Browse Source

prefetching for dgemm_beta

tags/v0.3.8^2
w00421467 6 years ago
parent
commit
3ccf8885ac
1 changed files with 14 additions and 9 deletions
  1. +14
    -9
      kernel/arm64/dgemm_beta.S

+ 14
- 9
kernel/arm64/dgemm_beta.S View File

@@ -43,7 +43,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define betaV0 v11.d[0]
#define I x16

#define size 128
#define prfm_size 640
#define calc_size 128

/**************************************************************************************
* Macro definitions
@@ -119,27 +120,31 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ldp q2, q3, [A02]
ldp q4, q5, [A03]
ldp q6, q7, [A04]
fmul v0.2d, v0.2d, betaV0
fmul v1.2d, v1.2d, betaV0
fmul v2.2d, v2.2d, betaV0
fmul v3.2d, v3.2d, betaV0

prfm PLDL1KEEP, [A01, prfm_size]

fmul v4.2d, v4.2d, betaV0
fmul v5.2d, v5.2d, betaV0

prfm PLDL1KEEP, [A03, prfm_size]

fmul v6.2d, v6.2d, betaV0
fmul v7.2d, v7.2d, betaV0

st1 {v0.2d, v1.2d}, [A01]
add A01, A01, size
add A01, A01, calc_size
st1 {v2.2d, v3.2d}, [A02]
add A02, A02, size
add A02, A02, calc_size
st1 {v4.2d, v5.2d}, [A03]
add A03, A03, size
add A03, A03, calc_size
st1 {v6.2d, v7.2d}, [A04]
add A04, A04, size
add A04, A04, calc_size

subs I , I , #1
bne .Lgemm_beta_03


Loading…
Cancel
Save