1@ This file was created from a .asm file 2@ using the ads2gas.pl script. 3.syntax unified 4@ 5@ Copyright (c) 2013 The WebM project authors. All Rights Reserved. 6@ 7@ Use of this source code is governed by a BSD-style license 8@ that can be found in the LICENSE file in the root of the source 9@ tree. An additional intellectual property rights grant can be found 10@ in the file PATENTS. All contributing project authors may 11@ be found in the AUTHORS file in the root of the source tree. 12@ 13 14 .global vpx_convolve_avg_neon 15 .type vpx_convolve_avg_neon, function 16 .arm 17 .eabi_attribute 24, 1 @Tag_ABI_align_needed 18 .eabi_attribute 25, 1 @Tag_ABI_align_preserved 19 20 .text 21 .p2align 2 22 23vpx_convolve_avg_neon: @ PROC 24 push {r4-r6, lr} 25 ldrd r4, r5, [sp, #36] 26 mov r6, r2 27 28 cmp r4, #32 29 bgt avg64 30 beq avg32 31 cmp r4, #8 32 bgt avg16 33 beq avg8 34 b avg4 35 36avg64: 37 sub lr, r1, #32 38 sub r4, r3, #32 39avg64_h: 40 pld [r0, r1, lsl #1] 41 vld1.8 {q0-q1}, [r0]! 42 vld1.8 {q2-q3}, [r0], lr 43 pld [r2, r3] 44 vld1.8 {q8-q9}, [r6,:128]! 45 vld1.8 {q10-q11}, [r6,:128], r4 46 vrhadd.u8 q0, q0, q8 47 vrhadd.u8 q1, q1, q9 48 vrhadd.u8 q2, q2, q10 49 vrhadd.u8 q3, q3, q11 50 vst1.8 {q0-q1}, [r2,:128]! 51 vst1.8 {q2-q3}, [r2,:128], r4 52 subs r5, r5, #1 53 bgt avg64_h 54 pop {r4-r6, pc} 55 56avg32: 57 vld1.8 {q0-q1}, [r0], r1 58 vld1.8 {q2-q3}, [r0], r1 59 vld1.8 {q8-q9}, [r6,:128], r3 60 vld1.8 {q10-q11}, [r6,:128], r3 61 pld [r0] 62 vrhadd.u8 q0, q0, q8 63 pld [r0, r1] 64 vrhadd.u8 q1, q1, q9 65 pld [r6] 66 vrhadd.u8 q2, q2, q10 67 pld [r6, r3] 68 vrhadd.u8 q3, q3, q11 69 vst1.8 {q0-q1}, [r2,:128], r3 70 vst1.8 {q2-q3}, [r2,:128], r3 71 subs r5, r5, #2 72 bgt avg32 73 pop {r4-r6, pc} 74 75avg16: 76 vld1.8 {q0}, [r0], r1 77 vld1.8 {q1}, [r0], r1 78 vld1.8 {q2}, [r6,:128], r3 79 vld1.8 {q3}, [r6,:128], r3 80 pld [r0] 81 pld [r0, r1] 82 vrhadd.u8 q0, q0, q2 83 pld [r6] 84 pld [r6, r3] 85 vrhadd.u8 q1, q1, q3 86 vst1.8 {q0}, [r2,:128], r3 87 vst1.8 {q1}, [r2,:128], r3 88 subs r5, r5, #2 89 bgt avg16 90 pop {r4-r6, pc} 91 92avg8: 93 vld1.8 {d0}, [r0], r1 94 vld1.8 {d1}, [r0], r1 95 vld1.8 {d2}, [r6,:64], r3 96 vld1.8 {d3}, [r6,:64], r3 97 pld [r0] 98 pld [r0, r1] 99 vrhadd.u8 q0, q0, q1 100 pld [r6] 101 pld [r6, r3] 102 vst1.8 {d0}, [r2,:64], r3 103 vst1.8 {d1}, [r2,:64], r3 104 subs r5, r5, #2 105 bgt avg8 106 pop {r4-r6, pc} 107 108avg4: 109 vld1.32 {d0[0]}, [r0], r1 110 vld1.32 {d0[1]}, [r0], r1 111 vld1.32 {d2[0]}, [r6,:32], r3 112 vld1.32 {d2[1]}, [r6,:32], r3 113 vrhadd.u8 d0, d0, d2 114 vst1.32 {d0[0]}, [r2,:32], r3 115 vst1.32 {d0[1]}, [r2,:32], r3 116 subs r5, r5, #2 117 bgt avg4 118 pop {r4-r6, pc} 119.size vpx_convolve_avg_neon, .-vpx_convolve_avg_neon @ ENDP 120 121 .section .note.GNU-stack,"",%progbits 122