xref: /aosp_15_r20/external/libmpeg2/common/arm/icv_sad_a9.s (revision a97c2a1f0a796dc32bed80d3353c69c5fc07c750)
1*a97c2a1fSXin Li@/******************************************************************************
2*a97c2a1fSXin Li@ *
3*a97c2a1fSXin Li@ * Copyright (C) 2015 The Android Open Source Project
4*a97c2a1fSXin Li@ *
5*a97c2a1fSXin Li@ * Licensed under the Apache License, Version 2.0 (the "License");
6*a97c2a1fSXin Li@ * you may not use this file except in compliance with the License.
7*a97c2a1fSXin Li@ * You may obtain a copy of the License at:
8*a97c2a1fSXin Li@ *
9*a97c2a1fSXin Li@ * http://www.apache.org/licenses/LICENSE-2.0
10*a97c2a1fSXin Li@ *
11*a97c2a1fSXin Li@ * Unless required by applicable law or agreed to in writing, software
12*a97c2a1fSXin Li@ * distributed under the License is distributed on an "AS IS" BASIS,
13*a97c2a1fSXin Li@ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14*a97c2a1fSXin Li@ * See the License for the specific language governing permissions and
15*a97c2a1fSXin Li@ * limitations under the License.
16*a97c2a1fSXin Li@ *
17*a97c2a1fSXin Li@ *****************************************************************************
18*a97c2a1fSXin Li@ * Originally developed and contributed by Ittiam Systems Pvt. Ltd, Bangalore
19*a97c2a1fSXin Li@*/
20*a97c2a1fSXin Li
21*a97c2a1fSXin Li
22*a97c2a1fSXin Li@******************************************************************************
23*a97c2a1fSXin Li@*
24*a97c2a1fSXin Li@*
25*a97c2a1fSXin Li@* @brief
26*a97c2a1fSXin Li@*  This file contains definitions of routines for SAD caclulation
27*a97c2a1fSXin Li@*
28*a97c2a1fSXin Li@* @author
29*a97c2a1fSXin Li@*  Ittiam
30*a97c2a1fSXin Li@*
31*a97c2a1fSXin Li@* @par List of Functions:
32*a97c2a1fSXin Li@*  - icv_sad_8x4_a9()
33*a97c2a1fSXin Li@*
34*a97c2a1fSXin Li@* @remarks
35*a97c2a1fSXin Li@*  None
36*a97c2a1fSXin Li@*
37*a97c2a1fSXin Li@*******************************************************************************
38*a97c2a1fSXin Li
39*a97c2a1fSXin Li
40*a97c2a1fSXin Li@******************************************************************************
41*a97c2a1fSXin Li@*
42*a97c2a1fSXin Li@*  @brief computes distortion (SAD) between 2 8x4  blocks
43*a97c2a1fSXin Li@*
44*a97c2a1fSXin Li@*
45*a97c2a1fSXin Li@*  @par   Description
46*a97c2a1fSXin Li@*   This functions computes SAD between 2 8x4 blocks.
47*a97c2a1fSXin Li@*
48*a97c2a1fSXin Li@* @param[in] pu1_src
49*a97c2a1fSXin Li@*  UWORD8 pointer to the source
50*a97c2a1fSXin Li@*
51*a97c2a1fSXin Li@* @param[out] pu1_ref
52*a97c2a1fSXin Li@*  UWORD8 pointer to the reference buffer
53*a97c2a1fSXin Li@*
54*a97c2a1fSXin Li@* @param[in] src_strd
55*a97c2a1fSXin Li@*  integer source stride
56*a97c2a1fSXin Li@*
57*a97c2a1fSXin Li@* @param[in] ref_strd
58*a97c2a1fSXin Li@*  integer reference stride
59*a97c2a1fSXin Li@*
60*a97c2a1fSXin Li@* @param[in] wd
61*a97c2a1fSXin Li@*  Width (assumed to be 8)
62*a97c2a1fSXin Li@*
63*a97c2a1fSXin Li@* @param[in] ht
64*a97c2a1fSXin Li@*  Height (assumed to be 4)
65*a97c2a1fSXin Li@*
66*a97c2a1fSXin Li@* @returns
67*a97c2a1fSXin Li@*  SAD value in r0
68*a97c2a1fSXin Li@*
69*a97c2a1fSXin Li@* @remarks
70*a97c2a1fSXin Li@*
71*a97c2a1fSXin Li@******************************************************************************
72*a97c2a1fSXin Li
73*a97c2a1fSXin Li    .global icv_sad_8x4_a9
74*a97c2a1fSXin Li
75*a97c2a1fSXin Liicv_sad_8x4_a9:
76*a97c2a1fSXin Li
77*a97c2a1fSXin Li    push          {lr}
78*a97c2a1fSXin Li
79*a97c2a1fSXin Li    vld1.8        d4, [r0], r2
80*a97c2a1fSXin Li    vld1.8        d5, [r1], r3
81*a97c2a1fSXin Li
82*a97c2a1fSXin Li    vld1.8        d6, [r0], r2
83*a97c2a1fSXin Li    vabdl.u8      q0, d5, d4
84*a97c2a1fSXin Li
85*a97c2a1fSXin Li    vld1.8        d7, [r1], r3
86*a97c2a1fSXin Li    vabal.u8      q0, d7, d6
87*a97c2a1fSXin Li
88*a97c2a1fSXin Li    vld1.8        d4, [r0], r2
89*a97c2a1fSXin Li    vld1.8        d5, [r1], r3
90*a97c2a1fSXin Li
91*a97c2a1fSXin Li    vld1.8        d6, [r0], r2
92*a97c2a1fSXin Li    vabal.u8      q0, d5, d4
93*a97c2a1fSXin Li
94*a97c2a1fSXin Li    vld1.8        d7, [r1], r3
95*a97c2a1fSXin Li    vabal.u8      q0, d7, d6
96*a97c2a1fSXin Li
97*a97c2a1fSXin Li    vadd.i16      d0, d1, d0
98*a97c2a1fSXin Li    vpaddl.u16    d0, d0
99*a97c2a1fSXin Li    vpaddl.u32    d0, d0
100*a97c2a1fSXin Li
101*a97c2a1fSXin Li    vmov.32       r0, d0[0]
102*a97c2a1fSXin Li
103*a97c2a1fSXin Li    pop           {pc}
104