1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/common/arm/neon/vp9_copy_neon.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,84 @@
1.4 +;
1.5 +; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
1.6 +;
1.7 +; Use of this source code is governed by a BSD-style license
1.8 +; that can be found in the LICENSE file in the root of the source
1.9 +; tree. An additional intellectual property rights grant can be found
1.10 +; in the file PATENTS. All contributing project authors may
1.11 +; be found in the AUTHORS file in the root of the source tree.
1.12 +;
1.13 +
1.14 + EXPORT |vp9_convolve_copy_neon|
1.15 + ARM
1.16 + REQUIRE8
1.17 + PRESERVE8
1.18 +
1.19 + AREA ||.text||, CODE, READONLY, ALIGN=2
1.20 +
1.21 +|vp9_convolve_copy_neon| PROC
1.22 + push {r4-r5, lr}
1.23 + ldrd r4, r5, [sp, #28]
1.24 +
1.25 + cmp r4, #32
1.26 + bgt copy64
1.27 + beq copy32
1.28 + cmp r4, #8
1.29 + bgt copy16
1.30 + beq copy8
1.31 + b copy4
1.32 +
1.33 +copy64
1.34 + sub lr, r1, #32
1.35 + sub r3, r3, #32
1.36 +copy64_h
1.37 + pld [r0, r1, lsl #1]
1.38 + vld1.8 {q0-q1}, [r0]!
1.39 + vld1.8 {q2-q3}, [r0], lr
1.40 + vst1.8 {q0-q1}, [r2@128]!
1.41 + vst1.8 {q2-q3}, [r2@128], r3
1.42 + subs r5, r5, #1
1.43 + bgt copy64_h
1.44 + pop {r4-r5, pc}
1.45 +
1.46 +copy32
1.47 + pld [r0, r1, lsl #1]
1.48 + vld1.8 {q0-q1}, [r0], r1
1.49 + pld [r0, r1, lsl #1]
1.50 + vld1.8 {q2-q3}, [r0], r1
1.51 + vst1.8 {q0-q1}, [r2@128], r3
1.52 + vst1.8 {q2-q3}, [r2@128], r3
1.53 + subs r5, r5, #2
1.54 + bgt copy32
1.55 + pop {r4-r5, pc}
1.56 +
1.57 +copy16
1.58 + pld [r0, r1, lsl #1]
1.59 + vld1.8 {q0}, [r0], r1
1.60 + pld [r0, r1, lsl #1]
1.61 + vld1.8 {q1}, [r0], r1
1.62 + vst1.8 {q0}, [r2@128], r3
1.63 + vst1.8 {q1}, [r2@128], r3
1.64 + subs r5, r5, #2
1.65 + bgt copy16
1.66 + pop {r4-r5, pc}
1.67 +
1.68 +copy8
1.69 + pld [r0, r1, lsl #1]
1.70 + vld1.8 {d0}, [r0], r1
1.71 + pld [r0, r1, lsl #1]
1.72 + vld1.8 {d2}, [r0], r1
1.73 + vst1.8 {d0}, [r2@64], r3
1.74 + vst1.8 {d2}, [r2@64], r3
1.75 + subs r5, r5, #2
1.76 + bgt copy8
1.77 + pop {r4-r5, pc}
1.78 +
1.79 +copy4
1.80 + ldr r12, [r0], r1
1.81 + str r12, [r2], r3
1.82 + subs r5, r5, #1
1.83 + bgt copy4
1.84 + pop {r4-r5, pc}
1.85 + ENDP
1.86 +
1.87 + END
