occurs if C is packed (which is usually not the case)

Moved the enabling of the tree barriers into bli_config.h
Fed the default MR and NR for double precision into bli_get_range instead of the number 8
This commit is contained in:
Tyler Smith
2014-04-04 15:09:10 -05:00
parent 575fb9b0b0
commit 5ec93bd9a7
30 changed files with 114 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
config/mic/bli_config.h
View File

@@ -36,6 +36,9 @@
#define BLIS_CONFIG_H
#define BLIS_TREE_BARRIER
#define BLIS_TREE_BARRIER_ARITY 4
// -- OPERATING SYSTEM ---------------------------------------------------------

33
frame/1m/packm/bli_packm_unb_var1.c
View File

@@ -56,7 +56,8 @@ static FUNCPTR_T GENARRAY(ftypes,packm_unb_var1);
void bli_packm_unb_var1( obj_t* c,
obj_t* p )
obj_t* p,
packm_thrinfo_t* thread )
{
num_t dt_cp = bli_obj_datatype( *c );
@@ -98,20 +99,22 @@ void bli_packm_unb_var1( obj_t* c,
// function pointer.
f = ftypes[dt_cp];
// Invoke the function.
f( strucc,
diagoffc,
diagc,
uploc,
transc,
densify,
m_p,
n_p,
m_max_p,
n_max_p,
buf_kappa,
buf_c, rs_c, cs_c,
buf_p, rs_p, cs_p );
if( thread_am_ochief( thread ) ) {
// Invoke the function.
f( strucc,
diagoffc,
diagc,
uploc,
transc,
densify,
m_p,
n_p,
m_max_p,
n_max_p,
buf_kappa,
buf_c, rs_c, cs_c,
buf_p, rs_p, cs_p );
}
}

3
frame/1m/packm/bli_packm_unb_var1.h
View File

@@ -33,7 +33,8 @@
*/
void bli_packm_unb_var1( obj_t* c,
obj_t* p );
obj_t* p,
packm_thrinfo_t* thread );
#undef GENTPROT

12
frame/1m/unpackm/bli_unpackm_int.c
View File

@@ -49,7 +49,8 @@ static FUNCPTR_T vars[2][3] =
void bli_unpackm_int( obj_t* p,
obj_t* a,
unpackm_t* cntl )
unpackm_t* cntl,
packm_thrinfo_t* thread )
{
// The unpackm operation consists of an optional post-process: castm.
// (This post-process is analogous to the castm pre-process in packm.)
@@ -122,9 +123,12 @@ void bli_unpackm_int( obj_t* p,
f = vars[n][i];
// Invoke the variant.
f( p,
&c,
cntl );
if( thread_am_ochief( thread ) ) {
f( p,
&c,
cntl );
}
thread_obarrier( thread );
// Now, if necessary, we cast the contents of c to matrix a. If casting
// was not necessary, then we are done because the call to the unpackm

3
frame/1m/unpackm/bli_unpackm_int.h
View File

@@ -34,7 +34,8 @@
void bli_unpackm_int( obj_t* p,
obj_t* a,
unpackm_t* cntl );
unpackm_t* cntl,
packm_thrinfo_t* thread );
/*
void bli_unpackm_init_cast( obj_t* p,

3
frame/2/ger/bli_ger_blk_var1.c
View File

@@ -91,7 +91,8 @@ void bli_ger_blk_var1( obj_t* alpha,
// Copy/unpack A1 (if A1 was packed).
bli_unpackm_int( &a1_pack, &a1,
cntl_sub_unpackm_a( cntl ) );
cntl_sub_unpackm_a( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back

3
frame/2/ger/bli_ger_blk_var2.c
View File

@@ -91,7 +91,8 @@ void bli_ger_blk_var2( obj_t* alpha,
// Copy/unpack A1 (if A1 was packed).
bli_unpackm_int( &a1_pack, &a1,
cntl_sub_unpackm_a( cntl ) );
cntl_sub_unpackm_a( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back

3
frame/2/her/bli_her_blk_var1.c
View File

@@ -113,7 +113,8 @@ void bli_her_blk_var1( conj_t conjh,
// Copy/unpack C11 (if C11 was packed).
bli_unpackm_int( &c11_pack, &c11,
cntl_sub_unpackm_c11( cntl ) );
cntl_sub_unpackm_c11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back

3
frame/2/her/bli_her_blk_var2.c
View File

@@ -113,7 +113,8 @@ void bli_her_blk_var2( conj_t conjh,
// Copy/unpack C11 (if C11 was packed).
bli_unpackm_int( &c11_pack, &c11,
cntl_sub_unpackm_c11( cntl ) );
cntl_sub_unpackm_c11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back

3
frame/2/her2/bli_her2_blk_var1.c
View File

@@ -137,7 +137,8 @@ void bli_her2_blk_var1( conj_t conjh,
// Copy/unpack C11 (if C11 was packed).
bli_unpackm_int( &c11_pack, &c11,
cntl_sub_unpackm_c11( cntl ) );
cntl_sub_unpackm_c11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back

3
frame/2/her2/bli_her2_blk_var2.c
View File

@@ -140,7 +140,8 @@ void bli_her2_blk_var2( conj_t conjh,
// Copy/unpack C11 (if C11 was packed).
bli_unpackm_int( &c11_pack, &c11,
cntl_sub_unpackm_c11( cntl ) );
cntl_sub_unpackm_c11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back

3
frame/2/her2/bli_her2_blk_var3.c
View File

@@ -140,7 +140,8 @@ void bli_her2_blk_var3( conj_t conjh,
// Copy/unpack C11 (if C11 was packed).
bli_unpackm_int( &c11_pack, &c11,
cntl_sub_unpackm_c11( cntl ) );
cntl_sub_unpackm_c11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back

3
frame/2/her2/bli_her2_blk_var4.c
View File

@@ -137,7 +137,8 @@ void bli_her2_blk_var4( conj_t conjh,
// Copy/unpack C11 (if C11 was packed).
bli_unpackm_int( &c11_pack, &c11,
cntl_sub_unpackm_c11( cntl ) );
cntl_sub_unpackm_c11( cntl ),
&BLIS_PACKM_SINGLE_THREADED );
}
// If any packing buffers were acquired within packm, release them back

10
frame/3/gemm/bli_gemm_blk_var1f.c
View File

@@ -131,13 +131,9 @@ void bli_gemm_blk_var1f( obj_t* a,
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
if( thread_am_ichief( thread ) ) {
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
//Barrier to make sure unpacking is done before next iteration's packing of C
//Somehow, we'd like to make this a noop if packing isn't done.
thread_ibarrier( thread );
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
gemm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back

10
frame/3/gemm/bli_gemm_blk_var2f.c
View File

@@ -130,13 +130,9 @@ void bli_gemm_blk_var2f( obj_t* a,
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
if( thread_am_ichief( thread ) ) {
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
//Barrier to make sure unpacking is done before next iteration's packing of C
//Somehow, we'd like to make this a noop if packing isn't done.
thread_ibarrier( thread );
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
gemm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back

10
frame/3/gemm/bli_gemm_blk_var3f.c
View File

@@ -139,14 +139,14 @@ void bli_gemm_blk_var3f( obj_t* a,
thread_obarrier( thread );
// Unpack C (if C was packed).
if( thread_am_ochief( thread ) ){
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ) );
bli_obj_release_pack( c_pack );
}
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
gemm_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) )
bli_obj_release_pack( c_pack );
if( thread_am_ichief( thread ) ){
bli_obj_release_pack( a1_pack );
bli_obj_release_pack( b1_pack );

9
frame/3/herk/bli_herk_blk_var1f.c
View File

@@ -126,12 +126,9 @@ void bli_herk_blk_var1f( obj_t* a,
herk_thread_sub_herk( thread ) );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
if( thread_am_ichief( thread ) ) {
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
thread_ibarrier( thread );
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
herk_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back

8
frame/3/herk/bli_herk_blk_var2f.c
View File

@@ -142,11 +142,9 @@ void bli_herk_blk_var2f( obj_t* a,
herk_thread_sub_herk( thread ) );
// Unpack C1 (if C1 was packed).
if( thread_am_ichief( thread ) ) {
bli_unpackm_int( c1S_pack, &c1S,
cntl_sub_unpackm_c( cntl ) );
}
thread_ibarrier( thread );
bli_unpackm_int( c1S_pack, &c1S,
cntl_sub_unpackm_c( cntl ),
herk_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back

11
frame/3/herk/bli_herk_blk_var3f.c
View File

@@ -136,14 +136,15 @@ void bli_herk_blk_var3f( obj_t* a,
thread_obarrier( thread );
// Unpack C (if C was packed).
if( thread_am_ochief( thread ) ) {
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ) );
bli_obj_release_pack( c_pack );
}
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
herk_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) ) {
bli_obj_release_pack( c_pack );
}
if( thread_am_ichief( thread ) ) {
bli_obj_release_pack( a1_pack );
bli_obj_release_pack( ah1_pack );

11
frame/3/trmm/bli_trmm_blk_var1f.c
View File

@@ -137,14 +137,9 @@ void bli_trmm_blk_var1f( obj_t* a,
trmm_thread_sub_trmm( thread ) );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
if( thread_am_ichief( thread ) ) {
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
//Barrier to make sure unpacking is done before next iteration's packing of C
//Somehow, we'd like to make this a noop if packing isn't done.
thread_ibarrier( thread );
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back

13
frame/3/trmm/bli_trmm_blk_var2b.c
View File

@@ -83,7 +83,7 @@ void bli_trmm_blk_var2b( obj_t* a,
n_trans = bli_obj_width_after_trans( *b );
dim_t start, end;
//bli_get_range( thread, 0, n_trans, 8, &start, &end );
bli_get_range_weighted( thread, 0, n_trans, BLIS_DEFAULT_NC_D, 0, &start, &end );
bli_get_range_weighted( thread, 0, n_trans, BLIS_DEFAULT_NC_D, bli_obj_is_upper( *c ), &start, &end );
// Partition along the n dimension.
for ( i = start; i < end; i += b_alg )
@@ -127,14 +127,9 @@ void bli_trmm_blk_var2b( obj_t* a,
trmm_thread_sub_trmm( thread ) );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
if( thread_am_ichief( thread ) ) {
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
//Barrier to make sure unpacking is done before next iteration's packing of C
//Somehow, we'd like to make this a noop if packing isn't done.
thread_ibarrier( thread );
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back

13
frame/3/trmm/bli_trmm_blk_var2f.c
View File

@@ -83,7 +83,7 @@ void bli_trmm_blk_var2f( obj_t* a,
n_trans = bli_obj_width_after_trans( *b );
dim_t start, end;
//bli_get_range( thread, 0, n_trans, 8, &start, &end );
bli_get_range_weighted( thread, 0, n_trans, BLIS_DEFAULT_NC_D, 1, &start, &end );
bli_get_range_weighted( thread, 0, n_trans, BLIS_DEFAULT_NC_D, bli_obj_is_lower( *c ), &start, &end );
// Partition along the n dimension.
for ( i = start; i < end; i += b_alg )
@@ -127,14 +127,9 @@ void bli_trmm_blk_var2f( obj_t* a,
trmm_thread_sub_trmm( thread ) );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
if( thread_am_ichief( thread ) ) {
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
//Barrier to make sure unpacking is done before next iteration's packing of C
//Somehow, we'd like to make this a noop if packing isn't done.
thread_ibarrier( thread );
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back

11
frame/3/trmm/bli_trmm_blk_var3b.c
View File

@@ -126,14 +126,15 @@ void bli_trmm_blk_var3b( obj_t* a,
thread_obarrier( thread );
// Unpack C (if C was packed).
if( thread_am_ochief( thread ) ){
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ) );
bli_obj_release_pack( c_pack );
}
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) ){
bli_obj_release_pack( c_pack );
}
if( thread_am_ichief( thread ) ){
bli_obj_release_pack( a1_pack );
bli_obj_release_pack( b1_pack );

11
frame/3/trmm/bli_trmm_blk_var3f.c
View File

@@ -126,14 +126,15 @@ void bli_trmm_blk_var3f( obj_t* a,
thread_obarrier( thread );
// Unpack C (if C was packed).
if( thread_am_ochief( thread ) ){
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ) );
bli_obj_release_pack( c_pack );
}
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) ){
bli_obj_release_pack( c_pack );
}
if( thread_am_ichief( thread ) ){
bli_obj_release_pack( a1_pack );
bli_obj_release_pack( b1_pack );

11
frame/3/trsm/bli_trsm_blk_var2b.c
View File

@@ -127,14 +127,9 @@ void bli_trsm_blk_var2b( obj_t* a,
trsm_thread_sub_trsm( thread ) );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
if( thread_am_ichief( thread ) ) {
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
//Barrier to make sure unpacking is done before next iteration's packing of C
//Somehow, we'd like to make this a noop if packing isn't done.
thread_ibarrier( thread );
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
trsm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back

11
frame/3/trsm/bli_trsm_blk_var2f.c
View File

@@ -128,14 +128,9 @@ void bli_trsm_blk_var2f( obj_t* a,
trsm_thread_sub_trsm( thread ) );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
if( thread_am_ichief( thread ) ) {
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
//Barrier to make sure unpacking is done before next iteration's packing of C
//Somehow, we'd like to make this a noop if packing isn't done.
thread_ibarrier( thread );
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
trsm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back

11
frame/3/trsm/bli_trsm_blk_var3b.c
View File

@@ -137,14 +137,15 @@ void bli_trsm_blk_var3b( obj_t* a,
thread_obarrier( thread );
// Unpack C (if C was packed).
if( thread_am_ochief( thread ) ) {
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ) );
bli_obj_release_pack( c_pack );
}
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
trsm_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) ) {
bli_obj_release_pack( c_pack );
}
if( thread_am_ichief( thread ) ) {
bli_obj_release_pack( a1_pack );
bli_obj_release_pack( b1_pack );

11
frame/3/trsm/bli_trsm_blk_var3f.c
View File

@@ -137,14 +137,15 @@ void bli_trsm_blk_var3f( obj_t* a,
thread_obarrier( thread );
// Unpack C (if C was packed).
if( thread_am_ochief( thread ) ) {
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ) );
bli_obj_release_pack( c_pack );
}
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
trsm_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) ) {
bli_obj_release_pack( c_pack );
}
if( thread_am_ichief( thread ) ) {
bli_obj_release_pack( a1_pack );
bli_obj_release_pack( b1_pack );

2
frame/base/bli_threading.c
View File

@@ -35,7 +35,7 @@
#include "blis.h"
#ifdef BLIS_TREE_BARRIER
barrier_t* bli_free_barrier_tree( barrier_t* barrier )
void bli_free_barrier_tree( barrier_t* barrier )
{
if( barrier == NULL )
return;

3
frame/base/bli_threading.h
View File

@@ -34,9 +34,6 @@
#ifndef BLIS_THREADING_H
#define BLIS_THREADING_H
#define BLIS_TREE_BARRIER
#define BLIS_TREE_BARRIER_ARITY 4
#ifdef BLIS_TREE_BARRIER
struct barrier_s
{