michael@0: ;
michael@0: ;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
michael@0: ;
michael@0: ;  Use of this source code is governed by a BSD-style license
michael@0: ;  that can be found in the LICENSE file in the root of the source
michael@0: ;  tree. An additional intellectual property rights grant can be found
michael@0: ;  in the file PATENTS.  All contributing project authors may
michael@0: ;  be found in the AUTHORS file in the root of the source tree.
michael@0: ;
michael@0: 
michael@0:     EXPORT  |vp9_convolve_copy_neon|
michael@0:     ARM
michael@0:     REQUIRE8
michael@0:     PRESERVE8
michael@0: 
michael@0:     AREA ||.text||, CODE, READONLY, ALIGN=2
michael@0: 
michael@0: |vp9_convolve_copy_neon| PROC
michael@0:     push                {r4-r5, lr}
michael@0:     ldrd                r4, r5, [sp, #28]
michael@0: 
michael@0:     cmp                 r4, #32
michael@0:     bgt                 copy64
michael@0:     beq                 copy32
michael@0:     cmp                 r4, #8
michael@0:     bgt                 copy16
michael@0:     beq                 copy8
michael@0:     b                   copy4
michael@0: 
michael@0: copy64
michael@0:     sub                 lr, r1, #32
michael@0:     sub                 r3, r3, #32
michael@0: copy64_h
michael@0:     pld                 [r0, r1, lsl #1]
michael@0:     vld1.8              {q0-q1}, [r0]!
michael@0:     vld1.8              {q2-q3}, [r0], lr
michael@0:     vst1.8              {q0-q1}, [r2@128]!
michael@0:     vst1.8              {q2-q3}, [r2@128], r3
michael@0:     subs                r5, r5, #1
michael@0:     bgt                 copy64_h
michael@0:     pop                 {r4-r5, pc}
michael@0: 
michael@0: copy32
michael@0:     pld                 [r0, r1, lsl #1]
michael@0:     vld1.8              {q0-q1}, [r0], r1
michael@0:     pld                 [r0, r1, lsl #1]
michael@0:     vld1.8              {q2-q3}, [r0], r1
michael@0:     vst1.8              {q0-q1}, [r2@128], r3
michael@0:     vst1.8              {q2-q3}, [r2@128], r3
michael@0:     subs                r5, r5, #2
michael@0:     bgt                 copy32
michael@0:     pop                 {r4-r5, pc}
michael@0: 
michael@0: copy16
michael@0:     pld                 [r0, r1, lsl #1]
michael@0:     vld1.8              {q0}, [r0], r1
michael@0:     pld                 [r0, r1, lsl #1]
michael@0:     vld1.8              {q1}, [r0], r1
michael@0:     vst1.8              {q0}, [r2@128], r3
michael@0:     vst1.8              {q1}, [r2@128], r3
michael@0:     subs                r5, r5, #2
michael@0:     bgt                 copy16
michael@0:     pop                 {r4-r5, pc}
michael@0: 
michael@0: copy8
michael@0:     pld                 [r0, r1, lsl #1]
michael@0:     vld1.8              {d0}, [r0], r1
michael@0:     pld                 [r0, r1, lsl #1]
michael@0:     vld1.8              {d2}, [r0], r1
michael@0:     vst1.8              {d0}, [r2@64], r3
michael@0:     vst1.8              {d2}, [r2@64], r3
michael@0:     subs                r5, r5, #2
michael@0:     bgt                 copy8
michael@0:     pop                 {r4-r5, pc}
michael@0: 
michael@0: copy4
michael@0:     ldr                 r12, [r0], r1
michael@0:     str                 r12, [r2], r3
michael@0:     subs                r5, r5, #1
michael@0:     bgt                 copy4
michael@0:     pop                 {r4-r5, pc}
michael@0:     ENDP
michael@0: 
michael@0:     END