frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 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 | |
fbarchard@google.com | 2b9c210 | 2012-03-22 22:36:44 +0000 | [diff] [blame] | 11 | #include "source/row.h" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 12 | |
fbarchard@google.com | 17f198c | 2012-01-04 02:21:05 +0000 | [diff] [blame] | 13 | #include "libyuv/basic_types.h" |
| 14 | |
fbarchard@google.com | fe5ff7e | 2011-12-10 07:45:58 +0000 | [diff] [blame] | 15 | #ifdef __cplusplus |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 16 | namespace libyuv { |
fbarchard@google.com | fe5ff7e | 2011-12-10 07:45:58 +0000 | [diff] [blame] | 17 | extern "C" { |
| 18 | #endif |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 19 | |
fbarchard@google.com | 15c3d45 | 2011-11-17 22:13:17 +0000 | [diff] [blame] | 20 | #if defined(__ARM_NEON__) && !defined(YUV_DISABLE_ASM) |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 21 | |
fbarchard@google.com | 8536b2f | 2012-02-21 21:02:54 +0000 | [diff] [blame] | 22 | static const uvec8 vtbl_4x4_transpose = |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 23 | { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 }; |
| 24 | |
| 25 | void TransposeWx8_NEON(const uint8* src, int src_stride, |
| 26 | uint8* dst, int dst_stride, |
| 27 | int width) { |
fbarchard@google.com | 5b22506 | 2012-03-29 02:19:26 +0000 | [diff] [blame^] | 28 | asm volatile ( |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 29 | // loops are on blocks of 8. loop will stop when |
| 30 | // counter gets to or below 0. starting the counter |
| 31 | // at w-8 allow for this |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 32 | "sub %4, #8 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 33 | |
| 34 | // handle 8x8 blocks. this should be the majority of the plane |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 35 | "1: \n" |
| 36 | "mov r9, %0 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 37 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 38 | "vld1.8 {d0}, [r9], %1 \n" |
| 39 | "vld1.8 {d1}, [r9], %1 \n" |
| 40 | "vld1.8 {d2}, [r9], %1 \n" |
| 41 | "vld1.8 {d3}, [r9], %1 \n" |
| 42 | "vld1.8 {d4}, [r9], %1 \n" |
| 43 | "vld1.8 {d5}, [r9], %1 \n" |
| 44 | "vld1.8 {d6}, [r9], %1 \n" |
| 45 | "vld1.8 {d7}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 46 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 47 | "vtrn.8 d1, d0 \n" |
| 48 | "vtrn.8 d3, d2 \n" |
| 49 | "vtrn.8 d5, d4 \n" |
| 50 | "vtrn.8 d7, d6 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 51 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 52 | "vtrn.16 d1, d3 \n" |
| 53 | "vtrn.16 d0, d2 \n" |
| 54 | "vtrn.16 d5, d7 \n" |
| 55 | "vtrn.16 d4, d6 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 56 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 57 | "vtrn.32 d1, d5 \n" |
| 58 | "vtrn.32 d0, d4 \n" |
| 59 | "vtrn.32 d3, d7 \n" |
| 60 | "vtrn.32 d2, d6 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 61 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 62 | "vrev16.8 q0, q0 \n" |
| 63 | "vrev16.8 q1, q1 \n" |
| 64 | "vrev16.8 q2, q2 \n" |
| 65 | "vrev16.8 q3, q3 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 66 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 67 | "mov r9, %2 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 68 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 69 | "vst1.8 {d1}, [r9], %3 \n" |
| 70 | "vst1.8 {d0}, [r9], %3 \n" |
| 71 | "vst1.8 {d3}, [r9], %3 \n" |
| 72 | "vst1.8 {d2}, [r9], %3 \n" |
| 73 | "vst1.8 {d5}, [r9], %3 \n" |
| 74 | "vst1.8 {d4}, [r9], %3 \n" |
| 75 | "vst1.8 {d7}, [r9], %3 \n" |
| 76 | "vst1.8 {d6}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 77 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 78 | "add %0, #8 \n" // src += 8 |
| 79 | "add %2, %2, %3, lsl #3 \n" // dst += 8 * dst_stride |
| 80 | "subs %4, #8 \n" // w -= 8 |
| 81 | "bge 1b \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 82 | |
| 83 | // add 8 back to counter. if the result is 0 there are |
| 84 | // no residuals. |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 85 | "adds %4, #8 \n" |
| 86 | "beq 4f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 87 | |
| 88 | // some residual, so between 1 and 7 lines left to transpose |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 89 | "cmp %4, #2 \n" |
| 90 | "blt 3f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 91 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 92 | "cmp %4, #4 \n" |
| 93 | "blt 2f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 94 | |
| 95 | // 4x8 block |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 96 | "mov r9, %0 \n" |
| 97 | "vld1.32 {d0[0]}, [r9], %1 \n" |
| 98 | "vld1.32 {d0[1]}, [r9], %1 \n" |
| 99 | "vld1.32 {d1[0]}, [r9], %1 \n" |
| 100 | "vld1.32 {d1[1]}, [r9], %1 \n" |
| 101 | "vld1.32 {d2[0]}, [r9], %1 \n" |
| 102 | "vld1.32 {d2[1]}, [r9], %1 \n" |
| 103 | "vld1.32 {d3[0]}, [r9], %1 \n" |
| 104 | "vld1.32 {d3[1]}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 105 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 106 | "mov r9, %2 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 107 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 108 | "vld1.8 {q3}, [%5] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 109 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 110 | "vtbl.8 d4, {d0, d1}, d6 \n" |
| 111 | "vtbl.8 d5, {d0, d1}, d7 \n" |
| 112 | "vtbl.8 d0, {d2, d3}, d6 \n" |
| 113 | "vtbl.8 d1, {d2, d3}, d7 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 114 | |
| 115 | // TODO: rework shuffle above to write |
| 116 | // out with 4 instead of 8 writes |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 117 | "vst1.32 {d4[0]}, [r9], %3 \n" |
| 118 | "vst1.32 {d4[1]}, [r9], %3 \n" |
| 119 | "vst1.32 {d5[0]}, [r9], %3 \n" |
| 120 | "vst1.32 {d5[1]}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 121 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 122 | "add r9, %2, #4 \n" |
| 123 | "vst1.32 {d0[0]}, [r9], %3 \n" |
| 124 | "vst1.32 {d0[1]}, [r9], %3 \n" |
| 125 | "vst1.32 {d1[0]}, [r9], %3 \n" |
| 126 | "vst1.32 {d1[1]}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 127 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 128 | "add %0, #4 \n" // src += 4 |
| 129 | "add %2, %2, %3, lsl #2 \n" // dst += 4 * dst_stride |
| 130 | "subs %4, #4 \n" // w -= 4 |
| 131 | "beq 4f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 132 | |
| 133 | // some residual, check to see if it includes a 2x8 block, |
| 134 | // or less |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 135 | "cmp %4, #2 \n" |
| 136 | "blt 3f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 137 | |
| 138 | // 2x8 block |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 139 | "2: \n" |
| 140 | "mov r9, %0 \n" |
| 141 | "vld1.16 {d0[0]}, [r9], %1 \n" |
| 142 | "vld1.16 {d1[0]}, [r9], %1 \n" |
| 143 | "vld1.16 {d0[1]}, [r9], %1 \n" |
| 144 | "vld1.16 {d1[1]}, [r9], %1 \n" |
| 145 | "vld1.16 {d0[2]}, [r9], %1 \n" |
| 146 | "vld1.16 {d1[2]}, [r9], %1 \n" |
| 147 | "vld1.16 {d0[3]}, [r9], %1 \n" |
| 148 | "vld1.16 {d1[3]}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 149 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 150 | "vtrn.8 d0, d1 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 151 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 152 | "mov r9, %2 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 153 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 154 | "vst1.64 {d0}, [r9], %3 \n" |
| 155 | "vst1.64 {d1}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 156 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 157 | "add %0, #2 \n" // src += 2 |
| 158 | "add %2, %2, %3, lsl #1 \n" // dst += 2 * dst_stride |
| 159 | "subs %4, #2 \n" // w -= 2 |
| 160 | "beq 4f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 161 | |
| 162 | // 1x8 block |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 163 | "3: \n" |
| 164 | "vld1.8 {d0[0]}, [%0], %1 \n" |
| 165 | "vld1.8 {d0[1]}, [%0], %1 \n" |
| 166 | "vld1.8 {d0[2]}, [%0], %1 \n" |
| 167 | "vld1.8 {d0[3]}, [%0], %1 \n" |
| 168 | "vld1.8 {d0[4]}, [%0], %1 \n" |
| 169 | "vld1.8 {d0[5]}, [%0], %1 \n" |
| 170 | "vld1.8 {d0[6]}, [%0], %1 \n" |
| 171 | "vld1.8 {d0[7]}, [%0] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 172 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 173 | "vst1.64 {d0}, [%2] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 174 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 175 | "4: \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 176 | |
| 177 | : "+r"(src), // %0 |
| 178 | "+r"(src_stride), // %1 |
| 179 | "+r"(dst), // %2 |
| 180 | "+r"(dst_stride), // %3 |
| 181 | "+r"(width) // %4 |
| 182 | : "r"(vtbl_4x4_transpose) // %5 |
| 183 | : "memory", "cc", "r9", "q0", "q1", "q2", "q3" |
| 184 | ); |
| 185 | } |
| 186 | |
fbarchard@google.com | 8536b2f | 2012-02-21 21:02:54 +0000 | [diff] [blame] | 187 | static const uvec8 vtbl_4x4_transpose_di = |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 188 | { 0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15 }; |
| 189 | |
| 190 | void TransposeUVWx8_NEON(const uint8* src, int src_stride, |
| 191 | uint8* dst_a, int dst_stride_a, |
| 192 | uint8* dst_b, int dst_stride_b, |
| 193 | int width) { |
fbarchard@google.com | 5b22506 | 2012-03-29 02:19:26 +0000 | [diff] [blame^] | 194 | asm volatile ( |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 195 | // loops are on blocks of 8. loop will stop when |
| 196 | // counter gets to or below 0. starting the counter |
| 197 | // at w-8 allow for this |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 198 | "sub %6, #8 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 199 | |
| 200 | // handle 8x8 blocks. this should be the majority of the plane |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 201 | "1: \n" |
| 202 | "mov r9, %0 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 203 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 204 | "vld2.8 {d0, d1}, [r9], %1 \n" |
| 205 | "vld2.8 {d2, d3}, [r9], %1 \n" |
| 206 | "vld2.8 {d4, d5}, [r9], %1 \n" |
| 207 | "vld2.8 {d6, d7}, [r9], %1 \n" |
| 208 | "vld2.8 {d16, d17}, [r9], %1 \n" |
| 209 | "vld2.8 {d18, d19}, [r9], %1 \n" |
| 210 | "vld2.8 {d20, d21}, [r9], %1 \n" |
| 211 | "vld2.8 {d22, d23}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 212 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 213 | "vtrn.8 q1, q0 \n" |
| 214 | "vtrn.8 q3, q2 \n" |
| 215 | "vtrn.8 q9, q8 \n" |
| 216 | "vtrn.8 q11, q10 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 217 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 218 | "vtrn.16 q1, q3 \n" |
| 219 | "vtrn.16 q0, q2 \n" |
| 220 | "vtrn.16 q9, q11 \n" |
| 221 | "vtrn.16 q8, q10 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 222 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 223 | "vtrn.32 q1, q9 \n" |
| 224 | "vtrn.32 q0, q8 \n" |
| 225 | "vtrn.32 q3, q11 \n" |
| 226 | "vtrn.32 q2, q10 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 227 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 228 | "vrev16.8 q0, q0 \n" |
| 229 | "vrev16.8 q1, q1 \n" |
| 230 | "vrev16.8 q2, q2 \n" |
| 231 | "vrev16.8 q3, q3 \n" |
| 232 | "vrev16.8 q8, q8 \n" |
| 233 | "vrev16.8 q9, q9 \n" |
| 234 | "vrev16.8 q10, q10 \n" |
| 235 | "vrev16.8 q11, q11 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 236 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 237 | "mov r9, %2 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 238 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 239 | "vst1.8 {d2}, [r9], %3 \n" |
| 240 | "vst1.8 {d0}, [r9], %3 \n" |
| 241 | "vst1.8 {d6}, [r9], %3 \n" |
| 242 | "vst1.8 {d4}, [r9], %3 \n" |
| 243 | "vst1.8 {d18}, [r9], %3 \n" |
| 244 | "vst1.8 {d16}, [r9], %3 \n" |
| 245 | "vst1.8 {d22}, [r9], %3 \n" |
| 246 | "vst1.8 {d20}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 247 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 248 | "mov r9, %4 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 249 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 250 | "vst1.8 {d3}, [r9], %5 \n" |
| 251 | "vst1.8 {d1}, [r9], %5 \n" |
| 252 | "vst1.8 {d7}, [r9], %5 \n" |
| 253 | "vst1.8 {d5}, [r9], %5 \n" |
| 254 | "vst1.8 {d19}, [r9], %5 \n" |
| 255 | "vst1.8 {d17}, [r9], %5 \n" |
| 256 | "vst1.8 {d23}, [r9], %5 \n" |
| 257 | "vst1.8 {d21}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 258 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 259 | "add %0, #8*2 \n" // src += 8*2 |
| 260 | "add %2, %2, %3, lsl #3 \n" // dst_a += 8 * dst_stride_a |
| 261 | "add %4, %4, %5, lsl #3 \n" // dst_b += 8 * dst_stride_b |
| 262 | "subs %6, #8 \n" // w -= 8 |
| 263 | "bge 1b \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 264 | |
| 265 | // add 8 back to counter. if the result is 0 there are |
| 266 | // no residuals. |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 267 | "adds %6, #8 \n" |
| 268 | "beq 4f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 269 | |
| 270 | // some residual, so between 1 and 7 lines left to transpose |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 271 | "cmp %6, #2 \n" |
| 272 | "blt 3f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 273 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 274 | "cmp %6, #4 \n" |
| 275 | "blt 2f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 276 | |
| 277 | //TODO(frkoenig) : clean this up |
| 278 | // 4x8 block |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 279 | "mov r9, %0 \n" |
| 280 | "vld1.64 {d0}, [r9], %1 \n" |
| 281 | "vld1.64 {d1}, [r9], %1 \n" |
| 282 | "vld1.64 {d2}, [r9], %1 \n" |
| 283 | "vld1.64 {d3}, [r9], %1 \n" |
| 284 | "vld1.64 {d4}, [r9], %1 \n" |
| 285 | "vld1.64 {d5}, [r9], %1 \n" |
| 286 | "vld1.64 {d6}, [r9], %1 \n" |
| 287 | "vld1.64 {d7}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 288 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 289 | "vld1.8 {q15}, [%7] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 290 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 291 | "vtrn.8 q0, q1 \n" |
| 292 | "vtrn.8 q2, q3 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 293 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 294 | "vtbl.8 d16, {d0, d1}, d30 \n" |
| 295 | "vtbl.8 d17, {d0, d1}, d31 \n" |
| 296 | "vtbl.8 d18, {d2, d3}, d30 \n" |
| 297 | "vtbl.8 d19, {d2, d3}, d31 \n" |
| 298 | "vtbl.8 d20, {d4, d5}, d30 \n" |
| 299 | "vtbl.8 d21, {d4, d5}, d31 \n" |
| 300 | "vtbl.8 d22, {d6, d7}, d30 \n" |
| 301 | "vtbl.8 d23, {d6, d7}, d31 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 302 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 303 | "mov r9, %2 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 304 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 305 | "vst1.32 {d16[0]}, [r9], %3 \n" |
| 306 | "vst1.32 {d16[1]}, [r9], %3 \n" |
| 307 | "vst1.32 {d17[0]}, [r9], %3 \n" |
| 308 | "vst1.32 {d17[1]}, [r9], %3 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 309 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 310 | "add r9, %2, #4 \n" |
| 311 | "vst1.32 {d20[0]}, [r9], %3 \n" |
| 312 | "vst1.32 {d20[1]}, [r9], %3 \n" |
| 313 | "vst1.32 {d21[0]}, [r9], %3 \n" |
| 314 | "vst1.32 {d21[1]}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 315 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 316 | "mov r9, %4 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 317 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 318 | "vst1.32 {d18[0]}, [r9], %5 \n" |
| 319 | "vst1.32 {d18[1]}, [r9], %5 \n" |
| 320 | "vst1.32 {d19[0]}, [r9], %5 \n" |
| 321 | "vst1.32 {d19[1]}, [r9], %5 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 322 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 323 | "add r9, %4, #4 \n" |
| 324 | "vst1.32 {d22[0]}, [r9], %5 \n" |
| 325 | "vst1.32 {d22[1]}, [r9], %5 \n" |
| 326 | "vst1.32 {d23[0]}, [r9], %5 \n" |
| 327 | "vst1.32 {d23[1]}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 328 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 329 | "add %0, #4*2 \n" // src += 4 * 2 |
| 330 | "add %2, %2, %3, lsl #2 \n" // dst_a += 4 * dst_stride_a |
| 331 | "add %4, %4, %5, lsl #2 \n" // dst_b += 4 * dst_stride_b |
| 332 | "subs %6, #4 \n" // w -= 4 |
| 333 | "beq 4f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 334 | |
| 335 | // some residual, check to see if it includes a 2x8 block, |
| 336 | // or less |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 337 | "cmp %6, #2 \n" |
| 338 | "blt 3f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 339 | |
| 340 | // 2x8 block |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 341 | "2: \n" |
| 342 | "mov r9, %0 \n" |
| 343 | "vld2.16 {d0[0], d2[0]}, [r9], %1 \n" |
| 344 | "vld2.16 {d1[0], d3[0]}, [r9], %1 \n" |
| 345 | "vld2.16 {d0[1], d2[1]}, [r9], %1 \n" |
| 346 | "vld2.16 {d1[1], d3[1]}, [r9], %1 \n" |
| 347 | "vld2.16 {d0[2], d2[2]}, [r9], %1 \n" |
| 348 | "vld2.16 {d1[2], d3[2]}, [r9], %1 \n" |
| 349 | "vld2.16 {d0[3], d2[3]}, [r9], %1 \n" |
| 350 | "vld2.16 {d1[3], d3[3]}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 351 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 352 | "vtrn.8 d0, d1 \n" |
| 353 | "vtrn.8 d2, d3 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 354 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 355 | "mov r9, %2 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 356 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 357 | "vst1.64 {d0}, [r9], %3 \n" |
| 358 | "vst1.64 {d2}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 359 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 360 | "mov r9, %4 \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 361 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 362 | "vst1.64 {d1}, [r9], %5 \n" |
| 363 | "vst1.64 {d3}, [r9] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 364 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 365 | "add %0, #2*2 \n" // src += 2 * 2 |
| 366 | "add %2, %2, %3, lsl #1 \n" // dst_a += 2 * dst_stride_a |
| 367 | "add %4, %4, %5, lsl #1 \n" // dst_b += 2 * dst_stride_b |
| 368 | "subs %6, #2 \n" // w -= 2 |
| 369 | "beq 4f \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 370 | |
| 371 | // 1x8 block |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 372 | "3: \n" |
| 373 | "vld2.8 {d0[0], d1[0]}, [%0], %1 \n" |
| 374 | "vld2.8 {d0[1], d1[1]}, [%0], %1 \n" |
| 375 | "vld2.8 {d0[2], d1[2]}, [%0], %1 \n" |
| 376 | "vld2.8 {d0[3], d1[3]}, [%0], %1 \n" |
| 377 | "vld2.8 {d0[4], d1[4]}, [%0], %1 \n" |
| 378 | "vld2.8 {d0[5], d1[5]}, [%0], %1 \n" |
| 379 | "vld2.8 {d0[6], d1[6]}, [%0], %1 \n" |
| 380 | "vld2.8 {d0[7], d1[7]}, [%0] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 381 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 382 | "vst1.64 {d0}, [%2] \n" |
| 383 | "vst1.64 {d1}, [%4] \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 384 | |
fbarchard@google.com | f7a5048 | 2011-11-10 22:41:20 +0000 | [diff] [blame] | 385 | "4: \n" |
frkoenig@google.com | f7e74a1 | 2011-11-03 22:41:59 +0000 | [diff] [blame] | 386 | |
| 387 | : "+r"(src), // %0 |
| 388 | "+r"(src_stride), // %1 |
| 389 | "+r"(dst_a), // %2 |
| 390 | "+r"(dst_stride_a), // %3 |
| 391 | "+r"(dst_b), // %4 |
| 392 | "+r"(dst_stride_b), // %5 |
| 393 | "+r"(width) // %6 |
| 394 | : "r"(vtbl_4x4_transpose_di)// %7 |
| 395 | : "memory", "cc", "r9", |
| 396 | "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11" |
| 397 | ); |
| 398 | } |
| 399 | #endif |
fbarchard@google.com | fe5ff7e | 2011-12-10 07:45:58 +0000 | [diff] [blame] | 400 | |
| 401 | #ifdef __cplusplus |
| 402 | } // extern "C" |
| 403 | } // namespace libyuv |
| 404 | #endif |