1 @/***************************************************************************** 2 @* 3 @* Copyright (C) 2012 Ittiam Systems Pvt Ltd, Bangalore 4 @* 5 @* Licensed under the Apache License, Version 2.0 (the "License"); 6 @* you may not use this file except in compliance with the License. 7 @* You may obtain a copy of the License at: 8 @* 9 @* http://www.apache.org/licenses/LICENSE-2.0 10 @* 11 @* Unless required by applicable law or agreed to in writing, software 12 @* distributed under the License is distributed on an "AS IS" BASIS, 13 @* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 14 @* See the License for the specific language governing permissions and 15 @* limitations under the License. 16 @* 17 @*****************************************************************************/ 18 @/** 19 @******************************************************************************* 20 @* ,:file 21 @* ihevc_sao_band_offset_chroma.s 22 @* 23 @* ,:brief 24 @* Contains function definitions for inter prediction interpolation. 25 @* Functions are coded using NEON intrinsics and can be compiled using@ ARM 26 @* RVCT 27 @* 28 @* ,:author 29 @* Parthiban V 30 @* 31 @* ,:par List of Functions: 32 @* 33 @* 34 @* ,:remarks 35 @* None 36 @* 37 @******************************************************************************* 38 @*/ 39 @void ihevc_sao_band_offset_chroma(UWORD8 *pu1_src, 40 @ WORD32 src_strd, 41 @ UWORD8 *pu1_src_left, 42 @ UWORD8 *pu1_src_top, 43 @ UWORD8 *pu1_src_top_left, 44 @ WORD32 sao_band_pos_u, 45 @ WORD32 sao_band_pos_v, 46 @ WORD8 *pi1_sao_offset_u, 47 @ WORD8 *pi1_sao_offset_v, 48 @ WORD32 wd, 49 @ WORD32 ht) 50 @ 51 @**************Variables Vs Registers***************************************** 52 @r0 => *pu1_src 53 @r1 => src_strd 54 @r2 => *pu1_src_left 55 @r3 => *pu1_src_top 56 @r4 => *pu1_src_top_left 57 @r5 => sao_band_pos_u 58 @r6 => sao_band_pos_v 59 @r7 => *pi1_sao_offset_u 60 @r8 => *pi1_sao_offset_v 61 @r9 => wd 62 @r10=> ht 63 64 .text 65 .p2align 2 66 67 .extern gu1_table_band_idx 68 .globl ihevc_sao_band_offset_chroma_a9q 69 70 gu1_table_band_idx_addr_1: 71 .long gu1_table_band_idx - ulbl1 - 8 72 73 gu1_table_band_idx_addr_2: 74 .long gu1_table_band_idx - ulbl2 - 8 75 76 ihevc_sao_band_offset_chroma_a9q: 77 78 STMFD sp!, {r4-r12, r14} @stack stores the values of the arguments 79 LDR r4,[sp,#40] @Loads pu1_src_top_left 80 LDR r10,[sp,#64] @Loads ht 81 82 LDR r9,[sp,#60] @Loads wd 83 MOV r11,r10 @Move the ht to r9 for loop counter 84 85 ADD r12,r0,r9 @pu1_src[row * src_strd + (wd)] 86 LDR r14, gu1_table_band_idx_addr_1 87 ulbl1: 88 add r14,r14,pc 89 SUB r12,r12,#2 @wd-2 90 91 SRC_LEFT_LOOP: 92 LDRH r5,[r12],r1 @Load the value 93 SUBS r11,r11,#1 @Decrement the loop counter 94 STRH r5,[r2],#2 @Store the value in pu1_src_left pointer 95 BNE SRC_LEFT_LOOP 96 97 LDR r5,[sp,#44] @Loads sao_band_pos_u 98 VLD1.8 D1,[r14]! @band_table_u.val[0] 99 ADD r12,r3,r9 @pu1_src_top[wd] 100 101 LDRH r11,[r12,#-2] 102 VLD1.8 D2,[r14]! @band_table_u.val[1] 103 LSL r6,r5,#3 @sao_band_pos_u 104 105 STRH r11,[r4] @store to pu1_src_top_left[0] 106 VLD1.8 D3,[r14]! @band_table_u.val[2] 107 LDR r7,[sp,#52] @Loads pi1_sao_offset_u 108 109 SUB r4,r10,#1 @ht-1 110 VDUP.8 D31,r6 @band_pos_u 111 MUL r4,r4,r1 @ht-1 * src_strd 112 113 ADD r4,r4,r0 @pu1_src[(ht - 1) * src_strd] 114 VLD1.8 D4,[r14]! @band_table_u.val[3] 115 MOV r11,r9 @Move the wd to r9 for loop counter 116 117 SRC_TOP_LOOP: @wd is always multiple of 8 118 VLD1.8 D0,[r4]! @Load pu1_src[(ht - 1) * src_strd + col] 119 SUBS r11,r11,#8 @Decrement the loop counter by 8 120 VST1.8 D0,[r3]! @Store to pu1_src_top[col] 121 BNE SRC_TOP_LOOP 122 123 VLD1.8 D30,[r7] @pi1_sao_offset_u load 124 VADD.I8 D5,D1,D31 @band_table_u.val[0] = vadd_u8(band_table_u.val[0], sao_band_pos_u) 125 126 VDUP.8 D29,D30[1] @vdup_n_u8(pi1_sao_offset_u[1]) 127 VADD.I8 D6,D2,D31 @band_table_u.val[1] = vadd_u8(band_table_u.val[1], sao_band_pos_u) 128 129 VDUP.8 D28,D30[2] @vdup_n_u8(pi1_sao_offset_u[2]) 130 VADD.I8 D7,D3,D31 @band_table_u.val[2] = vadd_u8(band_table_u.val[2], sao_band_pos_u) 131 132 VDUP.8 D27,D30[3] @vdup_n_u8(pi1_sao_offset_u[3]) 133 VADD.I8 D8,D4,D31 @band_table_u.val[3] = vadd_u8(band_table_u.val[3], sao_band_pos_u) 134 135 CMP r5,#28 136 VDUP.8 D26,D30[4] @vdup_n_u8(pi1_sao_offset_u[4]) 137 LDR r14, gu1_table_band_idx_addr_2 138 ulbl2: 139 add r14,r14,pc 140 141 VMOV.I8 D30,#16 @vdup_n_u8(16) 142 VADD.I8 D1,D5,D29 @band_table_u.val[0] = vadd_u8(band_table_u.val[0], vdup_n_u8(pi1_sao_offset_u[1])) 143 144 VLD1.8 D9,[r14]! @band_table_v.val[0] 145 VADD.I8 D2,D6,D28 @band_table_u.val[1] = vadd_u8(band_table_u.val[1], vdup_n_u8(pi1_sao_offset_u[2])) 146 147 VLD1.8 D10,[r14]! @band_table_v.val[1] 148 VADD.I8 D3,D7,D27 @band_table_u.val[2] = vadd_u8(band_table_u.val[2], vdup_n_u8(pi1_sao_offset_u[3])) 149 150 LDR r6,[sp,#48] @Loads sao_band_pos_v 151 VADD.I8 D4,D8,D26 @band_table_u.val[3] = vadd_u8(band_table_u.val[3], vdup_n_u8(pi1_sao_offset_u[4])) 152 LSL r11,r6,#3 @sao_band_pos_v 153 154 BLT SAO_BAND_POS_U_0 155 156 SAO_BAND_POS_U_28: @case 28 157 VCLE.U8 D13,D4,D30 @vcle_u8(band_table.val[3], vdup_n_u8(16)) 158 BNE SAO_BAND_POS_U_29 159 160 VORR.U8 D4,D4,D13 @band_table.val[3] = vorr_u8(band_table.val[3], au1_cmp) 161 B SWITCH_BREAK_U 162 163 SAO_BAND_POS_U_29: @case 29 164 CMP r5,#29 165 166 VCLE.U8 D14,D3,D30 @vcle_u8(band_table.val[2], vdup_n_u8(16)) 167 BNE SAO_BAND_POS_U_30 168 VORR.U8 D3,D3,D14 @band_table.val[2] = vorr_u8(band_table.val[2], au1_cmp) 169 170 VAND.U8 D4,D4,D13 @band_table.val[3] = vand_u8(band_table.val[3], au1_cmp) 171 B SWITCH_BREAK_U 172 173 SAO_BAND_POS_U_30: @case 30 174 CMP r5,#30 175 176 VCLE.U8 D15,D2,D30 @vcle_u8(band_table.val[1], vdup_n_u8(16)) 177 BNE SAO_BAND_POS_U_31 178 VORR.U8 D2,D2,D15 @band_table.val[1] = vorr_u8(band_table.val[1], au1_cmp) 179 180 VAND.U8 D3,D3,D14 @band_table.val[2] = vand_u8(band_table.val[2], au1_cmp) 181 182 SAO_BAND_POS_U_31: @case 31 183 CMP r5,#31 184 BNE SWITCH_BREAK_U 185 186 VCLE.U8 D16,D1,D30 @vcle_u8(band_table.val[0], vdup_n_u8(16)) 187 VORR.U8 D1,D1,D16 @band_table.val[0] = vorr_u8(band_table.val[0], au1_cmp) 188 189 VAND.U8 D2,D2,D15 @band_table.val[1] = vand_u8(band_table.val[1], au1_cmp) 190 B SWITCH_BREAK_U 191 192 SAO_BAND_POS_U_0: 193 CMP r5,#0 @case 0 194 BNE SWITCH_BREAK_U 195 196 VCLE.U8 D16,D1,D30 @vcle_u8(band_table.val[0], vdup_n_u8(16)) 197 VAND.U8 D1,D1,D16 @band_table.val[0] = vand_u8(band_table.val[0], au1_cmp) 198 199 SWITCH_BREAK_U: 200 VDUP.8 D30,r11 @band_pos_v 201 LDR r8,[sp,#56] @Loads pi1_sao_offset_v 202 203 VLD1.8 D11,[r14]! @band_table_v.val[2] 204 VADD.I8 D13,D9,D30 @band_table_v.val[0] = vadd_u8(band_table_v.val[0], band_pos_v) 205 206 VLD1.8 D12,[r14]! @band_table_v.val[3] 207 VADD.I8 D14,D10,D30 @band_table_v.val[1] = vadd_u8(band_table_v.val[1], band_pos_v) 208 209 VLD1.8 D25,[r8] @pi1_sao_offset_v load 210 VADD.I8 D15,D11,D30 @band_table_v.val[2] = vadd_u8(band_table_v.val[2], band_pos_v) 211 212 VDUP.8 D29,D25[1] @vdup_n_u8(pi1_sao_offset_v[1]) 213 VADD.I8 D16,D12,D30 @band_table_v.val[3] = vadd_u8(band_table_v.val[3], band_pos_v) 214 215 VDUP.8 D28,D25[2] @vdup_n_u8(pi1_sao_offset_v[2]) 216 VADD.I8 D9,D13,D29 @band_table_v.val[0] = vadd_u8(band_table_v.val[0], vdup_n_u8(pi1_sao_offset_v[1])) 217 218 VDUP.8 D27,D25[3] @vdup_n_u8(pi1_sao_offset_v[3]) 219 VADD.I8 D10,D14,D28 @band_table_v.val[1] = vadd_u8(band_table_v.val[1], vdup_n_u8(pi1_sao_offset_v[2])) 220 221 VDUP.8 D26,D25[4] @vdup_n_u8(pi1_sao_offset_v[4]) 222 VADD.I8 D11,D15,D27 @band_table_v.val[2] = vadd_u8(band_table_v.val[2], vdup_n_u8(pi1_sao_offset_v[3])) 223 224 VMOV.I8 D29,#16 @vdup_n_u8(16) 225 VADD.I8 D12,D16,D26 @band_table_v.val[3] = vadd_u8(band_table_v.val[3], vdup_n_u8(pi1_sao_offset_v[4])) 226 AND r12,r9,#0xf 227 228 CMP r6,#28 229 BLT SAO_BAND_POS_V_0 230 231 SAO_BAND_POS_V_28: @case 28 232 VCLE.U8 D17,D12,D29 @vcle_u8(band_table.val[3], vdup_n_u8(16)) 233 BNE SAO_BAND_POS_V_29 234 VORR.U8 D12,D12,D17 @band_table.val[3] = vorr_u8(band_table.val[3], au1_cmp) 235 B SWITCH_BREAK_V 236 237 SAO_BAND_POS_V_29: @case 29 238 CMP r6,#29 239 240 VCLE.U8 D18,D11,D29 @vcle_u8(band_table.val[2], vdup_n_u8(16)) 241 BNE SAO_BAND_POS_V_30 242 VORR.U8 D11,D11,D18 @band_table.val[2] = vorr_u8(band_table.val[2], au1_cmp) 243 244 VAND.U8 D12,D12,D17 @band_table.val[3] = vand_u8(band_table.val[3], au1_cmp) 245 B SWITCH_BREAK_V 246 247 SAO_BAND_POS_V_30: @case 30 248 CMP r6,#30 249 250 VCLE.U8 D19,D10,D29 @vcle_u8(band_table.val[1], vdup_n_u8(16)) 251 BNE SAO_BAND_POS_V_31 252 VORR.U8 D10,D10,D19 @band_table.val[1] = vorr_u8(band_table.val[1], au1_cmp) 253 254 VAND.U8 D11,D11,D18 @band_table.val[2] = vand_u8(band_table.val[2], au1_cmp) 255 B SWITCH_BREAK_V 256 257 SAO_BAND_POS_V_31: @case 31 258 CMP r6,#31 259 BNE SWITCH_BREAK_V 260 261 VCLE.U8 D20,D9,D29 @vcle_u8(band_table.val[0], vdup_n_u8(16)) 262 VORR.U8 D9,D9,D20 @band_table.val[0] = vorr_u8(band_table.val[0], au1_cmp) 263 264 VAND.U8 D10,D10,D19 @band_table.val[1] = vand_u8(band_table.val[1], au1_cmp) 265 B SWITCH_BREAK_V 266 267 SAO_BAND_POS_V_0: 268 CMP r6,#0 @case 0 269 BNE SWITCH_BREAK_V 270 271 VCLE.U8 D20,D9,D29 @vcle_u8(band_table.val[0], vdup_n_u8(16)) 272 VAND.U8 D9,D9,D20 @band_table.val[0] = vand_u8(band_table.val[0], au1_cmp) 273 274 SWITCH_BREAK_V: 275 CMP r9,#16 276 MOV r4,r0 @pu1_src_cpy 277 BLT WIDTH_RESIDUE 278 279 WIDTH_LOOP: @Width is assigned to be multiple of 16 280 MOV r4,r0 @pu1_src_cpy 281 MOV r11,r10 @move ht 282 ADD r5,r4,r1 283 284 HEIGHT_LOOP: @unrolled for 4 rows 285 ADD r6,r5,r1 286 VLD2.8 {D5,D6},[r4] @vld1q_u8(pu1_src_cpy) 287 ADD r7,r6,r1 288 289 VLD2.8 {D13,D14},[r5] @vld1q_u8(pu1_src_cpy) 290 VSUB.I8 D7,D5,D31 @vsub_u8(au1_cur_row_deint.val[0], band_pos_u) 291 292 VLD2.8 {D17,D18},[r6] @vld1q_u8(pu1_src_cpy) 293 VSUB.I8 D8,D6,D30 @vsub_u8(au1_cur_row_deint.val[1], band_pos_v) 294 295 VLD2.8 {D21,D22},[r7] @vld1q_u8(pu1_src_cpy) 296 VSUB.I8 D15,D13,D31 @vsub_u8(au1_cur_row_deint.val[0], band_pos_u) 297 298 VTBX.8 D5,{D1-D4},D7 @vtbx4_u8(au1_cur_row_deint.val[0], band_table_u, vsub_u8(au1_cur_row_deint.val[0], band_pos_u)) 299 VSUB.I8 D16,D14,D30 @vsub_u8(au1_cur_row_deint.val[1], band_pos_v) 300 301 VTBX.8 D6,{D9-D12},D8 @vtbx4_u8(au1_cur_row_deint.val[1], band_table_v, vsub_u8(au1_cur_row_deint.val[1], band_pos_v)) 302 VSUB.I8 D19,D17,D31 @vsub_u8(au1_cur_row_deint.val[0], band_pos_u) 303 304 VTBX.8 D13,{D1-D4},D15 @vtbx4_u8(au1_cur_row_deint.val[0], band_table_u, vsub_u8(au1_cur_row_deint.val[0], band_pos_u)) 305 VSUB.I8 D20,D18,D30 @vsub_u8(au1_cur_row_deint.val[1], band_pos_v) 306 307 VTBX.8 D14,{D9-D12},D16 @vtbx4_u8(au1_cur_row_deint.val[1], band_table_v, vsub_u8(au1_cur_row_deint.val[1], band_pos_v)) 308 VSUB.I8 D23,D21,D31 @vsub_u8(au1_cur_row_deint.val[0], band_pos_u) 309 310 VST2.8 {D5,D6},[r4] @vst1q_u8(pu1_src_cpy, au1_cur_row) 311 VSUB.I8 D24,D22,D30 @vsub_u8(au1_cur_row_deint.val[1], band_pos_v) 312 313 SUBS r11,r11,#4 @Decrement the ht loop count by 4 314 VTBX.8 D17,{D1-D4},D19 @vtbx4_u8(au1_cur_row_deint.val[0], band_table_u, vsub_u8(au1_cur_row_deint.val[0], band_pos_u)) 315 316 VST2.8 {D13,D14},[r5] @vst1q_u8(pu1_src_cpy, au1_cur_row) 317 318 VTBX.8 D18,{D9-D12},D20 @vtbx4_u8(au1_cur_row_deint.val[1], band_table_v, vsub_u8(au1_cur_row_deint.val[1], band_pos_v)) 319 VTBX.8 D21,{D1-D4},D23 @vtbx4_u8(au1_cur_row_deint.val[0], band_table_u, vsub_u8(au1_cur_row_deint.val[0], band_pos_u)) 320 VTBX.8 D22,{D9-D12},D24 @vtbx4_u8(au1_cur_row_deint.val[1], band_table_v, vsub_u8(au1_cur_row_deint.val[1], band_pos_v)) 321 322 VST2.8 {D17,D18},[r6],r1 @vst1q_u8(pu1_src_cpy, au1_cur_row) 323 324 ADD r4,r6,r1 325 VST2.8 {D21,D22},[r7] @vst1q_u8(pu1_src_cpy, au1_cur_row) 326 ADD r5,r4,r1 327 328 BNE HEIGHT_LOOP 329 330 SUB r9,r9,#16 @Decrement the width loop by 16 331 ADD r0,r0,#16 332 CMP r9,#8 333 BGT WIDTH_LOOP 334 BLT END_LOOP 335 MOV r4,r0 @pu1_src_cpy 336 337 WIDTH_RESIDUE: @If width is not multiple of 16 338 ADD r5,r4,r1 339 VLD2.8 {D5,D6},[r4] @vld1q_u8(pu1_src_cpy) 340 ADD r6,r5,r1 341 342 ADD r7,r6,r1 343 VLD2.8 {D13,D14},[r5] @vld1q_u8(pu1_src_cpy) 344 VSUB.I8 D7,D5,D31 @vsub_u8(au1_cur_row_deint.val[0], band_pos_u) 345 346 VLD2.8 {D17,D18},[r6] @vld1q_u8(pu1_src_cpy) 347 VSUB.I8 D8,D6,D30 @vsub_u8(au1_cur_row_deint.val[1], band_pos_v) 348 349 VTBX.8 D5,{D1-D4},D7 @vtbx4_u8(au1_cur_row_deint.val[0], band_table_u, vsub_u8(au1_cur_row_deint.val[0], band_pos_u)) 350 VSUB.I8 D15,D13,D31 @vsub_u8(au1_cur_row_deint.val[0], band_pos_u) 351 352 VTBX.8 D6,{D9-D12},D8 @vtbx4_u8(au1_cur_row_deint.val[1], band_table_v, vsub_u8(au1_cur_row_deint.val[1], band_pos_v)) 353 VSUB.I8 D16,D14,D30 @vsub_u8(au1_cur_row_deint.val[1], band_pos_v) 354 355 VLD2.8 {D21,D22},[r7] @vld1q_u8(pu1_src_cpy) 356 VSUB.I8 D19,D17,D31 @vsub_u8(au1_cur_row_deint.val[0], band_pos_u) 357 358 VTBX.8 D13,{D1-D4},D15 @vtbx4_u8(au1_cur_row_deint.val[0], band_table_u, vsub_u8(au1_cur_row_deint.val[0], band_pos_u)) 359 VSUB.I8 D20,D18,D30 @vsub_u8(au1_cur_row_deint.val[1], band_pos_v) 360 361 VTBX.8 D14,{D9-D12},D16 @vtbx4_u8(au1_cur_row_deint.val[1], band_table_v, vsub_u8(au1_cur_row_deint.val[1], band_pos_v)) 362 VZIP.8 D5,D6 363 364 VTBX.8 D17,{D1-D4},D19 @vtbx4_u8(au1_cur_row_deint.val[0], band_table_u, vsub_u8(au1_cur_row_deint.val[0], band_pos_u)) 365 VSUB.I8 D23,D21,D31 @vsub_u8(au1_cur_row_deint.val[0], band_pos_u) 366 367 VST1.8 {D5},[r4] @vst1q_u8(pu1_src_cpy, au1_cur_row) 368 VZIP.8 D13,D14 369 370 VTBX.8 D18,{D9-D12},D20 @vtbx4_u8(au1_cur_row_deint.val[1], band_table_v, vsub_u8(au1_cur_row_deint.val[1], band_pos_v)) 371 VSUB.I8 D24,D22,D30 @vsub_u8(au1_cur_row_deint.val[1], band_pos_v) 372 373 VST1.8 {D13},[r5] @vst1q_u8(pu1_src_cpy, au1_cur_row) 374 SUBS r10,r10,#4 @Decrement the ht loop count by 4 375 376 VTBX.8 D21,{D1-D4},D23 @vtbx4_u8(au1_cur_row_deint.val[0], band_table_u, vsub_u8(au1_cur_row_deint.val[0], band_pos_u)) 377 VZIP.8 D17,D18 378 379 VTBX.8 D22,{D9-D12},D24 @vtbx4_u8(au1_cur_row_deint.val[1], band_table_v, vsub_u8(au1_cur_row_deint.val[1], band_pos_v)) 380 VST1.8 {D17},[r6],r1 @vst1q_u8(pu1_src_cpy, au1_cur_row) 381 VZIP.8 D21,D22 382 383 ADD r4,r6,r1 384 VST1.8 {D21},[r7] @vst1q_u8(pu1_src_cpy, au1_cur_row) 385 ADD r5,r4,r1 386 387 BNE WIDTH_RESIDUE 388 389 END_LOOP: 390 LDMFD sp!,{r4-r12,r15} @Reload the registers from SP 391 392 393 394
