1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/encoder/arm/neon/vp8_memcpy_neon.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,70 @@
1.4 +;
1.5 +; Copyright (c) 2010 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 +
1.15 + EXPORT |vp8_memcpy_partial_neon|
1.16 +
1.17 + ARM
1.18 + REQUIRE8
1.19 + PRESERVE8
1.20 +
1.21 + AREA ||.text||, CODE, READONLY, ALIGN=2
1.22 +;=========================================
1.23 +;this is not a full memcpy function!!!
1.24 +;void vp8_memcpy_partial_neon(unsigned char *dst_ptr, unsigned char *src_ptr,
1.25 +; int sz);
1.26 +|vp8_memcpy_partial_neon| PROC
1.27 + ;pld [r1] ;preload pred data
1.28 + ;pld [r1, #128]
1.29 + ;pld [r1, #256]
1.30 + ;pld [r1, #384]
1.31 +
1.32 + mov r12, r2, lsr #8 ;copy 256 bytes data at one time
1.33 +
1.34 +memcpy_neon_loop
1.35 + vld1.8 {q0, q1}, [r1]! ;load src data
1.36 + subs r12, r12, #1
1.37 + vld1.8 {q2, q3}, [r1]!
1.38 + vst1.8 {q0, q1}, [r0]! ;copy to dst_ptr
1.39 + vld1.8 {q4, q5}, [r1]!
1.40 + vst1.8 {q2, q3}, [r0]!
1.41 + vld1.8 {q6, q7}, [r1]!
1.42 + vst1.8 {q4, q5}, [r0]!
1.43 + vld1.8 {q8, q9}, [r1]!
1.44 + vst1.8 {q6, q7}, [r0]!
1.45 + vld1.8 {q10, q11}, [r1]!
1.46 + vst1.8 {q8, q9}, [r0]!
1.47 + vld1.8 {q12, q13}, [r1]!
1.48 + vst1.8 {q10, q11}, [r0]!
1.49 + vld1.8 {q14, q15}, [r1]!
1.50 + vst1.8 {q12, q13}, [r0]!
1.51 + vst1.8 {q14, q15}, [r0]!
1.52 +
1.53 + ;pld [r1] ;preload pred data -- need to adjust for real device
1.54 + ;pld [r1, #128]
1.55 + ;pld [r1, #256]
1.56 + ;pld [r1, #384]
1.57 +
1.58 + bne memcpy_neon_loop
1.59 +
1.60 + ands r3, r2, #0xff ;extra copy
1.61 + beq done_copy_neon_loop
1.62 +
1.63 +extra_copy_neon_loop
1.64 + vld1.8 {q0}, [r1]! ;load src data
1.65 + subs r3, r3, #16
1.66 + vst1.8 {q0}, [r0]!
1.67 + bne extra_copy_neon_loop
1.68 +
1.69 +done_copy_neon_loop
1.70 + bx lr
1.71 + ENDP
1.72 +
1.73 + END
