fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright 2011 The LibYuv Project Authors. All rights reserved. |
| 3 | * |
| 4 | * Use of this source code is governed by a BSD-style license |
| 5 | * that can be found in the LICENSE file in the root of the source |
| 6 | * tree. An additional intellectual property rights grant can be found |
| 7 | * in the file PATENTS. All contributing project authors may |
| 8 | * be found in the AUTHORS file in the root of the source tree. |
| 9 | */ |
| 10 | |
| 11 | #include "libyuv/basic_types.h" |
| 12 | #include "libyuv/row.h" |
| 13 | |
| 14 | #ifdef __cplusplus |
| 15 | namespace libyuv { |
| 16 | extern "C" { |
| 17 | #endif |
| 18 | |
| 19 | // This module is for GCC Neon |
| 20 | #if !defined(YUV_DISABLE_ASM) && defined(__ARM_NEON__) |
| 21 | |
| 22 | /** |
| 23 | * NEON downscalers with interpolation. |
| 24 | * |
| 25 | * Provided by Fritz Koenig |
| 26 | * |
| 27 | */ |
| 28 | |
| 29 | void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */, |
| 30 | uint8* dst, int dst_width) { |
| 31 | asm volatile ( |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 32 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 33 | // load even pixels into q0, odd into q1 |
| 34 | "vld2.u8 {q0,q1}, [%0]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 35 | "subs %2, %2, #16 \n" // 16 processed per loop |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 36 | "vst1.u8 {q0}, [%1]! \n" // store even pixels |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 37 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 38 | : "+r"(src_ptr), // %0 |
| 39 | "+r"(dst), // %1 |
| 40 | "+r"(dst_width) // %2 |
| 41 | : |
| 42 | : "q0", "q1" // Clobber List |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 43 | ); |
| 44 | } |
| 45 | |
| 46 | void ScaleRowDown2Int_NEON(const uint8* src_ptr, ptrdiff_t src_stride, |
| 47 | uint8* dst, int dst_width) { |
| 48 | asm volatile ( |
| 49 | // change the stride to row 2 pointer |
| 50 | "add %1, %0 \n" |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 51 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 52 | "vld1.u8 {q0,q1}, [%0]! \n" // load row 1 and post inc |
| 53 | "vld1.u8 {q2,q3}, [%1]! \n" // load row 2 and post inc |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 54 | "subs %3, %3, #16 \n" // 16 processed per loop |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 55 | "vpaddl.u8 q0, q0 \n" // row 1 add adjacent |
| 56 | "vpaddl.u8 q1, q1 \n" |
| 57 | "vpadal.u8 q0, q2 \n" // row 2 add adjacent + row1 |
| 58 | "vpadal.u8 q1, q3 \n" |
| 59 | "vrshrn.u16 d0, q0, #2 \n" // downshift, round and pack |
| 60 | "vrshrn.u16 d1, q1, #2 \n" |
| 61 | "vst1.u8 {q0}, [%2]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 62 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 63 | : "+r"(src_ptr), // %0 |
| 64 | "+r"(src_stride), // %1 |
| 65 | "+r"(dst), // %2 |
| 66 | "+r"(dst_width) // %3 |
| 67 | : |
| 68 | : "q0", "q1", "q2", "q3" // Clobber List |
| 69 | ); |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 70 | } |
| 71 | |
| 72 | void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */, |
| 73 | uint8* dst_ptr, int dst_width) { |
| 74 | asm volatile ( |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 75 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 76 | "vld2.u8 {d0, d1}, [%0]! \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 77 | "subs %2, #4 \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 78 | "vtrn.u8 d1, d0 \n" |
| 79 | "vshrn.u16 d0, q0, #8 \n" |
| 80 | "vst1.u32 {d0[1]}, [%1]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 81 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 82 | : "+r"(src_ptr), // %0 |
| 83 | "+r"(dst_ptr), // %1 |
| 84 | "+r"(dst_width) // %2 |
| 85 | : |
| 86 | : "q0", "q1", "memory", "cc" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 87 | ); |
| 88 | } |
| 89 | |
| 90 | void ScaleRowDown4Int_NEON(const uint8* src_ptr, ptrdiff_t src_stride, |
| 91 | uint8* dst_ptr, int dst_width) { |
| 92 | asm volatile ( |
| 93 | "add r4, %0, %3 \n" |
| 94 | "add r5, r4, %3 \n" |
| 95 | "add %3, r5, %3 \n" |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 96 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 97 | "vld1.u8 {q0}, [%0]! \n" // load up 16x4 |
| 98 | "vld1.u8 {q1}, [r4]! \n" |
| 99 | "vld1.u8 {q2}, [r5]! \n" |
| 100 | "vld1.u8 {q3}, [%3]! \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 101 | "subs %2, #4 \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 102 | "vpaddl.u8 q0, q0 \n" |
| 103 | "vpadal.u8 q0, q1 \n" |
| 104 | "vpadal.u8 q0, q2 \n" |
| 105 | "vpadal.u8 q0, q3 \n" |
| 106 | "vpaddl.u16 q0, q0 \n" |
| 107 | "vrshrn.u32 d0, q0, #4 \n" // divide by 16 w/rounding |
| 108 | "vmovn.u16 d0, q0 \n" |
| 109 | "vst1.u32 {d0[0]}, [%1]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 110 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 111 | : "+r"(src_ptr), // %0 |
| 112 | "+r"(dst_ptr), // %1 |
| 113 | "+r"(dst_width) // %2 |
| 114 | : "r"(src_stride) // %3 |
| 115 | : "r4", "r5", "q0", "q1", "q2", "q3", "memory", "cc" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 116 | ); |
| 117 | } |
| 118 | |
| 119 | // Down scale from 4 to 3 pixels. Use the neon multilane read/write |
| 120 | // to load up the every 4th pixel into a 4 different registers. |
| 121 | // Point samples 32 pixels to 24 pixels. |
| 122 | void ScaleRowDown34_NEON(const uint8* src_ptr, |
| 123 | ptrdiff_t /* src_stride */, |
| 124 | uint8* dst_ptr, int dst_width) { |
| 125 | asm volatile ( |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 126 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 127 | "vld4.u8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 128 | "subs %2, #24 \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 129 | "vmov d2, d3 \n" // order d0, d1, d2 |
| 130 | "vst3.u8 {d0, d1, d2}, [%1]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 131 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 132 | : "+r"(src_ptr), // %0 |
| 133 | "+r"(dst_ptr), // %1 |
| 134 | "+r"(dst_width) // %2 |
| 135 | : |
| 136 | : "d0", "d1", "d2", "d3", "memory", "cc" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 137 | ); |
| 138 | } |
| 139 | |
| 140 | void ScaleRowDown34_0_Int_NEON(const uint8* src_ptr, |
| 141 | ptrdiff_t src_stride, |
| 142 | uint8* dst_ptr, int dst_width) { |
| 143 | asm volatile ( |
| 144 | "vmov.u8 d24, #3 \n" |
| 145 | "add %3, %0 \n" |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 146 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 147 | "vld4.u8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 |
| 148 | "vld4.u8 {d4, d5, d6, d7}, [%3]! \n" // src line 1 |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 149 | "subs %2, #24 \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 150 | |
| 151 | // filter src line 0 with src line 1 |
| 152 | // expand chars to shorts to allow for room |
| 153 | // when adding lines together |
| 154 | "vmovl.u8 q8, d4 \n" |
| 155 | "vmovl.u8 q9, d5 \n" |
| 156 | "vmovl.u8 q10, d6 \n" |
| 157 | "vmovl.u8 q11, d7 \n" |
| 158 | |
| 159 | // 3 * line_0 + line_1 |
| 160 | "vmlal.u8 q8, d0, d24 \n" |
| 161 | "vmlal.u8 q9, d1, d24 \n" |
| 162 | "vmlal.u8 q10, d2, d24 \n" |
| 163 | "vmlal.u8 q11, d3, d24 \n" |
| 164 | |
| 165 | // (3 * line_0 + line_1) >> 2 |
| 166 | "vqrshrn.u16 d0, q8, #2 \n" |
| 167 | "vqrshrn.u16 d1, q9, #2 \n" |
| 168 | "vqrshrn.u16 d2, q10, #2 \n" |
| 169 | "vqrshrn.u16 d3, q11, #2 \n" |
| 170 | |
| 171 | // a0 = (src[0] * 3 + s[1] * 1) >> 2 |
| 172 | "vmovl.u8 q8, d1 \n" |
| 173 | "vmlal.u8 q8, d0, d24 \n" |
| 174 | "vqrshrn.u16 d0, q8, #2 \n" |
| 175 | |
| 176 | // a1 = (src[1] * 1 + s[2] * 1) >> 1 |
| 177 | "vrhadd.u8 d1, d1, d2 \n" |
| 178 | |
| 179 | // a2 = (src[2] * 1 + s[3] * 3) >> 2 |
| 180 | "vmovl.u8 q8, d2 \n" |
| 181 | "vmlal.u8 q8, d3, d24 \n" |
| 182 | "vqrshrn.u16 d2, q8, #2 \n" |
| 183 | |
| 184 | "vst3.u8 {d0, d1, d2}, [%1]! \n" |
| 185 | |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 186 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 187 | : "+r"(src_ptr), // %0 |
| 188 | "+r"(dst_ptr), // %1 |
| 189 | "+r"(dst_width), // %2 |
| 190 | "+r"(src_stride) // %3 |
| 191 | : |
| 192 | : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 193 | ); |
| 194 | } |
| 195 | |
| 196 | void ScaleRowDown34_1_Int_NEON(const uint8* src_ptr, |
| 197 | ptrdiff_t src_stride, |
| 198 | uint8* dst_ptr, int dst_width) { |
| 199 | asm volatile ( |
| 200 | "vmov.u8 d24, #3 \n" |
| 201 | "add %3, %0 \n" |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 202 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 203 | "vld4.u8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 |
| 204 | "vld4.u8 {d4, d5, d6, d7}, [%3]! \n" // src line 1 |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 205 | "subs %2, #24 \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 206 | // average src line 0 with src line 1 |
| 207 | "vrhadd.u8 q0, q0, q2 \n" |
| 208 | "vrhadd.u8 q1, q1, q3 \n" |
| 209 | |
| 210 | // a0 = (src[0] * 3 + s[1] * 1) >> 2 |
| 211 | "vmovl.u8 q3, d1 \n" |
| 212 | "vmlal.u8 q3, d0, d24 \n" |
| 213 | "vqrshrn.u16 d0, q3, #2 \n" |
| 214 | |
| 215 | // a1 = (src[1] * 1 + s[2] * 1) >> 1 |
| 216 | "vrhadd.u8 d1, d1, d2 \n" |
| 217 | |
| 218 | // a2 = (src[2] * 1 + s[3] * 3) >> 2 |
| 219 | "vmovl.u8 q3, d2 \n" |
| 220 | "vmlal.u8 q3, d3, d24 \n" |
| 221 | "vqrshrn.u16 d2, q3, #2 \n" |
| 222 | |
| 223 | "vst3.u8 {d0, d1, d2}, [%1]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 224 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 225 | : "+r"(src_ptr), // %0 |
| 226 | "+r"(dst_ptr), // %1 |
| 227 | "+r"(dst_width), // %2 |
| 228 | "+r"(src_stride) // %3 |
| 229 | : |
| 230 | : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 231 | ); |
| 232 | } |
| 233 | |
| 234 | #define HAS_SCALEROWDOWN38_NEON |
| 235 | const uvec8 kShuf38 = |
| 236 | { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 }; |
| 237 | const uvec8 kShuf38_2 = |
| 238 | { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 }; |
| 239 | const vec16 kMult38_Div6 = |
| 240 | { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12, |
| 241 | 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 }; |
| 242 | const vec16 kMult38_Div9 = |
| 243 | { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18, |
| 244 | 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 }; |
| 245 | |
| 246 | // 32 -> 12 |
| 247 | void ScaleRowDown38_NEON(const uint8* src_ptr, |
| 248 | ptrdiff_t /* src_stride */, |
| 249 | uint8* dst_ptr, int dst_width) { |
| 250 | asm volatile ( |
| 251 | "vld1.u8 {q3}, [%3] \n" |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 252 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 253 | "vld1.u8 {d0, d1, d2, d3}, [%0]! \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 254 | "subs %2, #12 \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 255 | "vtbl.u8 d4, {d0, d1, d2, d3}, d6 \n" |
| 256 | "vtbl.u8 d5, {d0, d1, d2, d3}, d7 \n" |
| 257 | "vst1.u8 {d4}, [%1]! \n" |
| 258 | "vst1.u32 {d5[0]}, [%1]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 259 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 260 | : "+r"(src_ptr), // %0 |
| 261 | "+r"(dst_ptr), // %1 |
| 262 | "+r"(dst_width) // %2 |
| 263 | : "r"(&kShuf38) // %3 |
| 264 | : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 265 | ); |
| 266 | } |
| 267 | |
| 268 | // 32x3 -> 12x1 |
| 269 | void OMITFP ScaleRowDown38_3_Int_NEON(const uint8* src_ptr, |
| 270 | ptrdiff_t src_stride, |
| 271 | uint8* dst_ptr, int dst_width) { |
| 272 | asm volatile ( |
| 273 | "vld1.u16 {q13}, [%4] \n" |
| 274 | "vld1.u8 {q14}, [%5] \n" |
| 275 | "vld1.u8 {q15}, [%6] \n" |
| 276 | "add r4, %0, %3, lsl #1 \n" |
| 277 | "add %3, %0 \n" |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 278 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 279 | |
| 280 | // d0 = 00 40 01 41 02 42 03 43 |
| 281 | // d1 = 10 50 11 51 12 52 13 53 |
| 282 | // d2 = 20 60 21 61 22 62 23 63 |
| 283 | // d3 = 30 70 31 71 32 72 33 73 |
| 284 | "vld4.u8 {d0, d1, d2, d3}, [%0]! \n" |
| 285 | "vld4.u8 {d4, d5, d6, d7}, [%3]! \n" |
| 286 | "vld4.u8 {d16, d17, d18, d19}, [r4]! \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 287 | "subs %2, #12 \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 288 | |
| 289 | // Shuffle the input data around to get align the data |
| 290 | // so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7 |
| 291 | // d0 = 00 10 01 11 02 12 03 13 |
| 292 | // d1 = 40 50 41 51 42 52 43 53 |
| 293 | "vtrn.u8 d0, d1 \n" |
| 294 | "vtrn.u8 d4, d5 \n" |
| 295 | "vtrn.u8 d16, d17 \n" |
| 296 | |
| 297 | // d2 = 20 30 21 31 22 32 23 33 |
| 298 | // d3 = 60 70 61 71 62 72 63 73 |
| 299 | "vtrn.u8 d2, d3 \n" |
| 300 | "vtrn.u8 d6, d7 \n" |
| 301 | "vtrn.u8 d18, d19 \n" |
| 302 | |
| 303 | // d0 = 00+10 01+11 02+12 03+13 |
| 304 | // d2 = 40+50 41+51 42+52 43+53 |
| 305 | "vpaddl.u8 q0, q0 \n" |
| 306 | "vpaddl.u8 q2, q2 \n" |
| 307 | "vpaddl.u8 q8, q8 \n" |
| 308 | |
| 309 | // d3 = 60+70 61+71 62+72 63+73 |
| 310 | "vpaddl.u8 d3, d3 \n" |
| 311 | "vpaddl.u8 d7, d7 \n" |
| 312 | "vpaddl.u8 d19, d19 \n" |
| 313 | |
| 314 | // combine source lines |
| 315 | "vadd.u16 q0, q2 \n" |
| 316 | "vadd.u16 q0, q8 \n" |
| 317 | "vadd.u16 d4, d3, d7 \n" |
| 318 | "vadd.u16 d4, d19 \n" |
| 319 | |
| 320 | // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0] |
| 321 | // + s[6 + st * 1] + s[7 + st * 1] |
| 322 | // + s[6 + st * 2] + s[7 + st * 2]) / 6 |
| 323 | "vqrdmulh.s16 q2, q2, q13 \n" |
| 324 | "vmovn.u16 d4, q2 \n" |
| 325 | |
| 326 | // Shuffle 2,3 reg around so that 2 can be added to the |
| 327 | // 0,1 reg and 3 can be added to the 4,5 reg. This |
| 328 | // requires expanding from u8 to u16 as the 0,1 and 4,5 |
| 329 | // registers are already expanded. Then do transposes |
| 330 | // to get aligned. |
| 331 | // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33 |
| 332 | "vmovl.u8 q1, d2 \n" |
| 333 | "vmovl.u8 q3, d6 \n" |
| 334 | "vmovl.u8 q9, d18 \n" |
| 335 | |
| 336 | // combine source lines |
| 337 | "vadd.u16 q1, q3 \n" |
| 338 | "vadd.u16 q1, q9 \n" |
| 339 | |
| 340 | // d4 = xx 20 xx 30 xx 22 xx 32 |
| 341 | // d5 = xx 21 xx 31 xx 23 xx 33 |
| 342 | "vtrn.u32 d2, d3 \n" |
| 343 | |
| 344 | // d4 = xx 20 xx 21 xx 22 xx 23 |
| 345 | // d5 = xx 30 xx 31 xx 32 xx 33 |
| 346 | "vtrn.u16 d2, d3 \n" |
| 347 | |
| 348 | // 0+1+2, 3+4+5 |
| 349 | "vadd.u16 q0, q1 \n" |
| 350 | |
| 351 | // Need to divide, but can't downshift as the the value |
| 352 | // isn't a power of 2. So multiply by 65536 / n |
| 353 | // and take the upper 16 bits. |
| 354 | "vqrdmulh.s16 q0, q0, q15 \n" |
| 355 | |
| 356 | // Align for table lookup, vtbl requires registers to |
| 357 | // be adjacent |
| 358 | "vmov.u8 d2, d4 \n" |
| 359 | |
| 360 | "vtbl.u8 d3, {d0, d1, d2}, d28 \n" |
| 361 | "vtbl.u8 d4, {d0, d1, d2}, d29 \n" |
| 362 | |
| 363 | "vst1.u8 {d3}, [%1]! \n" |
| 364 | "vst1.u32 {d4[0]}, [%1]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 365 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 366 | : "+r"(src_ptr), // %0 |
| 367 | "+r"(dst_ptr), // %1 |
| 368 | "+r"(dst_width), // %2 |
| 369 | "+r"(src_stride) // %3 |
| 370 | : "r"(&kMult38_Div6), // %4 |
| 371 | "r"(&kShuf38_2), // %5 |
| 372 | "r"(&kMult38_Div9) // %6 |
| 373 | : "r4", "q0", "q1", "q2", "q3", "q8", "q9", |
| 374 | "q13", "q14", "q15", "memory", "cc" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 375 | ); |
| 376 | } |
| 377 | |
| 378 | // 32x2 -> 12x1 |
| 379 | void ScaleRowDown38_2_Int_NEON(const uint8* src_ptr, |
| 380 | ptrdiff_t src_stride, |
| 381 | uint8* dst_ptr, int dst_width) { |
| 382 | asm volatile ( |
| 383 | "vld1.u16 {q13}, [%4] \n" |
| 384 | "vld1.u8 {q14}, [%5] \n" |
| 385 | "add %3, %0 \n" |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 386 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 387 | |
| 388 | // d0 = 00 40 01 41 02 42 03 43 |
| 389 | // d1 = 10 50 11 51 12 52 13 53 |
| 390 | // d2 = 20 60 21 61 22 62 23 63 |
| 391 | // d3 = 30 70 31 71 32 72 33 73 |
| 392 | "vld4.u8 {d0, d1, d2, d3}, [%0]! \n" |
| 393 | "vld4.u8 {d4, d5, d6, d7}, [%3]! \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 394 | "subs %2, #12 \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 395 | |
| 396 | // Shuffle the input data around to get align the data |
| 397 | // so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7 |
| 398 | // d0 = 00 10 01 11 02 12 03 13 |
| 399 | // d1 = 40 50 41 51 42 52 43 53 |
| 400 | "vtrn.u8 d0, d1 \n" |
| 401 | "vtrn.u8 d4, d5 \n" |
| 402 | |
| 403 | // d2 = 20 30 21 31 22 32 23 33 |
| 404 | // d3 = 60 70 61 71 62 72 63 73 |
| 405 | "vtrn.u8 d2, d3 \n" |
| 406 | "vtrn.u8 d6, d7 \n" |
| 407 | |
| 408 | // d0 = 00+10 01+11 02+12 03+13 |
| 409 | // d2 = 40+50 41+51 42+52 43+53 |
| 410 | "vpaddl.u8 q0, q0 \n" |
| 411 | "vpaddl.u8 q2, q2 \n" |
| 412 | |
| 413 | // d3 = 60+70 61+71 62+72 63+73 |
| 414 | "vpaddl.u8 d3, d3 \n" |
| 415 | "vpaddl.u8 d7, d7 \n" |
| 416 | |
| 417 | // combine source lines |
| 418 | "vadd.u16 q0, q2 \n" |
| 419 | "vadd.u16 d4, d3, d7 \n" |
| 420 | |
| 421 | // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4 |
| 422 | "vqrshrn.u16 d4, q2, #2 \n" |
| 423 | |
| 424 | // Shuffle 2,3 reg around so that 2 can be added to the |
| 425 | // 0,1 reg and 3 can be added to the 4,5 reg. This |
| 426 | // requires expanding from u8 to u16 as the 0,1 and 4,5 |
| 427 | // registers are already expanded. Then do transposes |
| 428 | // to get aligned. |
| 429 | // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33 |
| 430 | "vmovl.u8 q1, d2 \n" |
| 431 | "vmovl.u8 q3, d6 \n" |
| 432 | |
| 433 | // combine source lines |
| 434 | "vadd.u16 q1, q3 \n" |
| 435 | |
| 436 | // d4 = xx 20 xx 30 xx 22 xx 32 |
| 437 | // d5 = xx 21 xx 31 xx 23 xx 33 |
| 438 | "vtrn.u32 d2, d3 \n" |
| 439 | |
| 440 | // d4 = xx 20 xx 21 xx 22 xx 23 |
| 441 | // d5 = xx 30 xx 31 xx 32 xx 33 |
| 442 | "vtrn.u16 d2, d3 \n" |
| 443 | |
| 444 | // 0+1+2, 3+4+5 |
| 445 | "vadd.u16 q0, q1 \n" |
| 446 | |
| 447 | // Need to divide, but can't downshift as the the value |
| 448 | // isn't a power of 2. So multiply by 65536 / n |
| 449 | // and take the upper 16 bits. |
| 450 | "vqrdmulh.s16 q0, q0, q13 \n" |
| 451 | |
| 452 | // Align for table lookup, vtbl requires registers to |
| 453 | // be adjacent |
| 454 | "vmov.u8 d2, d4 \n" |
| 455 | |
| 456 | "vtbl.u8 d3, {d0, d1, d2}, d28 \n" |
| 457 | "vtbl.u8 d4, {d0, d1, d2}, d29 \n" |
| 458 | |
| 459 | "vst1.u8 {d3}, [%1]! \n" |
| 460 | "vst1.u32 {d4[0]}, [%1]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 461 | "bgt 1b \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 462 | : "+r"(src_ptr), // %0 |
| 463 | "+r"(dst_ptr), // %1 |
| 464 | "+r"(dst_width), // %2 |
| 465 | "+r"(src_stride) // %3 |
| 466 | : "r"(&kMult38_Div6), // %4 |
| 467 | "r"(&kShuf38_2) // %5 |
| 468 | : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 469 | ); |
| 470 | } |
| 471 | |
| 472 | // 16x2 -> 16x1 |
| 473 | void ScaleFilterRows_NEON(uint8* dst_ptr, |
| 474 | const uint8* src_ptr, ptrdiff_t src_stride, |
| 475 | int dst_width, int source_y_fraction) { |
| 476 | asm volatile ( |
| 477 | "cmp %4, #0 \n" |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 478 | "beq 100f \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 479 | "add %2, %1 \n" |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 480 | "cmp %4, #64 \n" |
| 481 | "beq 75f \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 482 | "cmp %4, #128 \n" |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 483 | "beq 50f \n" |
| 484 | "cmp %4, #192 \n" |
| 485 | "beq 25f \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 486 | |
| 487 | "vdup.8 d5, %4 \n" |
| 488 | "rsb %4, #256 \n" |
| 489 | "vdup.8 d4, %4 \n" |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 490 | // General purpose row blend. |
fbarchard@google.com | 31d0503 | 2012-10-23 09:04:37 +0000 | [diff] [blame] | 491 | "1: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 492 | "vld1.u8 {q0}, [%1]! \n" |
| 493 | "vld1.u8 {q1}, [%2]! \n" |
| 494 | "subs %3, #16 \n" |
| 495 | "vmull.u8 q13, d0, d4 \n" |
| 496 | "vmull.u8 q14, d1, d4 \n" |
| 497 | "vmlal.u8 q13, d2, d5 \n" |
| 498 | "vmlal.u8 q14, d3, d5 \n" |
| 499 | "vrshrn.u16 d0, q13, #8 \n" |
| 500 | "vrshrn.u16 d1, q14, #8 \n" |
| 501 | "vst1.u8 {q0}, [%0]! \n" |
| 502 | "bgt 1b \n" |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 503 | "b 99f \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 504 | |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 505 | // Blend 25 / 75. |
| 506 | "25: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 507 | "vld1.u8 {q0}, [%1]! \n" |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 508 | "vld1.u8 {q1}, [%2]! \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 509 | "subs %3, #16 \n" |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 510 | "vrhadd.u8 q0, q1 \n" |
| 511 | "vrhadd.u8 q0, q1 \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 512 | "vst1.u8 {q0}, [%0]! \n" |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 513 | "bgt 25b \n" |
| 514 | "b 99f \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 515 | |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 516 | // Blend 50 / 50. |
| 517 | "50: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 518 | "vld1.u8 {q0}, [%1]! \n" |
| 519 | "vld1.u8 {q1}, [%2]! \n" |
| 520 | "subs %3, #16 \n" |
| 521 | "vrhadd.u8 q0, q1 \n" |
| 522 | "vst1.u8 {q0}, [%0]! \n" |
fbarchard@google.com | 66d16f4 | 2012-11-13 20:43:25 +0000 | [diff] [blame] | 523 | "bgt 50b \n" |
| 524 | "b 99f \n" |
| 525 | |
| 526 | // Blend 75 / 25. |
| 527 | "75: \n" |
| 528 | "vld1.u8 {q1}, [%1]! \n" |
| 529 | "vld1.u8 {q0}, [%2]! \n" |
| 530 | "subs %3, #16 \n" |
| 531 | "vrhadd.u8 q0, q1 \n" |
| 532 | "vrhadd.u8 q0, q1 \n" |
| 533 | "vst1.u8 {q0}, [%0]! \n" |
| 534 | "bgt 75b \n" |
| 535 | "b 99f \n" |
| 536 | |
| 537 | // Blend 100 / 0 - Copy row unchanged. |
| 538 | "100: \n" |
| 539 | "vld1.u8 {q0}, [%1]! \n" |
| 540 | "subs %3, #16 \n" |
| 541 | "vst1.u8 {q0}, [%0]! \n" |
| 542 | "bgt 100b \n" |
| 543 | |
| 544 | "99: \n" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 545 | "vst1.u8 {d1[7]}, [%0] \n" |
fbarchard@google.com | 8e50e61 | 2012-11-16 19:32:16 +0000 | [diff] [blame] | 546 | : "+r"(dst_ptr), // %0 |
| 547 | "+r"(src_ptr), // %1 |
| 548 | "+r"(src_stride), // %2 |
| 549 | "+r"(dst_width), // %3 |
| 550 | "+r"(source_y_fraction) // %4 |
| 551 | : |
| 552 | : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc" |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 553 | ); |
| 554 | } |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 555 | #endif // __ARM_NEON__ |
| 556 | |
| 557 | #ifdef __cplusplus |
| 558 | } // extern "C" |
| 559 | } // namespace libyuv |
| 560 | #endif |