Cortex-A57: Fix clang compilation errors
This commit is contained in:
@@ -192,164 +192,164 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.4s}, [pA_0]
|
||||
add pA_0, pA_0, #16
|
||||
|
||||
fmul v16.4s, v0.4s, v8.4s[0]
|
||||
fmul v20.4s, v0.4s, v8.4s[1]
|
||||
fmul v16.4s, v0.4s, v8.s[0]
|
||||
fmul v20.4s, v0.4s, v8.s[1]
|
||||
|
||||
ld1 {v2.4s}, [pA_1]
|
||||
add pA_1, pA_1, #16
|
||||
|
||||
fmul v24.4s, v0.4s, v8.4s[2]
|
||||
fmul v28.4s, v0.4s, v8.4s[3]
|
||||
fmul v24.4s, v0.4s, v8.s[2]
|
||||
fmul v28.4s, v0.4s, v8.s[3]
|
||||
|
||||
ld1 {v4.4s}, [pA_2]
|
||||
add pA_2, pA_2, #16
|
||||
|
||||
fmul v17.4s, v2.4s, v8.4s[0]
|
||||
fmul v21.4s, v2.4s, v8.4s[1]
|
||||
fmul v17.4s, v2.4s, v8.s[0]
|
||||
fmul v21.4s, v2.4s, v8.s[1]
|
||||
|
||||
ld1 {v6.4s}, [pA_3]
|
||||
add pA_3, pA_3, #16
|
||||
|
||||
fmul v25.4s, v2.4s, v8.4s[2]
|
||||
fmul v29.4s, v2.4s, v8.4s[3]
|
||||
fmul v25.4s, v2.4s, v8.s[2]
|
||||
fmul v29.4s, v2.4s, v8.s[3]
|
||||
|
||||
ld1 {v12.4s}, [pB] // for next round
|
||||
add pB, pB, #16
|
||||
|
||||
fmul v18.4s, v4.4s, v8.4s[0]
|
||||
fmul v19.4s, v6.4s, v8.4s[0]
|
||||
fmul v18.4s, v4.4s, v8.s[0]
|
||||
fmul v19.4s, v6.4s, v8.s[0]
|
||||
|
||||
ld1 {v1.4s}, [pA_0] // for next round
|
||||
add pA_0, pA_0, #16
|
||||
|
||||
fmul v22.4s, v4.4s, v8.4s[1]
|
||||
fmul v23.4s, v6.4s, v8.4s[1]
|
||||
fmul v22.4s, v4.4s, v8.s[1]
|
||||
fmul v23.4s, v6.4s, v8.s[1]
|
||||
|
||||
ld1 {v3.4s}, [pA_1] // for next round
|
||||
add pA_1, pA_1, #16
|
||||
|
||||
fmul v26.4s, v4.4s, v8.4s[2]
|
||||
fmul v27.4s, v6.4s, v8.4s[2]
|
||||
fmul v26.4s, v4.4s, v8.s[2]
|
||||
fmul v27.4s, v6.4s, v8.s[2]
|
||||
|
||||
ld1 {v5.4s}, [pA_2] // for next round
|
||||
add pA_2, pA_2, #16
|
||||
|
||||
fmul v30.4s, v4.4s, v8.4s[3]
|
||||
fmul v31.4s, v6.4s, v8.4s[3]
|
||||
fmul v30.4s, v4.4s, v8.s[3]
|
||||
fmul v31.4s, v6.4s, v8.s[3]
|
||||
|
||||
ld1 {v7.4s}, [pA_3] // for next round
|
||||
add pA_3, pA_3, #16
|
||||
.endm
|
||||
|
||||
.macro KERNEL16x4_M2
|
||||
fmla v16.4s, v1.4s, v12.4s[0]
|
||||
fmla v17.4s, v3.4s, v12.4s[0]
|
||||
fmla v16.4s, v1.4s, v12.s[0]
|
||||
fmla v17.4s, v3.4s, v12.s[0]
|
||||
|
||||
ld1 {v8.4s}, [pB] // for next round
|
||||
add pB, pB, #16
|
||||
|
||||
fmla v18.4s, v5.4s, v12.4s[0]
|
||||
fmla v19.4s, v7.4s, v12.4s[0]
|
||||
fmla v18.4s, v5.4s, v12.s[0]
|
||||
fmla v19.4s, v7.4s, v12.s[0]
|
||||
|
||||
ld1 {v0.4s}, [pA_0] // for next round
|
||||
add pA_0, pA_0, #16
|
||||
|
||||
fmla v20.4s, v1.4s, v12.4s[1]
|
||||
fmla v21.4s, v3.4s, v12.4s[1]
|
||||
fmla v20.4s, v1.4s, v12.s[1]
|
||||
fmla v21.4s, v3.4s, v12.s[1]
|
||||
|
||||
ld1 {v2.4s}, [pA_1] // for next round
|
||||
add pA_1, pA_1, #16
|
||||
|
||||
fmla v22.4s, v5.4s, v12.4s[1]
|
||||
fmla v23.4s, v7.4s, v12.4s[1]
|
||||
fmla v22.4s, v5.4s, v12.s[1]
|
||||
fmla v23.4s, v7.4s, v12.s[1]
|
||||
|
||||
ld1 {v4.4s}, [pA_2] // for next round
|
||||
add pA_2, pA_2, #16
|
||||
|
||||
fmla v24.4s, v1.4s, v12.4s[2]
|
||||
fmla v25.4s, v3.4s, v12.4s[2]
|
||||
fmla v24.4s, v1.4s, v12.s[2]
|
||||
fmla v25.4s, v3.4s, v12.s[2]
|
||||
|
||||
ld1 {v6.4s}, [pA_3] // for next round
|
||||
add pA_3, pA_3, #16
|
||||
|
||||
fmla v26.4s, v5.4s, v12.4s[2]
|
||||
fmla v27.4s, v7.4s, v12.4s[2]
|
||||
fmla v26.4s, v5.4s, v12.s[2]
|
||||
fmla v27.4s, v7.4s, v12.s[2]
|
||||
|
||||
prfm PLDL1KEEP, [pA_2, #512]
|
||||
|
||||
fmla v28.4s, v1.4s, v12.4s[3]
|
||||
fmla v29.4s, v3.4s, v12.4s[3]
|
||||
fmla v28.4s, v1.4s, v12.s[3]
|
||||
fmla v29.4s, v3.4s, v12.s[3]
|
||||
|
||||
prfm PLDL1KEEP, [pA_3, #512]
|
||||
|
||||
fmla v30.4s, v5.4s, v12.4s[3]
|
||||
fmla v31.4s, v7.4s, v12.4s[3]
|
||||
fmla v30.4s, v5.4s, v12.s[3]
|
||||
fmla v31.4s, v7.4s, v12.s[3]
|
||||
|
||||
prfm PLDL1KEEP, [pB, #512]
|
||||
.endm
|
||||
|
||||
.macro KERNEL16x4_M1
|
||||
fmla v16.4s, v0.4s, v8.4s[0]
|
||||
fmla v17.4s, v2.4s, v8.4s[0]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v17.4s, v2.4s, v8.s[0]
|
||||
|
||||
ld1 {v12.4s}, [pB] // for next round
|
||||
add pB, pB, #16
|
||||
|
||||
fmla v18.4s, v4.4s, v8.4s[0]
|
||||
fmla v19.4s, v6.4s, v8.4s[0]
|
||||
fmla v18.4s, v4.4s, v8.s[0]
|
||||
fmla v19.4s, v6.4s, v8.s[0]
|
||||
|
||||
ld1 {v1.4s}, [pA_0] // for next round
|
||||
add pA_0, pA_0, #16
|
||||
|
||||
fmla v20.4s, v0.4s, v8.4s[1]
|
||||
fmla v21.4s, v2.4s, v8.4s[1]
|
||||
fmla v20.4s, v0.4s, v8.s[1]
|
||||
fmla v21.4s, v2.4s, v8.s[1]
|
||||
|
||||
ld1 {v3.4s}, [pA_1] // for next round
|
||||
add pA_1, pA_1, #16
|
||||
|
||||
fmla v22.4s, v4.4s, v8.4s[1]
|
||||
fmla v23.4s, v6.4s, v8.4s[1]
|
||||
fmla v22.4s, v4.4s, v8.s[1]
|
||||
fmla v23.4s, v6.4s, v8.s[1]
|
||||
|
||||
ld1 {v5.4s}, [pA_2] // for next round
|
||||
add pA_2, pA_2, #16
|
||||
|
||||
fmla v24.4s, v0.4s, v8.4s[2]
|
||||
fmla v25.4s, v2.4s, v8.4s[2]
|
||||
fmla v24.4s, v0.4s, v8.s[2]
|
||||
fmla v25.4s, v2.4s, v8.s[2]
|
||||
|
||||
ld1 {v7.4s}, [pA_3] // for next round
|
||||
add pA_3, pA_3, #16
|
||||
|
||||
fmla v26.4s, v4.4s, v8.4s[2]
|
||||
fmla v27.4s, v6.4s, v8.4s[2]
|
||||
fmla v26.4s, v4.4s, v8.s[2]
|
||||
fmla v27.4s, v6.4s, v8.s[2]
|
||||
|
||||
prfm PLDL1KEEP, [pA_0, #512]
|
||||
|
||||
fmla v28.4s, v0.4s, v8.4s[3]
|
||||
fmla v29.4s, v2.4s, v8.4s[3]
|
||||
fmla v28.4s, v0.4s, v8.s[3]
|
||||
fmla v29.4s, v2.4s, v8.s[3]
|
||||
|
||||
prfm PLDL1KEEP, [pA_1, #512]
|
||||
|
||||
fmla v30.4s, v4.4s, v8.4s[3]
|
||||
fmla v31.4s, v6.4s, v8.4s[3]
|
||||
fmla v30.4s, v4.4s, v8.s[3]
|
||||
fmla v31.4s, v6.4s, v8.s[3]
|
||||
.endm
|
||||
|
||||
.macro KERNEL16x4_E
|
||||
fmla v16.4s, v1.4s, v12.4s[0]
|
||||
fmla v17.4s, v3.4s, v12.4s[0]
|
||||
fmla v18.4s, v5.4s, v12.4s[0]
|
||||
fmla v19.4s, v7.4s, v12.4s[0]
|
||||
fmla v20.4s, v1.4s, v12.4s[1]
|
||||
fmla v21.4s, v3.4s, v12.4s[1]
|
||||
fmla v22.4s, v5.4s, v12.4s[1]
|
||||
fmla v23.4s, v7.4s, v12.4s[1]
|
||||
fmla v24.4s, v1.4s, v12.4s[2]
|
||||
fmla v25.4s, v3.4s, v12.4s[2]
|
||||
fmla v26.4s, v5.4s, v12.4s[2]
|
||||
fmla v27.4s, v7.4s, v12.4s[2]
|
||||
fmla v28.4s, v1.4s, v12.4s[3]
|
||||
fmla v29.4s, v3.4s, v12.4s[3]
|
||||
fmla v30.4s, v5.4s, v12.4s[3]
|
||||
fmla v31.4s, v7.4s, v12.4s[3]
|
||||
fmla v16.4s, v1.4s, v12.s[0]
|
||||
fmla v17.4s, v3.4s, v12.s[0]
|
||||
fmla v18.4s, v5.4s, v12.s[0]
|
||||
fmla v19.4s, v7.4s, v12.s[0]
|
||||
fmla v20.4s, v1.4s, v12.s[1]
|
||||
fmla v21.4s, v3.4s, v12.s[1]
|
||||
fmla v22.4s, v5.4s, v12.s[1]
|
||||
fmla v23.4s, v7.4s, v12.s[1]
|
||||
fmla v24.4s, v1.4s, v12.s[2]
|
||||
fmla v25.4s, v3.4s, v12.s[2]
|
||||
fmla v26.4s, v5.4s, v12.s[2]
|
||||
fmla v27.4s, v7.4s, v12.s[2]
|
||||
fmla v28.4s, v1.4s, v12.s[3]
|
||||
fmla v29.4s, v3.4s, v12.s[3]
|
||||
fmla v30.4s, v5.4s, v12.s[3]
|
||||
fmla v31.4s, v7.4s, v12.s[3]
|
||||
.endm
|
||||
|
||||
.macro KERNEL16x4_SUB
|
||||
@@ -359,34 +359,34 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.4s}, [pA_0]
|
||||
add pA_0, pA_0, #16
|
||||
|
||||
fmla v16.4s, v0.4s, v8.4s[0]
|
||||
fmla v20.4s, v0.4s, v8.4s[1]
|
||||
fmla v24.4s, v0.4s, v8.4s[2]
|
||||
fmla v28.4s, v0.4s, v8.4s[3]
|
||||
fmla v16.4s, v0.4s, v8.s[0]
|
||||
fmla v20.4s, v0.4s, v8.s[1]
|
||||
fmla v24.4s, v0.4s, v8.s[2]
|
||||
fmla v28.4s, v0.4s, v8.s[3]
|
||||
|
||||
ld1 {v2.4s}, [pA_1]
|
||||
add pA_1, pA_1, #16
|
||||
|
||||
fmla v17.4s, v2.4s, v8.4s[0]
|
||||
fmla v21.4s, v2.4s, v8.4s[1]
|
||||
fmla v25.4s, v2.4s, v8.4s[2]
|
||||
fmla v29.4s, v2.4s, v8.4s[3]
|
||||
fmla v17.4s, v2.4s, v8.s[0]
|
||||
fmla v21.4s, v2.4s, v8.s[1]
|
||||
fmla v25.4s, v2.4s, v8.s[2]
|
||||
fmla v29.4s, v2.4s, v8.s[3]
|
||||
|
||||
ld1 {v4.4s}, [pA_2]
|
||||
add pA_2, pA_2, #16
|
||||
|
||||
fmla v18.4s, v4.4s, v8.4s[0]
|
||||
fmla v22.4s, v4.4s, v8.4s[1]
|
||||
fmla v26.4s, v4.4s, v8.4s[2]
|
||||
fmla v30.4s, v4.4s, v8.4s[3]
|
||||
fmla v18.4s, v4.4s, v8.s[0]
|
||||
fmla v22.4s, v4.4s, v8.s[1]
|
||||
fmla v26.4s, v4.4s, v8.s[2]
|
||||
fmla v30.4s, v4.4s, v8.s[3]
|
||||
|
||||
ld1 {v6.4s}, [pA_3]
|
||||
add pA_3, pA_3, #16
|
||||
|
||||
fmla v19.4s, v6.4s, v8.4s[0]
|
||||
fmla v23.4s, v6.4s, v8.4s[1]
|
||||
fmla v27.4s, v6.4s, v8.4s[2]
|
||||
fmla v31.4s, v6.4s, v8.4s[3]
|
||||
fmla v19.4s, v6.4s, v8.s[0]
|
||||
fmla v23.4s, v6.4s, v8.s[1]
|
||||
fmla v27.4s, v6.4s, v8.s[2]
|
||||
fmla v31.4s, v6.4s, v8.s[3]
|
||||
.endm
|
||||
|
||||
.macro SAVE16x4
|
||||
@@ -456,28 +456,28 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s, v1.2s}, [pA_0]
|
||||
add pA_0, pA_0, #16
|
||||
|
||||
fmla v16.2s, v0.2s, v8.2s[0]
|
||||
fmla v29.2s, v1.2s, v9.2s[1]
|
||||
fmla v20.2s, v0.2s, v8.2s[1]
|
||||
fmla v25.2s, v1.2s, v9.2s[0]
|
||||
fmla v16.2s, v0.2s, v8.s[0]
|
||||
fmla v29.2s, v1.2s, v9.s[1]
|
||||
fmla v20.2s, v0.2s, v8.s[1]
|
||||
fmla v25.2s, v1.2s, v9.s[0]
|
||||
|
||||
ld1 {v2.2s, v3.2s}, [pA_1]
|
||||
add pA_1, pA_1, #16
|
||||
|
||||
fmla v24.2s, v0.2s, v9.2s[0]
|
||||
fmla v21.2s, v1.2s, v8.2s[1]
|
||||
fmla v28.2s, v0.2s, v9.2s[1]
|
||||
fmla v17.2s, v1.2s, v8.2s[0]
|
||||
fmla v24.2s, v0.2s, v9.s[0]
|
||||
fmla v21.2s, v1.2s, v8.s[1]
|
||||
fmla v28.2s, v0.2s, v9.s[1]
|
||||
fmla v17.2s, v1.2s, v8.s[0]
|
||||
|
||||
fmla v18.2s, v2.2s, v8.2s[0]
|
||||
fmla v31.2s, v3.2s, v9.2s[1]
|
||||
fmla v22.2s, v2.2s, v8.2s[1]
|
||||
fmla v27.2s, v3.2s, v9.2s[0]
|
||||
fmla v18.2s, v2.2s, v8.s[0]
|
||||
fmla v31.2s, v3.2s, v9.s[1]
|
||||
fmla v22.2s, v2.2s, v8.s[1]
|
||||
fmla v27.2s, v3.2s, v9.s[0]
|
||||
|
||||
fmla v26.2s, v2.2s, v9.2s[0]
|
||||
fmla v23.2s, v3.2s, v8.2s[1]
|
||||
fmla v30.2s, v2.2s, v9.2s[1]
|
||||
fmla v19.2s, v3.2s, v8.2s[0]
|
||||
fmla v26.2s, v2.2s, v9.s[0]
|
||||
fmla v23.2s, v3.2s, v8.s[1]
|
||||
fmla v30.2s, v2.2s, v9.s[1]
|
||||
fmla v19.2s, v3.2s, v8.s[0]
|
||||
.endm
|
||||
|
||||
.macro SAVE8x4
|
||||
@@ -556,17 +556,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s, v1.2s}, [pA_0]
|
||||
add pA_0, pA_0, #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
|
||||
@@ -614,10 +614,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s}, [pA_0]
|
||||
add pA_0, pA_0, #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
|
||||
@@ -700,10 +700,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s, v1.2s}, [pA_0]
|
||||
add pA_0, pA_0, #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
|
||||
@@ -736,8 +736,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s}, [pA_0]
|
||||
add pA_0, pA_0, #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
|
||||
@@ -767,7 +767,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ldr s0 , [pA_0]
|
||||
add pA_0, pA_0, #4
|
||||
|
||||
fmla v16.2s, v8.2s, v0.2s[0]
|
||||
fmla v16.2s, v8.2s, v0.s[0]
|
||||
.endm
|
||||
|
||||
.macro SAVE1x2
|
||||
@@ -796,8 +796,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s, v1.2s}, [pA_0]
|
||||
add pA_0 , pA_0, #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
|
||||
@@ -825,7 +825,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
ld1 {v0.2s}, [pA_0]
|
||||
add pA_0 , pA_0, #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