Fixed iwalsh_neon build problems with RVDS4.1

rvct 4.1 was complaining about vstmia.16, store multiple expects 64 data type.
optimized the implementation.

Change-Id: I0701052cabd685c375637bbc3796ff6d88f5972c

1 files changed, 23 insertions, 39 deletions

diff --git a/vp8/common/arm/neon/iwalsh_neon.asm b/vp8/common/arm/neon/iwalsh_neon.asm
index 663bf390e..01c79d937 100644
--- a/vp8/common/arm/neon/iwalsh_neon.asm
+++ b/vp8/common/arm/neon/iwalsh_neon.asm
@@ -20,19 +20,16 @@
 |vp8_short_inv_walsh4x4_neon| PROC
 
     ; read in all four lines of values: d0->d3
-    vldm.64 r0, {q0, q1}
+    vld1.i16 {q0-q1}, [r0@128]
 
     ; first for loop
-
     vadd.s16 d4, d0, d3 ;a = [0] + [12]
-    vadd.s16 d5, d1, d2 ;b = [4] + [8]
-    vsub.s16 d6, d1, d2 ;c = [4] - [8]
-    vsub.s16 d7, d0, d3 ;d = [0] - [12]
+    vadd.s16 d6, d1, d2 ;b = [4] + [8]
+    vsub.s16 d5, d0, d3 ;d = [0] - [12]
+    vsub.s16 d7, d1, d2 ;c = [4] - [8]
 
-    vadd.s16 d0, d4, d5 ;a + b
-    vadd.s16 d1, d6, d7 ;c + d
-    vsub.s16 d2, d4, d5 ;a - b
-    vsub.s16 d3, d7, d6 ;d - c
+    vadd.s16 q0, q2, q3 ; a+b d+c
+    vsub.s16 q1, q2, q3 ; a-b d-c
 
     vtrn.32 d0, d2 ;d0:  0  1  8  9
                    ;d2:  2  3 10 11
@@ -47,29 +44,22 @@
     ; second for loop
 
     vadd.s16 d4, d0, d3 ;a = [0] + [3]
-    vadd.s16 d5, d1, d2 ;b = [1] + [2]
-    vsub.s16 d6, d1, d2 ;c = [1] - [2]
-    vsub.s16 d7, d0, d3 ;d = [0] - [3]
+    vadd.s16 d6, d1, d2 ;b = [1] + [2]
+    vsub.s16 d5, d0, d3 ;d = [0] - [3]
+    vsub.s16 d7, d1, d2 ;c = [1] - [2]
+
+    vmov.i16 q8, #3
 
-    vadd.s16 d0, d4, d5 ;e = a + b
-    vadd.s16 d1, d6, d7 ;f = c + d
-    vsub.s16 d2, d4, d5 ;g = a - b
-    vsub.s16 d3, d7, d6 ;h = d - c
+    vadd.s16 q0, q2, q3 ; a+b d+c
+    vsub.s16 q1, q2, q3 ; a-b d-c
 
-    vmov.i16 q2, #3
-    vadd.i16 q0, q0, q2 ;e/f += 3
-    vadd.i16 q1, q1, q2 ;g/h += 3
+    vadd.i16 q0, q0, q8 ;e/f += 3
+    vadd.i16 q1, q1, q8 ;g/h += 3
 
     vshr.s16 q0, q0, #3 ;e/f >> 3
     vshr.s16 q1, q1, #3 ;g/h >> 3
 
-    vtrn.32 d0, d2
-    vtrn.32 d1, d3
-    vtrn.16 d0, d1
-    vtrn.16 d2, d3
-
-    vstmia.16 r1!, {q0}
-    vstmia.16 r1!, {q1}
+    vst4.i16 {d0,d1,d2,d3}, [r1@128]
 
     bx lr
     ENDP    ; |vp8_short_inv_walsh4x4_neon|
@@ -77,19 +67,13 @@
 
 ;short vp8_short_inv_walsh4x4_1_neon(short *input, short *output)
 |vp8_short_inv_walsh4x4_1_neon| PROC
-    ; load a full line into a neon register
-    vld1.16  {q0}, [r0]
-    ; extract first element and replicate
-    vdup.16 q1, d0[0]
-    ; add 3 to all values
-    vmov.i16 q2, #3
-    vadd.i16 q3, q1, q2
-    ; right shift
-    vshr.s16 q3, q3, #3
-    ; write it back
-    vstmia.16 r1!, {q3}
-    vstmia.16 r1!, {q3}
-
+    ldrsh r2, [r0]          ; load input[0]
+    add r3, r2, #3          ; add 3
+    add r2, r1, #16         ; base for last 8 output
+    asr r0, r3, #3          ; right shift 3
+    vdup.16 q0, r0          ; load and duplicate
+    vst1.16 {q0}, [r1@128]  ; write back 8
+    vst1.16 {q0}, [r2@128]  ; write back last 8
     bx lr
     ENDP    ; |vp8_short_inv_walsh4x4_1_neon|