10d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar///***************************************************************************** 20d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 30d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* Copyright (C) 2012 Ittiam Systems Pvt Ltd, Bangalore 40d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 50d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* Licensed under the Apache License, Version 2.0 (the "License"); 60d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* you may not use this file except in compliance with the License. 70d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* You may obtain a copy of the License at: 80d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 90d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* http://www.apache.org/licenses/LICENSE-2.0 100d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 110d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* Unless required by applicable law or agreed to in writing, software 120d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* distributed under the License is distributed on an "AS IS" BASIS, 130d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 140d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* See the License for the specific language governing permissions and 150d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* limitations under the License. 160d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 170d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//*****************************************************************************/ 180d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar///** 190d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//******************************************************************************* 200d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* ,:file 210d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* ihevc_sao_edge_offset_class3.s 220d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 230d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* ,:brief 240d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* Contains function definitions for inter prediction interpolation. 250d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* Functions are coded using NEON intrinsics and can be compiled using@ ARM 260d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* RVCT 270d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 280d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* ,:author 290d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* Parthiban V 300d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 310d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* ,:par List of Functions: 320d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 330d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 340d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* ,:remarks 350d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* None 360d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//* 370d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//******************************************************************************* 380d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//*/ 390d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//void ihevc_sao_edge_offset_class3(UWORD8 *pu1_src, 400d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// WORD32 src_strd, 410d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// UWORD8 *pu1_src_left, 420d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// UWORD8 *pu1_src_top, 430d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// UWORD8 *pu1_src_top_left, 440d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// UWORD8 *pu1_src_top_right, 450d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// UWORD8 *pu1_src_bot_left, 460d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// UWORD8 *pu1_avail, 470d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// WORD8 *pi1_sao_offset, 480d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// WORD32 wd, 490d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// WORD32 ht) 500d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//**************Variables Vs Registers***************************************** 510d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//x0 => *pu1_src 520d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//x1 => src_strd 530d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//x2 => *pu1_src_left 540d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//x3 => *pu1_src_top 550d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//x4 => *pu1_src_top_left 560d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//x5 => *pu1_avail 570d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//x6 => *pi1_sao_offset 580d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//x7 => wd 590d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar//x8=> ht 600d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 610d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar.text 620d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar.p2align 2 630d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 640d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar.include "ihevc_neon_macros.s" 650d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 660d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar.globl gi1_table_edge_idx 670d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar.globl ihevc_sao_edge_offset_class3_av8 680d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 690d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakarihevc_sao_edge_offset_class3_av8: 700d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 710d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 720d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar // STMFD sp!,{x4-x12,x14} //stack stores the values of the arguments 730d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar stp x19, x20,[sp,#-16]! 740d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar stp x21, x22,[sp,#-16]! 750d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar stp x23, x24,[sp,#-16]! 760d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 770d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x19,x0 //Store pu1_src in sp 780d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x21,x6 //Store pu1_src_left in sp 790d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x22,x3 //Store pu1_src_top in sp 800d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x23,x7 //Store pu1_avail in sp 810d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x24,x4 //Store pu1_src_top_left in sp 820d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,x5 //Store pu1_src_top_right in sp 830d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x13,x6 //Store pu1_src_bot_left in sp 840d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 850d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x7 //Loads pu1_avail 860d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 879cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy LDR x6,[sp,#48] //Loads pi1_sao_offset 889cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy LDR w7,[sp,#56] //Loads wd 899cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy LDR w8,[sp,#64] //Loads ht 900d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 910d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x16,x7 // wd 920d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x17,x8 // ht 930d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 940d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x9,x7,#1 //wd - 1 950d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 960d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w10,[x3,x9] //pu1_src_top[wd - 1] 970d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 980d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x9,x7 //Move width to x9 for loop count 990d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1000d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB sp,sp,#0xA0 //Decrement the stack pointer to store some temp arr values 1010d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1020d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STRB w10,[sp] //u1_src_top_left_tmp = pu1_src_top[wd - 1] 1030d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x10,x8,#1 //ht-1 1040d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar madd x11, x10, x1, x0 //pu1_src[(ht - 1) * src_strd + col] 1050d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x12,sp,#0x02 //temp array 1060d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1070d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarAU1_SRC_TOP_LOOP: 1080d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v0.8b},[x11],#8 //pu1_src[(ht - 1) * src_strd + col] 1090d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x9,x9,#8 //Decrement the loop count by 8 1100d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ST1 {v0.8b},[x12],#8 //au1_src_top_tmp[col] = pu1_src[(ht - 1) * src_strd + col] 1110d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE AU1_SRC_TOP_LOOP 1120d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1130d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarPU1_AVAIL_5_LOOP: 1140d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w9,[x5,#5] //pu1_avail[5] 1150d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x9,#0 1160d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x10,x7,#1 //[wd - 1] 1170d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w9,[x0,x10] //u1_pos_0_0_tmp = pu1_src[wd - 1] 1180d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ PU1_AVAIL_6_LOOP 1190d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1200d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x11,x20 //Load pu1_src_top_right from sp 1210d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x10,x10,#1 //[wd - 1 - 1] 1220d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1230d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w11,[x11] //pu1_src_top_right[0] 1240d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x12,x9,x11 //pu1_src[wd - 1] - pu1_src_top_right[0] 1250d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1260d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x11,x0,x1 //pu1_src + src_strd 1270d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1280d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w14,[x11,x10] //pu1_src[wd - 1 - 1 + src_strd] 1290d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x12,#0 1300d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 1310d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x12, x20, x12,LT 1320d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x11,x9,x14 //pu1_src[wd - 1] - pu1_src[wd - 1 - 1 + src_strd] 1330d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1340d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 1350d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x12, x20, x12,GT //SIGN(pu1_src[wd - 1] - pu1_src_top_right[0]) 1360d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x11,#0 1370d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 1380d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x11, x20, x11,LT 1390d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 1400d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x11, x20, x11,GT //SIGN(pu1_src[wd - 1] - pu1_src[wd - 1 - 1 + src_strd]) 1410d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADRP x14, :got:gi1_table_edge_idx //table pointer 1420d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDR x14, [x14, #:got_lo12:gi1_table_edge_idx] 1430d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x11,x12,x11 //SIGN(pu1_src[wd - 1] - pu1_src_top_right[0]) + SIGN(pu1_src[wd - 1] - pu1_src[wd - 1 - 1 + src_strd]) 1440d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x11,x11,#2 //edge_idx 1450d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1460d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRSB x12,[x14,x11] //edge_idx = gi1_table_edge_idx[edge_idx] 1470d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x12,#0 //0 != edge_idx 1480d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ PU1_AVAIL_6_LOOP 1490d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRSB x10,[x6,x12] //pi1_sao_offset[edge_idx] 1500d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x9,x9,x10 //pu1_src[0] + pi1_sao_offset[edge_idx] 1510d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar mov x20,#255 1520d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar cmp x9,x20 1530d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x9, x20, x9, ge //u1_pos_0_0_tmp = CLIP3(pu1_src[0] + pi1_sao_offset[edge_idx], 0, (1 << bit_depth) - 1) 1540d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1550d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarPU1_AVAIL_6_LOOP: 1560d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w10,[x5,#6] //pu1_avail[6] 1570d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x11,x8,#1 //ht - 1 1580d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1590d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x10,#0 1600d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar madd x12, x11, x1, x0 //pu1_src[(ht - 1) * src_strd] 1610d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1620d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w10,[x12] //u1_pos_wd_ht_tmp = pu1_src[(ht - 1) * src_strd] 1630d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ PU1_AVAIL_3_LOOP 1640d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1650d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x14,x13 //Load pu1_src_bot_left from sp 1660d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x11,x12,x1 //pu1_src[(ht - 1) * src_strd) - src_strd] 1670d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1680d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w14,[x14] //Load pu1_src_bot_left[0] 1690d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x11,x11,#1 //pu1_src[(ht - 1) * src_strd + 1 - src_strd] 1700d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1710d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w11,[x11] //Load pu1_src[(ht - 1) * src_strd + 1 - src_strd] 1720d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x14,x10,x14 //pu1_src[(ht - 1) * src_strd] - pu1_src_bot_left[0] 1730d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1740d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x11,x10,x11 //pu1_src[(ht - 1) * src_strd] - pu1_src[(ht - 1) * src_strd + 1 - src_strd] 1750d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x11,#0 1760d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 1770d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x11, x20, x11,LT 1780d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 1790d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x11, x20, x11,GT //SIGN(pu1_src[(ht - 1) * src_strd] - pu1_src[(ht - 1) * src_strd + 1 - src_strd]) 1800d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1810d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x14,#0 1820d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 1830d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x14, x20, x14,LT 1840d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 1850d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x14, x20, x14,GT //SIGN(pu1_src[(ht - 1) * src_strd] - pu1_src_bot_left[0]) 1860d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1870d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x11,x11,x14 //Add 2 sign value 1880d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1890d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADRP x14, :got:gi1_table_edge_idx //table pointer 1900d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDR x14, [x14, #:got_lo12:gi1_table_edge_idx] 1910d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x11,x11,#2 //edge_idx 1920d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 1930d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRSB x12,[x14,x11] //edge_idx = gi1_table_edge_idx[edge_idx] 1940d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x12,#0 1950d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ PU1_AVAIL_3_LOOP 1960d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRSB x11,[x6,x12] //pi1_sao_offset[edge_idx] 1970d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x10,x10,x11 //pu1_src[(ht - 1) * src_strd] + pi1_sao_offset[edge_idx] 1980d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar mov x20,#255 1990d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar cmp x10,x20 2000d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x10, x20, x10, ge //u1_pos_wd_ht_tmp = CLIP3(pu1_src[(ht - 1) * src_strd] + pi1_sao_offset[edge_idx], 0, (1 << bit_depth) - 1) 2010d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2020d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarPU1_AVAIL_3_LOOP: 2030d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x21,x2 2040d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x12,x8 //Move ht 2050d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2060d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x14,x2 //Move pu1_src_left to pu1_src_left_cpy 2070d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movi v0.16b, #2 //const_2 = vdupq_n_s8(2) 2080d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w11,[x5,#3] //pu1_avail[3] 2090d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2100d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x11,#0 2110d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movi v2.8h, #0 //const_min_clip = vdupq_n_s16(0) 2120d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x20,x12,#1 //ht_tmp-- 2130d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x12, x20, x12,EQ 2140d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2150d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5,#2] //pu1_avail[2] 2160d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movi v4.8h, #255 //const_max_clip = vdupq_n_u16((1 << bit_depth) - 1) 2170d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x5,#0 2180d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2190d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x20,x0,x1 //pu1_src += src_strd 2200d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x0, x20, x0,EQ 2210d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v7.8b},[x6] //offset_tbl = vld1_s8(pi1_sao_offset) 2220d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x20,x12,#1 //ht_tmp-- 2230d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x12, x20, x12,EQ 2240d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2250d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADRP x6, :got:gi1_table_edge_idx //table pointer 2260d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDR x6, [x6, #:got_lo12:gi1_table_edge_idx] 2270d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2289cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy movi v1.16b, #0xFF //au1_mask = vdupq_n_s8(-1) 2290d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x20,x14,#1 //pu1_src_left_cpy += 1 2300d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x14, x20, x14,EQ 2310d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2320d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x15,x0 //Store pu1_src in sp 2330d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v6.8b},[x6] //edge_idx_tbl = vld1_s8(gi1_table_edge_idx) 2340d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x6,x7 //move wd to x6 loop_count 2350d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2360d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x7,#16 //Compare wd with 16 2370d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BLT WIDTH_RESIDUE //If not jump to WIDTH_RESIDUE where loop is unrolled for 8 case 2380d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x8,#4 //Compare ht with 4 2390d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BLE WD_16_HT_4_LOOP //If jump to WD_16_HT_4_LOOP 2400d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2410d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarWIDTH_LOOP_16: 2420d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x16 //Loads wd 2430d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2440d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //Loads pu1_avail 2450d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x6,x7 //col == wd 2460d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRb w20, [x5] //pu1_avail[0] 2470d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel w8,w20,w8,EQ 2480d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#-1 2490d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,NE 250d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v1.b[0], w8 //au1_mask = vsetq_lane_s8(-1, au1_mask, 0) 2510d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2520d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x6,#16 //if(col == 16) 2530d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SKIP_AU1_MASK_VAL 2540d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x5,#1] //pu1_avail[1] 255d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v1.b[15], w8 //au1_mask = vsetq_lane_s8(pu1_avail[1], au1_mask, 15) 2560d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2570d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSKIP_AU1_MASK_VAL: 2580d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x5,#2] //pu1_avail[2] 2590d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x8,#0 2600d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2610d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x4,x17 //Loads ht 2620d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x20,x0,x1 //pu1_src - src_strd 2630d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,EQ 2640d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2650d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x3, x8,NE 2660d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x5,sp,#0x42 //*au1_src_left_tmp 2670d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2680d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x16 //Loads wd 2690d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x8,#1 //pu1_src - src_strd + 1 2700d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2710d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x7,x7,x6 //(wd - col) 2729cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy LD1 {v3.16b},[x8] //pu1_top_row = vld1q_u8(pu1_src - src_strd + 1) 2730d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x3,x3,#16 2740d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2750d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x8,x19 //Loads *pu1_src 2769cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy LD1 {v5.16b},[x0] //pu1_cur_row = vld1q_u8(pu1_src) 2770d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x7,x7,#15 //15 + (wd - col) 2780d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2790d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x7,x8,x7 //pu1_src[0 * src_strd + 15 + (wd - col)] 2809cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v17.16b, v5.16b , v3.16b //vcgtq_u8(pu1_cur_row, pu1_top_row) 2810d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x5,#1 2820d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2830d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarAU1_SRC_LEFT_LOOP: 2840d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x7] //load the value and increment by src_strd 2850d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x7,x7,x1 2860d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x4,x4,#1 //decrement the loop count 2870d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STRB w8,[x5,#1]! //store it in the stack pointer 2880d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE AU1_SRC_LEFT_LOOP 2890d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2900d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movi v18.16b, #0 2919cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v16.16b, v3.16b , v5.16b //vcltq_u8(pu1_cur_row, pu1_top_row) 2920d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2930d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x0,x1 //I *pu1_src + src_strd 2949cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy SUB v17.16b, v16.16b , v17.16b //sign_up = vreinterpretq_s8_u8(vsubq_u8(cmp_lt, cmp_gt)) 2950d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x12 //row count, move ht_tmp to x7 2960d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 2970d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x12,x7 //I ht_tmp - row 2980d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v16.16b},[x8] //I pu1_next_row = vld1q_u8(pu1_src_cpy + src_strd) 2990d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x14,x5 //I pu1_src_left_cpy[ht_tmp - row] 3000d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3010d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x8,#1 //I pu1_src_left_cpy[ht_tmp - row + 1] 3020d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x8] 3030d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3040d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //I Loads pu1_avail 305d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v18.b[15], w8 //I vsetq_lane_u8 3060d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5,#2] //I pu1_avail[2] 3070d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3080d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar EXT v18.16b, v18.16b , v16.16b,#15 //I pu1_next_row_tmp = vextq_u8(pu1_next_row_tmp, pu1_next_row, 15) 3090d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x5,#0 //I 3100d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SIGN_UP_CHANGE_DONE //I 3110d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3120d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSIGN_UP_CHANGE: 3130d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x0,#15] //I pu1_src_cpy[15] 3140d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x0,x1 //I pu1_src_cpy[16 - src_strd] 3150d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3160d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5,#16] //I load the value 3170d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x8,x8,x5 //I pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd] 3180d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x8,#0 //I 3190d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 3200d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,LT //I 3210d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 3220d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,GT //I SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]) 323d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v17.b[15], w8 //I sign_up = vsetq_lane_s8(SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]), sign_up, 15) 3240d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3250d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSIGN_UP_CHANGE_DONE: 3269cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v3.16b, v5.16b , v18.16b //I vcgtq_u8(pu1_cur_row, pu1_next_row_tmp) 3279cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v18.16b, v18.16b , v5.16b //I vcltq_u8(pu1_cur_row, pu1_next_row_tmp) 3289cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy SUB v3.16b, v18.16b , v3.16b //I sign_down = vreinterpretq_s8_u8(vsubq_u8(cmp_lt, cmp_gt)) 3290d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3309cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy ADD v18.16b, v0.16b , v17.16b //I edge_idx = vaddq_s8(const_2, sign_up) 3319cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy ADD v18.16b, v18.16b , v3.16b //I edge_idx = vaddq_s8(edge_idx, sign_down) 3320d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v18.16b, {v6.16b},v18.16b //I vtbl1_s8(edge_idx_tbl, vget_low_s8(edge_idx)) 3339cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy NEG v17.16b, v3.16b //I sign_up = vnegq_s8(sign_down) 3340d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3359cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy EXT v17.16b, v17.16b , v17.16b,#1 //I sign_up = vextq_s8(sign_up, sign_up, 1) 3360d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v19.8b, {v6.16b},v19.8b //I vtbl1_s8(edge_idx_tbl, vget_high_s8(edge_idx)) 3370d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3389cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy Uxtl v20.8h, v5.8b //I pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pu1_cur_row))) 3399cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy AND v18.16b, v18.16b , v1.16b //I edge_idx = vandq_s8(edge_idx, au1_mask) 3400d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3419cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy TBL v3.16b, {v7.16b},v18.16b //I offset = vtbl1_s8(offset_tbl, vget_low_s8(edge_idx)) 3420d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3439cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy Uxtl2 v22.8h, v5.16b //I pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pu1_cur_row))) 3449cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy SADDW v20.8h, v20.8h , v3.8b //I pi2_tmp_cur_row.val[0] = vaddw_s8(pi2_tmp_cur_row.val[0], offset) 3450d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3460d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v20.8h, v20.8h , v2.8h //I pi2_tmp_cur_row.val[0] = vmaxq_s16(pi2_tmp_cur_row.val[0], const_min_clip) 3470d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v11.8b, {v7.16b},v19.8b //I offset = vtbl1_s8(offset_tbl, vget_high_s8(edge_idx)) 3480d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v20.8h, v20.8h , v4.8h //I pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[0]), const_max_clip)) 3490d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3509cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy mov v5.16b, v16.16b 3519cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy SADDW2 v22.8h, v22.8h , v3.16b //I pi2_tmp_cur_row.val[1] = vaddw_s8(pi2_tmp_cur_row.val[1], offset) 3520d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3530d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v22.8h, v22.8h , v2.8h //I pi2_tmp_cur_row.val[1] = vmaxq_s16(pi2_tmp_cur_row.val[1], const_min_clip) 3540d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v22.8h, v22.8h , v4.8h //I pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[1]), const_max_clip)) 3550d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3560d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x7,x7,#1 //I Decrement the ht_tmp loop count by 1 3570d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3580d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarPU1_SRC_LOOP: 3590d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x0,x1,LSL #1 //II *pu1_src + src_strd 3600d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn v20.8b, v20.8h //I vmovn_s16(pi2_tmp_cur_row.val[0]) 3610d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x12,x7 //II ht_tmp - row 3620d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3630d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x4,x0,x1 //II pu1_src_cpy[16 - src_strd] 3640d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn2 v20.16b, v22.8h //I vmovn_s16(pi2_tmp_cur_row.val[1]) 3650d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x2,x8,x1 //III *pu1_src + src_strd 3660d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3670d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w11,[x4,#15] //II pu1_src_cpy[15] 3680d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v16.16b},[x8] //II pu1_next_row = vld1q_u8(pu1_src_cpy + src_strd) 3690d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x7,x7,#1 //II Decrement the ht_tmp loop count by 1 3700d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3710d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x14,x5 //II pu1_src_left_cpy[ht_tmp - row] 3720d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v30.16b},[x2] //III pu1_next_row = vld1q_u8(pu1_src_cpy + src_strd) 3730d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x8,#1] 3740d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3750d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w4,[x0,#16] //II load the value 376d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v18.b[15], w8 //II vsetq_lane_u8 3770d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x11,x11,x4 //II pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd] 3780d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3790d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x11,#0 //II 3800d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ST1 { v20.16b},[x0],x1 //I vst1q_u8(pu1_src_cpy, pu1_cur_row) 3810d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x12,x7 //III ht_tmp - row 3820d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3830d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 3840d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x11, x20, x11,LT //II 3850d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar EXT v18.16b, v18.16b , v16.16b,#15 //II pu1_next_row_tmp = vextq_u8(pu1_next_row_tmp, pu1_next_row, 15) 3860d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 3870d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x11, x20, x11,GT //II SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]) 3880d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3890d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x14,x5 //III pu1_src_left_cpy[ht_tmp - row] 390d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v17.b[15], w11 //II sign_up = vsetq_lane_s8(SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]), sign_up, 15) 3910d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x7,#1 //III 3920d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 3930d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE NEXT_ROW_ELSE_2 //III 3940d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //III Loads pu1_avail 3950d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5,#3] //III pu1_avail[3] 3960d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x5,#0 //III 3970d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x20,x2,#2 //III pu1_src_cpy[src_strd - 1] 3980d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,NE 3990d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4000d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarNEXT_ROW_ELSE_2: 4010d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x8,#1] //III 4029cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v24.16b, v5.16b , v18.16b //II vcgtq_u8(pu1_cur_row, pu1_next_row_tmp) 4030d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x5,x0,x1 4040d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4050d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w2,[x5,#15] //III pu1_src_cpy[15] 4069cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v26.16b, v18.16b , v5.16b //II vcltq_u8(pu1_cur_row, pu1_next_row_tmp) 4070d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x0,#16] //III load the value 4080d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4090d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x2,x2,x5 //III pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd] 4100d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB v24.16b, v26.16b , v24.16b //II sign_down = vreinterpretq_s8_u8(vsubq_u8(cmp_lt, cmp_gt)) 4110d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x2,#0 //III 4120d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4130d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 4140d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x2, x20, x2,LT //III 415d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v18.b[15], w8 //III vsetq_lane_u8 4160d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 4170d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x2, x20, x2,GT //III SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]) 4180d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4190d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x7,x7,#1 //III Decrement the ht_tmp loop count by 1 4209cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy ADD v26.16b, v0.16b , v17.16b //II edge_idx = vaddq_s8(const_2, sign_up) 4210d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4229cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy NEG v17.16b, v24.16b //II sign_up = vnegq_s8(sign_down) 4230d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar EXT v18.16b, v18.16b , v30.16b,#15 //III pu1_next_row_tmp = vextq_u8(pu1_next_row_tmp, pu1_next_row, 15) 4240d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4250d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD v26.16b, v26.16b , v24.16b //II edge_idx = vaddq_s8(edge_idx, sign_down) 4260d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4279cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy EXT v17.16b, v17.16b , v17.16b,#1 //II sign_up = vextq_s8(sign_up, sign_up, 1) 4280d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v26.16b, {v6.16b},v26.16b //II vtbl1_s8(edge_idx_tbl, vget_low_s8(edge_idx)) 4299cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v3.16b, v16.16b , v18.16b //III vcgtq_u8(pu1_cur_row, pu1_next_row_tmp) 4300d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 431d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v17.b[15], w2 //III sign_up = vsetq_lane_s8(SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]), sign_up, 15) 4320d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v27.8b, {v6.16b},v27.8b //II vtbl1_s8(edge_idx_tbl, vget_high_s8(edge_idx)) 4330d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar cmhi v18.16b, v18.16b , v16.16b //III vcltq_u8(pu1_cur_row, pu1_next_row_tmp) 4340d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4359cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy Uxtl v28.8h, v5.8b //II pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pu1_cur_row))) 4369cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy AND v26.16b, v26.16b , v1.16b //II edge_idx = vandq_s8(edge_idx, au1_mask) 4370d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4389cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy SUB v3.16b, v18.16b , v3.16b //III sign_down = vreinterpretq_s8_u8(vsubq_u8(cmp_lt, cmp_gt)) 4390d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v24.16b, {v7.16b},v26.16b //II offset = vtbl1_s8(offset_tbl, vget_low_s8(edge_idx)) 4409cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy ADD v18.16b, v0.16b , v17.16b //III edge_idx = vaddq_s8(const_2, sign_up) 4410d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4429cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy ADD v18.16b, v18.16b , v3.16b //III edge_idx = vaddq_s8(edge_idx, sign_down) 4430d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v25.8b, {v7.16b},v27.8b //II offset = vtbl1_s8(offset_tbl, vget_high_s8(edge_idx)) 4449cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy NEG v17.16b, v3.16b //III sign_up = vnegq_s8(sign_down) 4450d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4460d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SADDW v28.8h, v28.8h , v24.8b //II pi2_tmp_cur_row.val[0] = vaddw_s8(pi2_tmp_cur_row.val[0], offset) 4470d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v18.16b, {v6.16b},v18.16b //III vtbl1_s8(edge_idx_tbl, vget_low_s8(edge_idx)) 4480d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v28.8h, v28.8h , v2.8h //II pi2_tmp_cur_row.val[0] = vmaxq_s16(pi2_tmp_cur_row.val[0], const_min_clip) 4490d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4509cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy EXT v17.16b, v17.16b , v17.16b,#1 //III sign_up = vextq_s8(sign_up, sign_up, 1) 4510d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v19.8b, {v6.16b},v19.8b //III vtbl1_s8(edge_idx_tbl, vget_high_s8(edge_idx)) 4520d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v28.8h, v28.8h , v4.8h //II pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[0]), const_max_clip)) 4530d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4549cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy Uxtl2 v26.8h, v5.16b //II pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pu1_cur_row))) 4559cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy AND v18.16b, v18.16b , v1.16b //III edge_idx = vandq_s8(edge_idx, au1_mask) 4560d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4570d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SADDW2 v26.8h, v26.8h , v24.16b //II pi2_tmp_cur_row.val[1] = vaddw_s8(pi2_tmp_cur_row.val[1], offset) 4589cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy TBL v3.16b, {v7.16b},v18.16b //III offset = vtbl1_s8(offset_tbl, vget_low_s8(edge_idx)) 4590d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v26.8h, v26.8h , v2.8h //II pi2_tmp_cur_row.val[1] = vmaxq_s16(pi2_tmp_cur_row.val[1], const_min_clip) 4600d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4610d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar Uxtl v20.8h, v16.8b //III pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pu1_cur_row))) 4620d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v26.8h, v26.8h , v4.8h //II pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[1]), const_max_clip)) 4630d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4649cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy SADDW v20.8h, v20.8h , v3.8b //III pi2_tmp_cur_row.val[0] = vaddw_s8(pi2_tmp_cur_row.val[0], offset) 4650d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v11.8b, {v7.16b},v19.8b //III offset = vtbl1_s8(offset_tbl, vget_high_s8(edge_idx)) 4660d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v20.8h, v20.8h , v2.8h //III pi2_tmp_cur_row.val[0] = vmaxq_s16(pi2_tmp_cur_row.val[0], const_min_clip) 4670d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4680d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar Uxtl2 v22.8h, v16.16b //III pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pu1_cur_row))) 4690d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v20.8h, v20.8h , v4.8h //III pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[0]), const_max_clip)) 4700d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4710d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn v28.8b, v28.8h //II vmovn_s16(pi2_tmp_cur_row.val[0]) 4729cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy SADDW2 v22.8h, v22.8h , v3.16b //III pi2_tmp_cur_row.val[1] = vaddw_s8(pi2_tmp_cur_row.val[1], offset) 4730d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4740d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn2 v28.16b, v26.8h //II vmovn_s16(pi2_tmp_cur_row.val[1]) 4750d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v22.8h, v22.8h , v2.8h //III pi2_tmp_cur_row.val[1] = vmaxq_s16(pi2_tmp_cur_row.val[1], const_min_clip) 4760d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4779cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy mov v5.16b, v30.16b //II pu1_cur_row = pu1_next_row 4780d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v22.8h, v22.8h , v4.8h //III pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[1]), const_max_clip)) 4790d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4800d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x7,#1 //III 4810d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ST1 { v28.16b},[x0],x1 //II vst1q_u8(pu1_src_cpy, pu1_cur_row) 4820d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BGT PU1_SRC_LOOP //If not equal jump to PU1_SRC_LOOP 4830d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BLT INNER_LOOP_DONE 4840d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4850d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x0,x1,LSL #1 //*pu1_src + src_strd 4860d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn v20.8b, v20.8h //III vmovn_s16(pi2_tmp_cur_row.val[0]) 4870d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //Loads pu1_avail 4880d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4890d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5,#3] //pu1_avail[3] 4900d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn2 v20.16b, v22.8h //III vmovn_s16(pi2_tmp_cur_row.val[1]) 4910d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x5,#0 4920d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4930d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x4,x0,x1 //pu1_src_cpy[16 - src_strd] 4940d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v16.16b},[x8] //pu1_next_row = vld1q_u8(pu1_src_cpy + src_strd) 4950d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x0,#16] //load the value 4960d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 4970d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ NEXT_ROW_ELSE_3 4980d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x8,x8,#1 4990d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x8] //pu1_src_cpy[src_strd - 1] 5000d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar B NEXT_ROW_POINTER_ASSIGNED_3 5010d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarNEXT_ROW_ELSE_3: 5020d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x11,x12,x7 //ht_tmp - row 5030d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x14,x11 //pu1_src_left_cpy[ht_tmp - row] 5040d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x8,#1 //pu1_src_left_cpy[ht_tmp - row + 1] 5050d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x8] 5060d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5070d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarNEXT_ROW_POINTER_ASSIGNED_3: 5080d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w11,[x4,#15] //pu1_src_cpy[15] 509d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v18.b[15], w8 //vsetq_lane_u8 5100d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x8,x11,x5 //pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd] 5110d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5120d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x8,#0 5130d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar EXT v18.16b, v18.16b , v16.16b,#15 //pu1_next_row_tmp = vextq_u8(pu1_next_row_tmp, pu1_next_row, 15) 5140d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 5150d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,LT 5160d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5170d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ST1 { v20.16b},[x0],x1 //III vst1q_u8(pu1_src_cpy, pu1_cur_row) 5189cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v24.16b, v5.16b , v18.16b //vcgtq_u8(pu1_cur_row, pu1_next_row_tmp) 5190d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5200d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 5210d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,GT //SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]) 5229cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v26.16b, v18.16b , v5.16b //vcltq_u8(pu1_cur_row, pu1_next_row_tmp) 5230d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 524d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v17.b[15], w8 //sign_up = vsetq_lane_s8(SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]), sign_up, 15) 5250d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB v24.16b, v26.16b , v24.16b //sign_down = vreinterpretq_s8_u8(vsubq_u8(cmp_lt, cmp_gt)) 5260d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5279cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy Uxtl v20.8h, v5.8b //pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pu1_cur_row))) 5289cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy ADD v26.16b, v0.16b , v17.16b //edge_idx = vaddq_s8(const_2, sign_up) 5290d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5309cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy Uxtl2 v22.8h, v5.16b //pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pu1_cur_row))) 5310d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD v26.16b, v26.16b , v24.16b //edge_idx = vaddq_s8(edge_idx, sign_down) 5320d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5330d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v26.16b, {v6.16b},v26.16b //vtbl1_s8(edge_idx_tbl, vget_low_s8(edge_idx)) 5340d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v27.8b, {v6.16b},v27.8b //vtbl1_s8(edge_idx_tbl, vget_high_s8(edge_idx)) 5350d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5369cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy AND v26.16b, v26.16b , v1.16b //edge_idx = vandq_s8(edge_idx, au1_mask) 5370d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5380d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v24.16b, {v7.16b},v26.16b //offset = vtbl1_s8(offset_tbl, vget_low_s8(edge_idx)) 5390d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5400d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SADDW v20.8h, v20.8h , v24.8b //pi2_tmp_cur_row.val[0] = vaddw_s8(pi2_tmp_cur_row.val[0], offset) 5410d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v25.8b, {v7.16b},v27.8b //offset = vtbl1_s8(offset_tbl, vget_high_s8(edge_idx)) 5420d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v20.8h, v20.8h , v2.8h //pi2_tmp_cur_row.val[0] = vmaxq_s16(pi2_tmp_cur_row.val[0], const_min_clip) 5430d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5440d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v20.8h, v20.8h , v4.8h //pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[0]), const_max_clip)) 5450d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5460d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SADDW2 v22.8h, v22.8h , v24.16b //pi2_tmp_cur_row.val[1] = vaddw_s8(pi2_tmp_cur_row.val[1], offset) 5470d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v22.8h, v22.8h , v2.8h //pi2_tmp_cur_row.val[1] = vmaxq_s16(pi2_tmp_cur_row.val[1], const_min_clip) 5480d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v22.8h, v22.8h , v4.8h //pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[1]), const_max_clip)) 5490d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5500d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarINNER_LOOP_DONE: 5510d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn v20.8b, v20.8h //vmovn_s16(pi2_tmp_cur_row.val[0]) 5520d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x8,x17 //Loads ht 5530d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5540d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn2 v20.16b, v22.8h //vmovn_s16(pi2_tmp_cur_row.val[1]) 5550d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x5,sp,#0x42 //*au1_src_left_tmp 5560d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5570d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ST1 { v20.16b},[x0],x1 //vst1q_u8(pu1_src_cpy, pu1_cur_row) 5580d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x2,x21 //Loads *pu1_src_left 5590d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSRC_LEFT_LOOP: 5600d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDR w7,[x5],#4 //au1_src_left_tmp[row] 5610d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x8,x8,#4 5620d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STR w7,[x2],#4 //pu1_src_left[row] = au1_src_left_tmp[row] 5630d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SRC_LEFT_LOOP 5640d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5650d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x6,x6,#16 //Decrement the wd loop count by 16 5660d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x6,#8 //Check whether residue remains 5670d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BLT RE_ASSINING_LOOP //Jump to re-assigning loop 5680d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x16 //Loads wd 5690d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x0,x15 //Loads *pu1_src 5700d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x7,x7,x6 5710d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x0,x0,x7 5720d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BGT WIDTH_LOOP_16 //If not equal jump to width_loop 5730d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ WIDTH_RESIDUE //If residue remains jump to residue loop 5740d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5750d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5760d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5770d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarWD_16_HT_4_LOOP: 5780d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //Loads pu1_avail 5790d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x16 //Loads wd 5800d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x6,x7 //col == wd 5810d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRb w20, [x5] //pu1_avail[0] 5820d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel w8,w20,w8,EQ 5830d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#-1 5840d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,NE 585d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v1.b[0], w8 //au1_mask = vsetq_lane_s8(-1, au1_mask, 0) 5860d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5870d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x6,#16 //if(col == 16) 5880d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SKIP_AU1_MASK_VAL_WD_16_HT_4 5890d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x5,#1] //pu1_avail[1] 590d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v1.b[15], w8 //au1_mask = vsetq_lane_s8(pu1_avail[1], au1_mask, 15) 5910d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5920d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSKIP_AU1_MASK_VAL_WD_16_HT_4: 5930d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x5,#2] //pu1_avail[2] 5940d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x8,#0 5950d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 5960d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x20,x0,x1 //pu1_src - src_strd 5970d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,EQ 5980d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x3, x8,NE 5990d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x8,#1 //pu1_src - src_strd + 1 6009cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy LD1 {v3.16b},[x8] //pu1_top_row = vld1q_u8(pu1_src - src_strd + 1) 6010d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6020d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x3,x3,#16 6030d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x5,sp,#0x42 //*au1_src_left_tmp 6040d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x4,x17 //Loads ht 6050d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x16 //Loads wd 6060d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x7,x7,x6 //(wd - col) 6070d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x7,x7,#15 //15 + (wd - col) 6080d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x8,x19 //Loads *pu1_src 6090d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x7,x8,x7 //pu1_src[0 * src_strd + 15 + (wd - col)] 6100d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x5,#1 6110d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6120d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarAU1_SRC_LEFT_LOOP_WD_16_HT_4: 6130d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x7] //load the value and increment by src_strd 6140d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x7,x7,x1 6150d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STRB w8,[x5,#1]! //store it in the stack pointer 6160d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x4,x4,#1 //decrement the loop count 6170d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE AU1_SRC_LEFT_LOOP_WD_16_HT_4 6180d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6199cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy LD1 {v5.16b},[x0] //pu1_cur_row = vld1q_u8(pu1_src) 6200d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6219cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v17.16b, v5.16b , v3.16b //vcgtq_u8(pu1_cur_row, pu1_top_row) 6229cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v16.16b, v3.16b , v5.16b //vcltq_u8(pu1_cur_row, pu1_top_row) 6239cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy SUB v17.16b, v16.16b , v17.16b //sign_up = vreinterpretq_s8_u8(vsubq_u8(cmp_lt, cmp_gt)) 6240d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movi v18.16b, #0 6250d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x12 //row count, move ht_tmp to x7 6260d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6270d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarPU1_SRC_LOOP_WD_16_HT_4: 6280d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x0,x1 //*pu1_src + src_strd 6290d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v16.16b},[x8] //pu1_next_row = vld1q_u8(pu1_src_cpy + src_strd) 6300d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //Loads pu1_avail 6310d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5,#3] //pu1_avail[3] 6320d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x5,#0 6330d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ NEXT_ROW_ELSE_WD_16_HT_4 6340d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x7,#1 6350d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x8,x8,#1 6360d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRb w20, [x8] //pu1_src_cpy[src_strd - 1] 6370d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel w8,w20,w8,EQ 6380d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ NEXT_ROW_POINTER_ASSIGNED_WD_16_HT_4 6390d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarNEXT_ROW_ELSE_WD_16_HT_4: 6400d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x12,x7 //ht_tmp - row 6410d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x14,x5 //pu1_src_left_cpy[ht_tmp - row] 6420d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x8,#1 //pu1_src_left_cpy[ht_tmp - row + 1] 6430d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x8] 6440d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6450d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarNEXT_ROW_POINTER_ASSIGNED_WD_16_HT_4: 646d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v18.b[15], w8 //vsetq_lane_u8 6470d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar EXT v18.16b, v18.16b , v16.16b,#15 //pu1_next_row_tmp = vextq_u8(pu1_next_row_tmp, pu1_next_row, 15) 6480d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6490d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x7,x12 6500d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SIGN_UP_CHANGE_WD_16_HT_4 6510d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //Loads pu1_avail 6520d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5,#2] //pu1_avail[2] 6530d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x5,#0 6540d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SIGN_UP_CHANGE_DONE_WD_16_HT_4 6550d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6560d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSIGN_UP_CHANGE_WD_16_HT_4: 6570d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x0,#15] //pu1_src_cpy[15] 6580d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x5,x0,#16 //pu1_src_cpy[16] 6590d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x5,x1 //pu1_src_cpy[16 - src_strd] 6600d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5] //load the value 6610d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x8,x8,x5 //pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd] 6620d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x8,#0 6630d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 6640d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,LT 6650d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 6660d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,GT //SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]) 667d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v17.b[15], w8 //sign_up = vsetq_lane_s8(SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]), sign_up, 15) 6680d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6690d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSIGN_UP_CHANGE_DONE_WD_16_HT_4: 6709cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v20.16b, v5.16b , v18.16b //vcgtq_u8(pu1_cur_row, pu1_next_row_tmp) 6719cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v22.16b, v18.16b , v5.16b //vcltq_u8(pu1_cur_row, pu1_next_row_tmp) 6720d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB v24.16b, v22.16b , v20.16b //sign_down = vreinterpretq_s8_u8(vsubq_u8(cmp_lt, cmp_gt)) 6730d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6749cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy ADD v26.16b, v0.16b , v17.16b //edge_idx = vaddq_s8(const_2, sign_up) 6750d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD v26.16b, v26.16b , v24.16b //edge_idx = vaddq_s8(edge_idx, sign_down) 6760d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v26.16b, {v6.16b},v26.16b //vtbl1_s8(edge_idx_tbl, vget_low_s8(edge_idx)) 6770d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v27.8b, {v6.16b},v27.8b //vtbl1_s8(edge_idx_tbl, vget_high_s8(edge_idx)) 6780d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6799cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy AND v26.16b, v26.16b , v1.16b //edge_idx = vandq_s8(edge_idx, au1_mask) 6800d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6819cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy NEG v17.16b, v24.16b //sign_up = vnegq_s8(sign_down) 6829cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy EXT v17.16b, v17.16b , v17.16b,#1 //sign_up = vextq_s8(sign_up, sign_up, 1) 6830d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6840d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v24.16b, {v7.16b},v26.16b //offset = vtbl1_s8(offset_tbl, vget_low_s8(edge_idx)) 6859cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy Uxtl v28.8h, v5.8b //pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pu1_cur_row))) 6860d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SADDW v28.8h, v28.8h , v24.8b //pi2_tmp_cur_row.val[0] = vaddw_s8(pi2_tmp_cur_row.val[0], offset) 6870d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v28.8h, v28.8h , v2.8h //pi2_tmp_cur_row.val[0] = vmaxq_s16(pi2_tmp_cur_row.val[0], const_min_clip) 6880d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v28.8h, v28.8h , v4.8h //pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[0]), const_max_clip)) 6890d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6900d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v25.8b, {v7.16b},v27.8b //offset = vtbl1_s8(offset_tbl, vget_high_s8(edge_idx)) 6919cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy Uxtl2 v30.8h, v5.16b //pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pu1_cur_row))) 6920d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SADDW2 v30.8h, v30.8h , v24.16b //pi2_tmp_cur_row.val[1] = vaddw_s8(pi2_tmp_cur_row.val[1], offset) 6930d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v30.8h, v30.8h , v2.8h //pi2_tmp_cur_row.val[1] = vmaxq_s16(pi2_tmp_cur_row.val[1], const_min_clip) 6940d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v30.8h, v30.8h , v4.8h //pi2_tmp_cur_row.val[1] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[1]), const_max_clip)) 6950d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6960d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn v28.8b, v28.8h //vmovn_s16(pi2_tmp_cur_row.val[0]) 6970d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn2 v28.16b, v30.8h //vmovn_s16(pi2_tmp_cur_row.val[1]) 6980d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 6990d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ST1 { v28.16b},[x0],x1 //vst1q_u8(pu1_src_cpy, pu1_cur_row) 7000d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7019cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy mov v5.16b, v16.16b //pu1_cur_row = pu1_next_row 7020d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x7,x7,#1 //Decrement the ht_tmp loop count by 1 7030d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE PU1_SRC_LOOP_WD_16_HT_4 //If not equal jump to PU1_SRC_LOOP_WD_16_HT_4 7040d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7050d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x8,x17 //Loads ht 7060d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x5,sp,#0x42 //*au1_src_left_tmp 7070d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x2,x21 //Loads *pu1_src_left 7080d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSRC_LEFT_LOOP_WD_16_HT_4: 7090d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDR w7,[x5],#4 //au1_src_left_tmp[row] 7100d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STR w7,[x2],#4 //pu1_src_left[row] = au1_src_left_tmp[row] 7110d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x8,x8,#4 7120d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SRC_LEFT_LOOP_WD_16_HT_4 7130d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7140d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x6,x6,#16 //Decrement the wd loop count by 16 7150d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BLE RE_ASSINING_LOOP //Jump to re-assigning loop 7160d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BGT WD_16_HT_4_LOOP //If not equal jump to width_loop 7170d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7180d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7190d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarWIDTH_RESIDUE: 7200d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x16 //Loads wd 7210d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //Loads pu1_avail 7220d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x6,x7 //wd_residue == wd 7230d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRb w20, [x5] //pu1_avail[0] 7240d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel w8,w20,w8,EQ 7250d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7260d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#-1 7270d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,NE 728d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v1.b[0], w8 //au1_mask = vsetq_lane_s8(-1, au1_mask, 0) 7290d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7300d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x5,#1] //pu1_avail[1] 731d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v1.b[7], w8 //au1_mask = vsetq_lane_s8(pu1_avail[1], au1_mask, 15) 7320d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7330d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarPU1_AVAIL_2_RESIDUE: 7340d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x5,#2] //pu1_avail[2] 7350d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x8,#0 7360d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7370d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x20,x0,x1 //pu1_src - src_strd 7380d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,EQ 7390d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x3, x8,NE 7400d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x8,#1 //pu1_src - src_strd + 1 7419cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy LD1 {v3.16b},[x8] //pu1_top_row = vld1q_u8(pu1_src - src_strd + 1) 7420d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7430d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7440d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x5,sp,#0x42 //*au1_src_left_tmp 7450d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x4,x17 //Loads ht 7460d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x16 //Loads wd 7470d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x8,x19 //Loads *pu1_src 7480d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x7,x7,#1 //(wd - 1) 7490d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x7,x8,x7 //pu1_src[0 * src_strd + (wd - 1)] 7500d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x5,#1 7510d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7520d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarAU1_SRC_LEFT_LOOP_RESIDUE: 7530d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x7] //load the value and increment by src_strd 7540d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x7,x7,x1 7550d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STRB w8,[x5,#1]! //store it in the stack pointer 7560d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x4,x4,#1 //decrement the loop count 7570d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE AU1_SRC_LEFT_LOOP_RESIDUE 7580d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7599cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy LD1 {v5.16b},[x0] //pu1_cur_row = vld1q_u8(pu1_src) 7600d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7619cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v17.16b, v5.16b , v3.16b //vcgtq_u8(pu1_cur_row, pu1_top_row) 7629cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v16.16b, v3.16b , v5.16b //vcltq_u8(pu1_cur_row, pu1_top_row) 7639cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy SUB v17.16b, v16.16b , v17.16b //sign_up = vreinterpretq_s8_u8(vsubq_u8(cmp_lt, cmp_gt)) 7640d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x12 //row count, move ht_tmp to x7 7650d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7660d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarPU1_SRC_LOOP_RESIDUE: 7670d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movi v18.16b, #0 7680d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x0,x1 //*pu1_src + src_strd 7690d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v16.16b},[x8] //pu1_next_row = vld1q_u8(pu1_src_cpy + src_strd) 7700d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //Loads pu1_avail 7710d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5,#3] //pu1_avail[3] 7720d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x5,#0 7730d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ NEXT_ROW_ELSE_RESIDUE 7740d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x7,#1 7750d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x8,x8,#1 7760d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRb w20, [x8] //pu1_src_cpy[src_strd - 1] 7770d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel w8,w20,w8,EQ 7780d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BEQ NEXT_ROW_POINTER_ASSIGNED_RESIDUE 7790d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarNEXT_ROW_ELSE_RESIDUE: 7800d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x12,x7 //ht_tmp - row 7810d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x14,x5 //pu1_src_left_cpy[ht_tmp - row] 7820d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x8,#1 //pu1_src_left_cpy[ht_tmp - row + 1] 7830d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x8] 7840d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7850d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarNEXT_ROW_POINTER_ASSIGNED_RESIDUE: 786d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v18.b[15], w8 //vsetq_lane_u8 7870d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar EXT v18.16b, v18.16b , v16.16b,#15 //pu1_next_row_tmp = vextq_u8(pu1_next_row_tmp, pu1_next_row, 15) 7880d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7890d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x7,x12 7900d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SIGN_UP_CHANGE_RESIDUE 7910d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x5,x23 //Loads pu1_avail 7920d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5,#2] //pu1_avail[2] 7930d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x5,#0 7940d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SIGN_UP_CHANGE_DONE_RESIDUE 7950d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 7960d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSIGN_UP_CHANGE_RESIDUE: 7970d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[x0,#15] //pu1_src_cpy[15] 7980d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x5,x0,#16 //pu1_src_cpy[16] 7990d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x5,x5,x1 //pu1_src_cpy[16 - src_strd] 8000d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w5,[x5] //load the value 8010d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x8,x8,x5 //pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd] 8020d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar CMP x8,#0 8030d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar movn x20,#0 8040d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,LT 8050d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x20,#1 8060d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar csel x8, x20, x8,GT //SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]) 807d91eedb8cdcdd3d4f23379517752d48fa5791604Bernhard Rosenkränzer mov v17.b[15], w8 //sign_up = vsetq_lane_s8(SIGN(pu1_src_cpy[15] - pu1_src_cpy[16 - src_strd]), sign_up, 15) 8080d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8090d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSIGN_UP_CHANGE_DONE_RESIDUE: 8109cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v20.16b, v5.16b , v18.16b //vcgtq_u8(pu1_cur_row, pu1_next_row_tmp) 8119cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy cmhi v22.16b, v18.16b , v5.16b //vcltq_u8(pu1_cur_row, pu1_next_row_tmp) 8120d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB v24.16b, v22.16b , v20.16b //sign_down = vreinterpretq_s8_u8(vsubq_u8(cmp_lt, cmp_gt)) 8130d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8149cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy ADD v26.16b, v0.16b , v17.16b //edge_idx = vaddq_s8(const_2, sign_up) 8150d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD v26.16b, v26.16b , v24.16b //edge_idx = vaddq_s8(edge_idx, sign_down) 8160d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v26.16b, {v6.16b},v26.16b //vtbl1_s8(edge_idx_tbl, vget_low_s8(edge_idx)) 8170d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar// TBL v27.8b, {v6.16b},v27.8b //vtbl1_s8(edge_idx_tbl, vget_high_s8(edge_idx)) 8180d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8199cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy AND v26.16b, v26.16b , v1.16b //edge_idx = vandq_s8(edge_idx, au1_mask) 8200d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8219cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy NEG v17.16b, v24.16b //sign_up = vnegq_s8(sign_down) 8229cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy EXT v17.16b, v17.16b , v17.16b,#1 //sign_up = vextq_s8(sign_up, sign_up, 1) 8230d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8240d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar TBL v24.8b, {v7.16b},v26.8b //offset = vtbl1_s8(offset_tbl, vget_low_s8(edge_idx)) 8259cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy Uxtl v28.8h, v5.8b //pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pu1_cur_row))) 8260d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SADDW v28.8h, v28.8h , v24.8b //pi2_tmp_cur_row.val[0] = vaddw_s8(pi2_tmp_cur_row.val[0], offset) 8270d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SMAX v28.8h, v28.8h , v2.8h //pi2_tmp_cur_row.val[0] = vmaxq_s16(pi2_tmp_cur_row.val[0], const_min_clip) 8280d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar UMIN v28.8h, v28.8h , v4.8h //pi2_tmp_cur_row.val[0] = vreinterpretq_s16_u16(vminq_u16(vreinterpretq_u16_s16(pi2_tmp_cur_row.val[0]), const_max_clip)) 8290d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8300d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar xtn v30.8b, v28.8h //vmovn_s16(pi2_tmp_cur_row.val[0]) 8310d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8320d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ST1 {v30.8b},[x0],x1 //vst1q_u8(pu1_src_cpy, pu1_cur_row) 8339cbd70a2930875be59d7df68136ac9a1a949a13dNaveen Kumar Ponnusamy mov v5.16b, v16.16b //pu1_cur_row = pu1_next_row 8340d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x7,x7,#1 8350d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE PU1_SRC_LOOP_RESIDUE 8360d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8370d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x8,x17 //Loads ht 8380d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x2,x21 //Loads *pu1_src_left 8390d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x5,sp,#0x42 //*au1_src_left_tmp 8400d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8410d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSRC_LEFT_LOOP_RESIDUE: 8420d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDR w7,[x5],#4 //au1_src_left_tmp[row] 8430d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x8,x8,#4 8440d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STR w7,[x2],#4 //pu1_src_left[row] = au1_src_left_tmp[row] 8450d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SRC_LEFT_LOOP_RESIDUE 8460d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8470d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8480d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarRE_ASSINING_LOOP: 8490d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x7,x16 //Loads wd 8500d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x0,x19 //Loads *pu1_src 8510d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8520d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x11,x17 //Loads ht 8530d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x0,x7 //pu1_src[wd] 8540d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8550d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x4,x24 //Loads pu1_src_top_left 8560d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x11,x11,#1 //ht - 1 8570d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8580d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUB x8,x8,#1 8590d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STRB w9,[x8] //pu1_src_org[wd - 1] = u1_pos_wd_0_tmp 8600d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x8,x8,#1 8610d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar madd x6, x11, x1, x0 //pu1_src_org[(ht - 1) * src_strd] 8620d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8630d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LDRB w8,[sp] //load u1_src_top_left_tmp from stack pointer 8640d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD x12,sp,#0x02 8650d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8660d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STRB w10,[x6] //pu1_src_org[wd - 1 + (ht - 1) * src_strd] = u1_pos_wd_ht_tmp 8670d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar STRB w8,[x4] //*pu1_src_top_left = u1_src_top_left_tmp 8680d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar MOV x3,x22 //Loads pu1_src_top 8690d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8700d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarSRC_TOP_LOOP: 8710d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar LD1 {v0.8b},[x12],#8 //pu1_src_top[col] = au1_src_top_tmp[col] 8720d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar SUBS x7,x7,#8 //Decrement the width 8730d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ST1 {v0.8b},[x3],#8 //pu1_src_top[col] = au1_src_top_tmp[col] 8740d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar BNE SRC_TOP_LOOP 8750d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8760d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish MahendrakarEND_LOOPS: 8770d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ADD sp,sp,#0xA0 8780d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar // LDMFD sp!,{x4-x12,x15} //Reload the registers from SP 8790d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ldp x23, x24,[sp], #16 8800d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ldp x21, x22,[sp], #16 8810d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ldp x19, x20,[sp], #16 8820d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar ret 8830d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8840d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 8850d8951cef4b1a1dbf4ff5ba3e8796cf1d4503098Harish Mahendrakar 886