Cortex-A57: Fix clang compilation errors
This commit is contained in:
Executable → Regular
+192
-192
@@ -161,25 +161,25 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v3.4s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmul v16.4s, v0.4s, v8.2s[0]
|
||||
fmul v17.4s, v1.4s, v8.2s[0]
|
||||
fmul v18.4s, v2.4s, v8.2s[0]
|
||||
fmul v19.4s, v3.4s, v8.2s[0]
|
||||
fmul v16.4s, v0.4s, v8.s[0]
|
||||
fmul v17.4s, v1.4s, v8.s[0]
|
||||
fmul v18.4s, v2.4s, v8.s[0]
|
||||
fmul v19.4s, v3.4s, v8.s[0]
|
||||
|
||||
fmul v20.4s, v0.4s, v8.2s[1]
|
||||
fmul v21.4s, v1.4s, v8.2s[1]
|
||||
fmul v22.4s, v2.4s, v8.2s[1]
|
||||
fmul v23.4s, v3.4s, v8.2s[1]
|
||||
fmul v20.4s, v0.4s, v8.s[1]
|
||||
fmul v21.4s, v1.4s, v8.s[1]
|
||||
fmul v22.4s, v2.4s, v8.s[1]
|
||||
fmul v23.4s, v3.4s, v8.s[1]
|
||||
|
||||
fmul v24.4s, v0.4s, v9.2s[0]
|
||||
fmul v25.4s, v1.4s, v9.2s[0]
|
||||
fmul v26.4s, v2.4s, v9.2s[0]
|
||||
fmul v27.4s, v3.4s, v9.2s[0]
|
||||
fmul v24.4s, v0.4s, v9.s[0]
|
||||
fmul v25.4s, v1.4s, v9.s[0]
|
||||
fmul v26.4s, v2.4s, v9.s[0]
|
||||
fmul v27.4s, v3.4s, v9.s[0]
|
||||
|
||||
fmul v28.4s, v0.4s, v9.2s[1]
|
||||
fmul v29.4s, v1.4s, v9.2s[1]
|
||||
fmul v30.4s, v2.4s, v9.2s[1]
|
||||
fmul v31.4s, v3.4s, v9.2s[1]
|
||||
fmul v28.4s, v0.4s, v9.s[1]
|
||||
fmul v29.4s, v1.4s, v9.s[1]
|
||||
fmul v30.4s, v2.4s, v9.s[1]
|
||||
fmul v31.4s, v3.4s, v9.s[1]
|
||||
|
||||
ld1 {v12.2s, v13.2s}, [pB]
|
||||
add pB, pB, #16
|
||||
@@ -194,25 +194,25 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
.endm
|
||||
|
||||
.macro KERNEL16x4_M1
|
||||
fmla v16.4s, v0.4s, v8.2s[0]
|
||||
fmla v17.4s, v1.4s, v8.2s[0]
|
||||
fmla v18.4s, v2.4s, v8.2s[0]
|
||||
fmla v19.4s, v3.4s, v8.2s[0]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v17.4s, v1.4s, v8.s[0]
|
||||
fmla v18.4s, v2.4s, v8.s[0]
|
||||
fmla v19.4s, v3.4s, v8.s[0]
|
||||
|
||||
fmla v20.4s, v0.4s, v8.2s[1]
|
||||
fmla v21.4s, v1.4s, v8.2s[1]
|
||||
fmla v22.4s, v2.4s, v8.2s[1]
|
||||
fmla v23.4s, v3.4s, v8.2s[1]
|
||||
fmla v20.4s, v0.4s, v8.s[1]
|
||||
fmla v21.4s, v1.4s, v8.s[1]
|
||||
fmla v22.4s, v2.4s, v8.s[1]
|
||||
fmla v23.4s, v3.4s, v8.s[1]
|
||||
|
||||
fmla v24.4s, v0.4s, v9.2s[0]
|
||||
fmla v25.4s, v1.4s, v9.2s[0]
|
||||
fmla v26.4s, v2.4s, v9.2s[0]
|
||||
fmla v27.4s, v3.4s, v9.2s[0]
|
||||
fmla v24.4s, v0.4s, v9.s[0]
|
||||
fmla v25.4s, v1.4s, v9.s[0]
|
||||
fmla v26.4s, v2.4s, v9.s[0]
|
||||
fmla v27.4s, v3.4s, v9.s[0]
|
||||
|
||||
fmla v28.4s, v0.4s, v9.2s[1]
|
||||
fmla v29.4s, v1.4s, v9.2s[1]
|
||||
fmla v30.4s, v2.4s, v9.2s[1]
|
||||
fmla v31.4s, v3.4s, v9.2s[1]
|
||||
fmla v28.4s, v0.4s, v9.s[1]
|
||||
fmla v29.4s, v1.4s, v9.s[1]
|
||||
fmla v30.4s, v2.4s, v9.s[1]
|
||||
fmla v31.4s, v3.4s, v9.s[1]
|
||||
|
||||
ld1 {v12.2s, v13.2s}, [pB]
|
||||
add pB, pB, #16
|
||||
@@ -227,25 +227,25 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
.endm
|
||||
|
||||
.macro KERNEL16x4_M2
|
||||
fmla v16.4s, v4.4s, v12.2s[0]
|
||||
fmla v17.4s, v5.4s, v12.2s[0]
|
||||
fmla v18.4s, v6.4s, v12.2s[0]
|
||||
fmla v19.4s, v7.4s, v12.2s[0]
|
||||
fmla v16.4s, v4.4s, v12.s[0]
|
||||
fmla v17.4s, v5.4s, v12.s[0]
|
||||
fmla v18.4s, v6.4s, v12.s[0]
|
||||
fmla v19.4s, v7.4s, v12.s[0]
|
||||
|
||||
fmla v20.4s, v4.4s, v12.2s[1]
|
||||
fmla v21.4s, v5.4s, v12.2s[1]
|
||||
fmla v22.4s, v6.4s, v12.2s[1]
|
||||
fmla v23.4s, v7.4s, v12.2s[1]
|
||||
fmla v20.4s, v4.4s, v12.s[1]
|
||||
fmla v21.4s, v5.4s, v12.s[1]
|
||||
fmla v22.4s, v6.4s, v12.s[1]
|
||||
fmla v23.4s, v7.4s, v12.s[1]
|
||||
|
||||
fmla v24.4s, v4.4s, v13.2s[0]
|
||||
fmla v25.4s, v5.4s, v13.2s[0]
|
||||
fmla v26.4s, v6.4s, v13.2s[0]
|
||||
fmla v27.4s, v7.4s, v13.2s[0]
|
||||
fmla v24.4s, v4.4s, v13.s[0]
|
||||
fmla v25.4s, v5.4s, v13.s[0]
|
||||
fmla v26.4s, v6.4s, v13.s[0]
|
||||
fmla v27.4s, v7.4s, v13.s[0]
|
||||
|
||||
fmla v28.4s, v4.4s, v13.2s[1]
|
||||
fmla v29.4s, v5.4s, v13.2s[1]
|
||||
fmla v30.4s, v6.4s, v13.2s[1]
|
||||
fmla v31.4s, v7.4s, v13.2s[1]
|
||||
fmla v28.4s, v4.4s, v13.s[1]
|
||||
fmla v29.4s, v5.4s, v13.s[1]
|
||||
fmla v30.4s, v6.4s, v13.s[1]
|
||||
fmla v31.4s, v7.4s, v13.s[1]
|
||||
|
||||
ld1 {v8.2s, v9.2s}, [pB]
|
||||
add pB, pB, #16
|
||||
@@ -260,25 +260,25 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
.endm
|
||||
|
||||
.macro KERNEL16x4_E
|
||||
fmla v16.4s, v4.4s, v12.2s[0]
|
||||
fmla v17.4s, v5.4s, v12.2s[0]
|
||||
fmla v18.4s, v6.4s, v12.2s[0]
|
||||
fmla v19.4s, v7.4s, v12.2s[0]
|
||||
fmla v16.4s, v4.4s, v12.s[0]
|
||||
fmla v17.4s, v5.4s, v12.s[0]
|
||||
fmla v18.4s, v6.4s, v12.s[0]
|
||||
fmla v19.4s, v7.4s, v12.s[0]
|
||||
|
||||
fmla v20.4s, v4.4s, v12.2s[1]
|
||||
fmla v21.4s, v5.4s, v12.2s[1]
|
||||
fmla v22.4s, v6.4s, v12.2s[1]
|
||||
fmla v23.4s, v7.4s, v12.2s[1]
|
||||
fmla v20.4s, v4.4s, v12.s[1]
|
||||
fmla v21.4s, v5.4s, v12.s[1]
|
||||
fmla v22.4s, v6.4s, v12.s[1]
|
||||
fmla v23.4s, v7.4s, v12.s[1]
|
||||
|
||||
fmla v24.4s, v4.4s, v13.2s[0]
|
||||
fmla v25.4s, v5.4s, v13.2s[0]
|
||||
fmla v26.4s, v6.4s, v13.2s[0]
|
||||
fmla v27.4s, v7.4s, v13.2s[0]
|
||||
fmla v24.4s, v4.4s, v13.s[0]
|
||||
fmla v25.4s, v5.4s, v13.s[0]
|
||||
fmla v26.4s, v6.4s, v13.s[0]
|
||||
fmla v27.4s, v7.4s, v13.s[0]
|
||||
|
||||
fmla v28.4s, v4.4s, v13.2s[1]
|
||||
fmla v29.4s, v5.4s, v13.2s[1]
|
||||
fmla v30.4s, v6.4s, v13.2s[1]
|
||||
fmla v31.4s, v7.4s, v13.2s[1]
|
||||
fmla v28.4s, v4.4s, v13.s[1]
|
||||
fmla v29.4s, v5.4s, v13.s[1]
|
||||
fmla v30.4s, v6.4s, v13.s[1]
|
||||
fmla v31.4s, v7.4s, v13.s[1]
|
||||
.endm
|
||||
|
||||
.macro KERNEL16x4_SUB
|
||||
@@ -293,25 +293,25 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v3.4s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v16.4s, v0.4s, v8.2s[0]
|
||||
fmla v17.4s, v1.4s, v8.2s[0]
|
||||
fmla v18.4s, v2.4s, v8.2s[0]
|
||||
fmla v19.4s, v3.4s, v8.2s[0]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v17.4s, v1.4s, v8.s[0]
|
||||
fmla v18.4s, v2.4s, v8.s[0]
|
||||
fmla v19.4s, v3.4s, v8.s[0]
|
||||
|
||||
fmla v20.4s, v0.4s, v8.2s[1]
|
||||
fmla v21.4s, v1.4s, v8.2s[1]
|
||||
fmla v22.4s, v2.4s, v8.2s[1]
|
||||
fmla v23.4s, v3.4s, v8.2s[1]
|
||||
fmla v20.4s, v0.4s, v8.s[1]
|
||||
fmla v21.4s, v1.4s, v8.s[1]
|
||||
fmla v22.4s, v2.4s, v8.s[1]
|
||||
fmla v23.4s, v3.4s, v8.s[1]
|
||||
|
||||
fmla v24.4s, v0.4s, v9.2s[0]
|
||||
fmla v25.4s, v1.4s, v9.2s[0]
|
||||
fmla v26.4s, v2.4s, v9.2s[0]
|
||||
fmla v27.4s, v3.4s, v9.2s[0]
|
||||
fmla v24.4s, v0.4s, v9.s[0]
|
||||
fmla v25.4s, v1.4s, v9.s[0]
|
||||
fmla v26.4s, v2.4s, v9.s[0]
|
||||
fmla v27.4s, v3.4s, v9.s[0]
|
||||
|
||||
fmla v28.4s, v0.4s, v9.2s[1]
|
||||
fmla v29.4s, v1.4s, v9.2s[1]
|
||||
fmla v30.4s, v2.4s, v9.2s[1]
|
||||
fmla v31.4s, v3.4s, v9.2s[1]
|
||||
fmla v28.4s, v0.4s, v9.s[1]
|
||||
fmla v29.4s, v1.4s, v9.s[1]
|
||||
fmla v30.4s, v2.4s, v9.s[1]
|
||||
fmla v31.4s, v3.4s, v9.s[1]
|
||||
.endm
|
||||
|
||||
.macro SAVE16x4
|
||||
@@ -369,14 +369,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v1.4s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmul v16.4s, v0.4s, v8.2s[0]
|
||||
fmul v17.4s, v1.4s, v8.2s[0]
|
||||
fmul v20.4s, v0.4s, v8.2s[1]
|
||||
fmul v21.4s, v1.4s, v8.2s[1]
|
||||
fmul v24.4s, v0.4s, v9.2s[0]
|
||||
fmul v25.4s, v1.4s, v9.2s[0]
|
||||
fmul v28.4s, v0.4s, v9.2s[1]
|
||||
fmul v29.4s, v1.4s, v9.2s[1]
|
||||
fmul v16.4s, v0.4s, v8.s[0]
|
||||
fmul v17.4s, v1.4s, v8.s[0]
|
||||
fmul v20.4s, v0.4s, v8.s[1]
|
||||
fmul v21.4s, v1.4s, v8.s[1]
|
||||
fmul v24.4s, v0.4s, v9.s[0]
|
||||
fmul v25.4s, v1.4s, v9.s[0]
|
||||
fmul v28.4s, v0.4s, v9.s[1]
|
||||
fmul v29.4s, v1.4s, v9.s[1]
|
||||
|
||||
ld1 {v12.2s, v13.2s}, [pB]
|
||||
add pB, pB, #16
|
||||
@@ -387,14 +387,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
.endm
|
||||
|
||||
.macro KERNEL8x4_M1
|
||||
fmla v16.4s, v0.4s, v8.2s[0]
|
||||
fmla v17.4s, v1.4s, v8.2s[0]
|
||||
fmla v20.4s, v0.4s, v8.2s[1]
|
||||
fmla v21.4s, v1.4s, v8.2s[1]
|
||||
fmla v24.4s, v0.4s, v9.2s[0]
|
||||
fmla v25.4s, v1.4s, v9.2s[0]
|
||||
fmla v28.4s, v0.4s, v9.2s[1]
|
||||
fmla v29.4s, v1.4s, v9.2s[1]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v17.4s, v1.4s, v8.s[0]
|
||||
fmla v20.4s, v0.4s, v8.s[1]
|
||||
fmla v21.4s, v1.4s, v8.s[1]
|
||||
fmla v24.4s, v0.4s, v9.s[0]
|
||||
fmla v25.4s, v1.4s, v9.s[0]
|
||||
fmla v28.4s, v0.4s, v9.s[1]
|
||||
fmla v29.4s, v1.4s, v9.s[1]
|
||||
|
||||
ld1 {v12.2s, v13.2s}, [pB]
|
||||
add pB, pB, #16
|
||||
@@ -405,14 +405,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
.endm
|
||||
|
||||
.macro KERNEL8x4_M2
|
||||
fmla v16.4s, v4.4s, v12.2s[0]
|
||||
fmla v17.4s, v5.4s, v12.2s[0]
|
||||
fmla v20.4s, v4.4s, v12.2s[1]
|
||||
fmla v21.4s, v5.4s, v12.2s[1]
|
||||
fmla v24.4s, v4.4s, v13.2s[0]
|
||||
fmla v25.4s, v5.4s, v13.2s[0]
|
||||
fmla v28.4s, v4.4s, v13.2s[1]
|
||||
fmla v29.4s, v5.4s, v13.2s[1]
|
||||
fmla v16.4s, v4.4s, v12.s[0]
|
||||
fmla v17.4s, v5.4s, v12.s[0]
|
||||
fmla v20.4s, v4.4s, v12.s[1]
|
||||
fmla v21.4s, v5.4s, v12.s[1]
|
||||
fmla v24.4s, v4.4s, v13.s[0]
|
||||
fmla v25.4s, v5.4s, v13.s[0]
|
||||
fmla v28.4s, v4.4s, v13.s[1]
|
||||
fmla v29.4s, v5.4s, v13.s[1]
|
||||
|
||||
ld1 {v8.2s, v9.2s}, [pB]
|
||||
add pB, pB, #16
|
||||
@@ -423,14 +423,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
.endm
|
||||
|
||||
.macro KERNEL8x4_E
|
||||
fmla v16.4s, v4.4s, v12.2s[0]
|
||||
fmla v17.4s, v5.4s, v12.2s[0]
|
||||
fmla v20.4s, v4.4s, v12.2s[1]
|
||||
fmla v21.4s, v5.4s, v12.2s[1]
|
||||
fmla v24.4s, v4.4s, v13.2s[0]
|
||||
fmla v25.4s, v5.4s, v13.2s[0]
|
||||
fmla v28.4s, v4.4s, v13.2s[1]
|
||||
fmla v29.4s, v5.4s, v13.2s[1]
|
||||
fmla v16.4s, v4.4s, v12.s[0]
|
||||
fmla v17.4s, v5.4s, v12.s[0]
|
||||
fmla v20.4s, v4.4s, v12.s[1]
|
||||
fmla v21.4s, v5.4s, v12.s[1]
|
||||
fmla v24.4s, v4.4s, v13.s[0]
|
||||
fmla v25.4s, v5.4s, v13.s[0]
|
||||
fmla v28.4s, v4.4s, v13.s[1]
|
||||
fmla v29.4s, v5.4s, v13.s[1]
|
||||
.endm
|
||||
|
||||
.macro KERNEL8x4_SUB
|
||||
@@ -441,14 +441,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v1.4s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v16.4s, v0.4s, v8.2s[0]
|
||||
fmla v17.4s, v1.4s, v8.2s[0]
|
||||
fmla v20.4s, v0.4s, v8.2s[1]
|
||||
fmla v21.4s, v1.4s, v8.2s[1]
|
||||
fmla v24.4s, v0.4s, v9.2s[0]
|
||||
fmla v25.4s, v1.4s, v9.2s[0]
|
||||
fmla v28.4s, v0.4s, v9.2s[1]
|
||||
fmla v29.4s, v1.4s, v9.2s[1]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v17.4s, v1.4s, v8.s[0]
|
||||
fmla v20.4s, v0.4s, v8.s[1]
|
||||
fmla v21.4s, v1.4s, v8.s[1]
|
||||
fmla v24.4s, v0.4s, v9.s[0]
|
||||
fmla v25.4s, v1.4s, v9.s[0]
|
||||
fmla v28.4s, v0.4s, v9.s[1]
|
||||
fmla v29.4s, v1.4s, v9.s[1]
|
||||
.endm
|
||||
|
||||
.macro SAVE8x4
|
||||
@@ -496,17 +496,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s, v1.2s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmul v16.2s, v0.2s, v8.2s[0]
|
||||
fmul v29.2s, v1.2s, v9.2s[1]
|
||||
fmul v16.2s, v0.2s, v8.s[0]
|
||||
fmul v29.2s, v1.2s, v9.s[1]
|
||||
|
||||
fmul v20.2s, v0.2s, v8.2s[1]
|
||||
fmul v25.2s, v1.2s, v9.2s[0]
|
||||
fmul v20.2s, v0.2s, v8.s[1]
|
||||
fmul v25.2s, v1.2s, v9.s[0]
|
||||
|
||||
fmul v24.2s, v0.2s, v9.2s[0]
|
||||
fmul v21.2s, v1.2s, v8.2s[1]
|
||||
fmul v24.2s, v0.2s, v9.s[0]
|
||||
fmul v21.2s, v1.2s, v8.s[1]
|
||||
|
||||
fmul v28.2s, v0.2s, v9.2s[1]
|
||||
fmul v17.2s, v1.2s, v8.2s[0]
|
||||
fmul v28.2s, v0.2s, v9.s[1]
|
||||
fmul v17.2s, v1.2s, v8.s[0]
|
||||
|
||||
ld1 {v12.2s, v13.2s}, [pB]
|
||||
add pB, pB, #16
|
||||
@@ -515,61 +515,61 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
.endm
|
||||
|
||||
.macro KERNEL4x4_M1
|
||||
fmla v16.2s, v0.2s, v8.2s[0]
|
||||
fmla v29.2s, v1.2s, v9.2s[1]
|
||||
fmla v16.2s, v0.2s, v8.s[0]
|
||||
fmla v29.2s, v1.2s, v9.s[1]
|
||||
|
||||
ld1 {v12.2s, v13.2s}, [pB] // For next round
|
||||
add pB, pB, #16
|
||||
|
||||
fmla v20.2s, v0.2s, v8.2s[1]
|
||||
fmla v25.2s, v1.2s, v9.2s[0]
|
||||
fmla v20.2s, v0.2s, v8.s[1]
|
||||
fmla v25.2s, v1.2s, v9.s[0]
|
||||
|
||||
ld1 {v4.2s, v5.2s}, [pA] // For next round
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v24.2s, v0.2s, v9.2s[0]
|
||||
fmla v21.2s, v1.2s, v8.2s[1]
|
||||
fmla v24.2s, v0.2s, v9.s[0]
|
||||
fmla v21.2s, v1.2s, v8.s[1]
|
||||
|
||||
prfm PLDL1KEEP, [pB, #512]
|
||||
|
||||
fmla v28.2s, v0.2s, v9.2s[1]
|
||||
fmla v17.2s, v1.2s, v8.2s[0]
|
||||
fmla v28.2s, v0.2s, v9.s[1]
|
||||
fmla v17.2s, v1.2s, v8.s[0]
|
||||
.endm
|
||||
|
||||
.macro KERNEL4x4_M2
|
||||
fmla v16.2s, v4.2s, v12.2s[0]
|
||||
fmla v29.2s, v5.2s, v13.2s[1]
|
||||
fmla v16.2s, v4.2s, v12.s[0]
|
||||
fmla v29.2s, v5.2s, v13.s[1]
|
||||
|
||||
ld1 {v8.2s, v9.2s}, [pB] // For next round
|
||||
add pB, pB, #16
|
||||
|
||||
fmla v20.2s, v4.2s, v12.2s[1]
|
||||
fmla v25.2s, v5.2s, v13.2s[0]
|
||||
fmla v20.2s, v4.2s, v12.s[1]
|
||||
fmla v25.2s, v5.2s, v13.s[0]
|
||||
|
||||
ld1 {v0.2s, v1.2s}, [pA] // For next round
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v24.2s, v4.2s, v13.2s[0]
|
||||
fmla v21.2s, v5.2s, v12.2s[1]
|
||||
fmla v24.2s, v4.2s, v13.s[0]
|
||||
fmla v21.2s, v5.2s, v12.s[1]
|
||||
|
||||
prfm PLDL1KEEP, [pA, #512]
|
||||
|
||||
fmla v28.2s, v4.2s, v13.2s[1]
|
||||
fmla v17.2s, v5.2s, v12.2s[0]
|
||||
fmla v28.2s, v4.2s, v13.s[1]
|
||||
fmla v17.2s, v5.2s, v12.s[0]
|
||||
.endm
|
||||
|
||||
.macro KERNEL4x4_E
|
||||
fmla v16.2s, v4.2s, v12.2s[0]
|
||||
fmla v29.2s, v5.2s, v13.2s[1]
|
||||
fmla v16.2s, v4.2s, v12.s[0]
|
||||
fmla v29.2s, v5.2s, v13.s[1]
|
||||
|
||||
fmla v20.2s, v4.2s, v12.2s[1]
|
||||
fmla v25.2s, v5.2s, v13.2s[0]
|
||||
fmla v20.2s, v4.2s, v12.s[1]
|
||||
fmla v25.2s, v5.2s, v13.s[0]
|
||||
|
||||
fmla v24.2s, v4.2s, v13.2s[0]
|
||||
fmla v21.2s, v5.2s, v12.2s[1]
|
||||
fmla v24.2s, v4.2s, v13.s[0]
|
||||
fmla v21.2s, v5.2s, v12.s[1]
|
||||
|
||||
fmla v28.2s, v4.2s, v13.2s[1]
|
||||
fmla v17.2s, v5.2s, v12.2s[0]
|
||||
fmla v28.2s, v4.2s, v13.s[1]
|
||||
fmla v17.2s, v5.2s, v12.s[0]
|
||||
.endm
|
||||
|
||||
.macro KERNEL4x4_SUB
|
||||
@@ -578,17 +578,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s, v1.2s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v16.2s, v0.2s, v8.2s[0]
|
||||
fmla v29.2s, v1.2s, v9.2s[1]
|
||||
fmla v16.2s, v0.2s, v8.s[0]
|
||||
fmla v29.2s, v1.2s, v9.s[1]
|
||||
|
||||
fmla v20.2s, v0.2s, v8.2s[1]
|
||||
fmla v25.2s, v1.2s, v9.2s[0]
|
||||
fmla v20.2s, v0.2s, v8.s[1]
|
||||
fmla v25.2s, v1.2s, v9.s[0]
|
||||
|
||||
fmla v24.2s, v0.2s, v9.2s[0]
|
||||
fmla v21.2s, v1.2s, v8.2s[1]
|
||||
fmla v24.2s, v0.2s, v9.s[0]
|
||||
fmla v21.2s, v1.2s, v8.s[1]
|
||||
|
||||
fmla v28.2s, v0.2s, v9.2s[1]
|
||||
fmla v17.2s, v1.2s, v8.2s[0]
|
||||
fmla v28.2s, v0.2s, v9.s[1]
|
||||
fmla v17.2s, v1.2s, v8.s[0]
|
||||
.endm
|
||||
|
||||
.macro SAVE4x4
|
||||
@@ -633,10 +633,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s}, [pA]
|
||||
add pA, pA, #8
|
||||
|
||||
fmla v16.2s, v0.2s, v8.2s[0]
|
||||
fmla v20.2s, v0.2s, v8.2s[1]
|
||||
fmla v24.2s, v0.2s, v9.2s[0]
|
||||
fmla v28.2s, v0.2s, v9.2s[1]
|
||||
fmla v16.2s, v0.2s, v8.s[0]
|
||||
fmla v20.2s, v0.2s, v8.s[1]
|
||||
fmla v24.2s, v0.2s, v9.s[0]
|
||||
fmla v28.2s, v0.2s, v9.s[1]
|
||||
.endm
|
||||
|
||||
.macro SAVE2x4
|
||||
@@ -718,15 +718,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v3.4s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v16.4s, v0.4s, v8.2s[0]
|
||||
fmla v17.4s, v1.4s, v8.2s[0]
|
||||
fmla v18.4s, v2.4s, v8.2s[0]
|
||||
fmla v19.4s, v3.4s, v8.2s[0]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v17.4s, v1.4s, v8.s[0]
|
||||
fmla v18.4s, v2.4s, v8.s[0]
|
||||
fmla v19.4s, v3.4s, v8.s[0]
|
||||
|
||||
fmla v20.4s, v0.4s, v8.2s[1]
|
||||
fmla v21.4s, v1.4s, v8.2s[1]
|
||||
fmla v22.4s, v2.4s, v8.2s[1]
|
||||
fmla v23.4s, v3.4s, v8.2s[1]
|
||||
fmla v20.4s, v0.4s, v8.s[1]
|
||||
fmla v21.4s, v1.4s, v8.s[1]
|
||||
fmla v22.4s, v2.4s, v8.s[1]
|
||||
fmla v23.4s, v3.4s, v8.s[1]
|
||||
.endm
|
||||
|
||||
.macro SAVE16x2
|
||||
@@ -764,11 +764,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v1.4s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v16.4s, v0.4s, v8.2s[0]
|
||||
fmla v17.4s, v1.4s, v8.2s[0]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v17.4s, v1.4s, v8.s[0]
|
||||
|
||||
fmla v20.4s, v0.4s, v8.2s[1]
|
||||
fmla v21.4s, v1.4s, v8.2s[1]
|
||||
fmla v20.4s, v0.4s, v8.s[1]
|
||||
fmla v21.4s, v1.4s, v8.s[1]
|
||||
.endm
|
||||
|
||||
.macro SAVE8x2
|
||||
@@ -802,10 +802,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s, v1.2s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v16.2s, v0.2s, v8.2s[0]
|
||||
fmla v17.2s, v1.2s, v8.2s[0]
|
||||
fmla v20.2s, v0.2s, v8.2s[1]
|
||||
fmla v21.2s, v1.2s, v8.2s[1]
|
||||
fmla v16.2s, v0.2s, v8.s[0]
|
||||
fmla v17.2s, v1.2s, v8.s[0]
|
||||
fmla v20.2s, v0.2s, v8.s[1]
|
||||
fmla v21.2s, v1.2s, v8.s[1]
|
||||
.endm
|
||||
|
||||
.macro SAVE4x2
|
||||
@@ -837,8 +837,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s}, [pA]
|
||||
add pA, pA, #8
|
||||
|
||||
fmla v16.2s, v0.2s, v8.2s[0]
|
||||
fmla v20.2s, v0.2s, v8.2s[1]
|
||||
fmla v16.2s, v0.2s, v8.s[0]
|
||||
fmla v20.2s, v0.2s, v8.s[1]
|
||||
.endm
|
||||
|
||||
.macro SAVE2x2
|
||||
@@ -866,7 +866,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ldr s0 , [pA]
|
||||
add pA, pA, #4
|
||||
|
||||
fmla v16.2s, v8.2s, v0.2s[0]
|
||||
fmla v16.2s, v8.2s, v0.s[0]
|
||||
.endm
|
||||
|
||||
.macro SAVE1x2
|
||||
@@ -901,10 +901,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v3.4s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v16.4s, v0.4s, v8.2s[0]
|
||||
fmla v17.4s, v1.4s, v8.2s[0]
|
||||
fmla v18.4s, v2.4s, v8.2s[0]
|
||||
fmla v19.4s, v3.4s, v8.2s[0]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v17.4s, v1.4s, v8.s[0]
|
||||
fmla v18.4s, v2.4s, v8.s[0]
|
||||
fmla v19.4s, v3.4s, v8.s[0]
|
||||
.endm
|
||||
|
||||
.macro SAVE16x1
|
||||
@@ -934,8 +934,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v1.4s}, [pA]
|
||||
add pA, pA, #16
|
||||
|
||||
fmla v16.4s, v0.4s, v8.2s[0]
|
||||
fmla v17.4s, v1.4s, v8.2s[0]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v17.4s, v1.4s, v8.s[0]
|
||||
.endm
|
||||
|
||||
.macro SAVE8x1
|
||||
@@ -961,8 +961,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s, v1.2s}, [pA]
|
||||
add pA , pA, #16
|
||||
|
||||
fmla v16.2s, v0.2s, v8.2s[0]
|
||||
fmla v17.2s, v1.2s, v8.2s[0]
|
||||
fmla v16.2s, v0.2s, v8.s[0]
|
||||
fmla v17.2s, v1.2s, v8.s[0]
|
||||
.endm
|
||||
|
||||
.macro SAVE4x1
|
||||
@@ -987,7 +987,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s}, [pA]
|
||||
add pA , pA, #8
|
||||
|
||||
fmla v16.2s, v0.2s, v8.2s[0]
|
||||
fmla v16.2s, v0.2s, v8.s[0]
|
||||
.endm
|
||||
|
||||
.macro SAVE2x1
|
||||
|
||||
Reference in New Issue
Block a user