floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

optimized dtrmm kernel for ARMV7

This commit is contained in:
wernsaar 2013-11-28 12:32:12 +01:00
parent 274304bd03
commit dec7ad0dfd
1 changed files with 13 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
kernel/arm/dtrmm_kernel_4x2_vfp.S
View File

@ -26,7 +26,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/ *****************************************************************************/
/************************************************************************************** /**************************************************************************************
* 2013/11/23 Saar * 2013/11/28 Saar
* BLASTEST : OK * BLASTEST : OK
* CTEST : OK * CTEST : OK
* TEST : OK * TEST : OK
@ -106,25 +106,26 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
.macro KERNEL4x2_SUB .macro KERNEL4x2_SUB
fldd d4 , [ BO ] fldd d4 , [ BO ]
fldd d5 , [ BO, #8 ]
fldd d0 , [ AO ] fldd d0 , [ AO ]
fldd d1 , [ AO, #8 ] fldd d1 , [ AO, #8 ]
fldd d2 , [ AO, #16 ] pld [ AO , #A_PRE ]
fldd d3 , [ AO, #24 ]
fmacd d8 , d0, d4 fmacd d8 , d0, d4
fldd d2 , [ AO, #16 ]
fmacd d9 , d1, d4 fmacd d9 , d1, d4
fldd d3 , [ AO, #24 ]
fmacd d10 , d2, d4 fmacd d10 , d2, d4
fldd d5 , [ BO, #8 ]
fmacd d11 , d3, d4 fmacd d11 , d3, d4
fmacd d12 , d0, d5 fmacd d12 , d0, d5
fmacd d13 , d1, d5 fmacd d13 , d1, d5
add AO , AO, #32
fmacd d14 , d2, d5 fmacd d14 , d2, d5
add BO , BO, #16
fmacd d15 , d3, d5 fmacd d15 , d3, d5
add AO , AO, #32
add BO , BO, #16
.endm .endm
@ -490,13 +491,18 @@ _L2_M4_20:
.align 5 .align 5
_L2_M4_22: _L2_M4_22:
pld [ BO , #B_PRE ]
KERNEL4x2_SUB KERNEL4x2_SUB
KERNEL4x2_SUB KERNEL4x2_SUB
pld [ BO , #B_PRE ]
KERNEL4x2_SUB KERNEL4x2_SUB
KERNEL4x2_SUB KERNEL4x2_SUB
pld [ BO , #B_PRE ]
KERNEL4x2_SUB KERNEL4x2_SUB
KERNEL4x2_SUB KERNEL4x2_SUB
pld [ BO , #B_PRE ]
KERNEL4x2_SUB KERNEL4x2_SUB
KERNEL4x2_SUB KERNEL4x2_SUB