1 /* 2 * Copyright (c) 2014 The WebM 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 <arm_neon.h> 12 #include "./vpx_config.h" 13 14 #if (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7)) 15 static INLINE void write_2x8(unsigned char *dst, int pitch, 16 const uint8x8x2_t result, 17 const uint8x8x2_t result2) { 18 vst2_lane_u8(dst, result, 0); 19 dst += pitch; 20 vst2_lane_u8(dst, result, 1); 21 dst += pitch; 22 vst2_lane_u8(dst, result, 2); 23 dst += pitch; 24 vst2_lane_u8(dst, result, 3); 25 dst += pitch; 26 vst2_lane_u8(dst, result, 4); 27 dst += pitch; 28 vst2_lane_u8(dst, result, 5); 29 dst += pitch; 30 vst2_lane_u8(dst, result, 6); 31 dst += pitch; 32 vst2_lane_u8(dst, result, 7); 33 dst += pitch; 34 35 vst2_lane_u8(dst, result2, 0); 36 dst += pitch; 37 vst2_lane_u8(dst, result2, 1); 38 dst += pitch; 39 vst2_lane_u8(dst, result2, 2); 40 dst += pitch; 41 vst2_lane_u8(dst, result2, 3); 42 dst += pitch; 43 vst2_lane_u8(dst, result2, 4); 44 dst += pitch; 45 vst2_lane_u8(dst, result2, 5); 46 dst += pitch; 47 vst2_lane_u8(dst, result2, 6); 48 dst += pitch; 49 vst2_lane_u8(dst, result2, 7); 50 } 51 #else 52 static INLINE void write_2x4(unsigned char *dst, int pitch, 53 const uint8x8x2_t result) { 54 /* 55 * uint8x8x2_t result 56 00 01 02 03 | 04 05 06 07 57 10 11 12 13 | 14 15 16 17 58 --- 59 * after vtrn_u8 60 00 10 02 12 | 04 14 06 16 61 01 11 03 13 | 05 15 07 17 62 */ 63 const uint8x8x2_t r01_u8 = vtrn_u8(result.val[0], 64 result.val[1]); 65 const uint16x4_t x_0_4 = vreinterpret_u16_u8(r01_u8.val[0]); 66 const uint16x4_t x_1_5 = vreinterpret_u16_u8(r01_u8.val[1]); 67 vst1_lane_u16((uint16_t *)dst, x_0_4, 0); 68 dst += pitch; 69 vst1_lane_u16((uint16_t *)dst, x_1_5, 0); 70 dst += pitch; 71 vst1_lane_u16((uint16_t *)dst, x_0_4, 1); 72 dst += pitch; 73 vst1_lane_u16((uint16_t *)dst, x_1_5, 1); 74 dst += pitch; 75 vst1_lane_u16((uint16_t *)dst, x_0_4, 2); 76 dst += pitch; 77 vst1_lane_u16((uint16_t *)dst, x_1_5, 2); 78 dst += pitch; 79 vst1_lane_u16((uint16_t *)dst, x_0_4, 3); 80 dst += pitch; 81 vst1_lane_u16((uint16_t *)dst, x_1_5, 3); 82 } 83 84 static INLINE void write_2x8(unsigned char *dst, int pitch, 85 const uint8x8x2_t result, 86 const uint8x8x2_t result2) { 87 write_2x4(dst, pitch, result); 88 dst += pitch * 8; 89 write_2x4(dst, pitch, result2); 90 } 91 #endif 92 93 94 #if (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7)) 95 static INLINE 96 uint8x8x4_t read_4x8(unsigned char *src, int pitch, uint8x8x4_t x) { 97 x = vld4_lane_u8(src, x, 0); 98 src += pitch; 99 x = vld4_lane_u8(src, x, 1); 100 src += pitch; 101 x = vld4_lane_u8(src, x, 2); 102 src += pitch; 103 x = vld4_lane_u8(src, x, 3); 104 src += pitch; 105 x = vld4_lane_u8(src, x, 4); 106 src += pitch; 107 x = vld4_lane_u8(src, x, 5); 108 src += pitch; 109 x = vld4_lane_u8(src, x, 6); 110 src += pitch; 111 x = vld4_lane_u8(src, x, 7); 112 return x; 113 } 114 #else 115 static INLINE 116 uint8x8x4_t read_4x8(unsigned char *src, int pitch, uint8x8x4_t x) { 117 const uint8x8_t a = vld1_u8(src); 118 const uint8x8_t b = vld1_u8(src + pitch * 1); 119 const uint8x8_t c = vld1_u8(src + pitch * 2); 120 const uint8x8_t d = vld1_u8(src + pitch * 3); 121 const uint8x8_t e = vld1_u8(src + pitch * 4); 122 const uint8x8_t f = vld1_u8(src + pitch * 5); 123 const uint8x8_t g = vld1_u8(src + pitch * 6); 124 const uint8x8_t h = vld1_u8(src + pitch * 7); 125 const uint32x2x2_t r04_u32 = vtrn_u32(vreinterpret_u32_u8(a), 126 vreinterpret_u32_u8(e)); 127 const uint32x2x2_t r15_u32 = vtrn_u32(vreinterpret_u32_u8(b), 128 vreinterpret_u32_u8(f)); 129 const uint32x2x2_t r26_u32 = vtrn_u32(vreinterpret_u32_u8(c), 130 vreinterpret_u32_u8(g)); 131 const uint32x2x2_t r37_u32 = vtrn_u32(vreinterpret_u32_u8(d), 132 vreinterpret_u32_u8(h)); 133 const uint16x4x2_t r02_u16 = vtrn_u16(vreinterpret_u16_u32(r04_u32.val[0]), 134 vreinterpret_u16_u32(r26_u32.val[0])); 135 const uint16x4x2_t r13_u16 = vtrn_u16(vreinterpret_u16_u32(r15_u32.val[0]), 136 vreinterpret_u16_u32(r37_u32.val[0])); 137 const uint8x8x2_t r01_u8 = vtrn_u8(vreinterpret_u8_u16(r02_u16.val[0]), 138 vreinterpret_u8_u16(r13_u16.val[0])); 139 const uint8x8x2_t r23_u8 = vtrn_u8(vreinterpret_u8_u16(r02_u16.val[1]), 140 vreinterpret_u8_u16(r13_u16.val[1])); 141 /* 142 * after vtrn_u32 143 00 01 02 03 | 40 41 42 43 144 10 11 12 13 | 50 51 52 53 145 20 21 22 23 | 60 61 62 63 146 30 31 32 33 | 70 71 72 73 147 --- 148 * after vtrn_u16 149 00 01 20 21 | 40 41 60 61 150 02 03 22 23 | 42 43 62 63 151 10 11 30 31 | 50 51 70 71 152 12 13 32 33 | 52 52 72 73 153 154 00 01 20 21 | 40 41 60 61 155 10 11 30 31 | 50 51 70 71 156 02 03 22 23 | 42 43 62 63 157 12 13 32 33 | 52 52 72 73 158 --- 159 * after vtrn_u8 160 00 10 20 30 | 40 50 60 70 161 01 11 21 31 | 41 51 61 71 162 02 12 22 32 | 42 52 62 72 163 03 13 23 33 | 43 53 63 73 164 */ 165 x.val[0] = r01_u8.val[0]; 166 x.val[1] = r01_u8.val[1]; 167 x.val[2] = r23_u8.val[0]; 168 x.val[3] = r23_u8.val[1]; 169 170 return x; 171 } 172 #endif 173 174 static INLINE void vp8_loop_filter_simple_vertical_edge_neon( 175 unsigned char *s, 176 int p, 177 const unsigned char *blimit) { 178 unsigned char *src1; 179 uint8x16_t qblimit, q0u8; 180 uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q11u8, q12u8, q14u8, q15u8; 181 int16x8_t q2s16, q13s16, q11s16; 182 int8x8_t d28s8, d29s8; 183 int8x16_t q2s8, q3s8, q10s8, q11s8, q14s8; 184 uint8x8x4_t d0u8x4; // d6, d7, d8, d9 185 uint8x8x4_t d1u8x4; // d10, d11, d12, d13 186 uint8x8x2_t d2u8x2; // d12, d13 187 uint8x8x2_t d3u8x2; // d14, d15 188 189 qblimit = vdupq_n_u8(*blimit); 190 191 src1 = s - 2; 192 d0u8x4 = read_4x8(src1, p, d0u8x4); 193 src1 += p * 8; 194 d1u8x4 = read_4x8(src1, p, d1u8x4); 195 196 q3u8 = vcombine_u8(d0u8x4.val[0], d1u8x4.val[0]); // d6 d10 197 q4u8 = vcombine_u8(d0u8x4.val[2], d1u8x4.val[2]); // d8 d12 198 q5u8 = vcombine_u8(d0u8x4.val[1], d1u8x4.val[1]); // d7 d11 199 q6u8 = vcombine_u8(d0u8x4.val[3], d1u8x4.val[3]); // d9 d13 200 201 q15u8 = vabdq_u8(q5u8, q4u8); 202 q14u8 = vabdq_u8(q3u8, q6u8); 203 204 q15u8 = vqaddq_u8(q15u8, q15u8); 205 q14u8 = vshrq_n_u8(q14u8, 1); 206 q0u8 = vdupq_n_u8(0x80); 207 q11s16 = vdupq_n_s16(3); 208 q15u8 = vqaddq_u8(q15u8, q14u8); 209 210 q3u8 = veorq_u8(q3u8, q0u8); 211 q4u8 = veorq_u8(q4u8, q0u8); 212 q5u8 = veorq_u8(q5u8, q0u8); 213 q6u8 = veorq_u8(q6u8, q0u8); 214 215 q15u8 = vcgeq_u8(qblimit, q15u8); 216 217 q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q4u8)), 218 vget_low_s8(vreinterpretq_s8_u8(q5u8))); 219 q13s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q4u8)), 220 vget_high_s8(vreinterpretq_s8_u8(q5u8))); 221 222 q14s8 = vqsubq_s8(vreinterpretq_s8_u8(q3u8), 223 vreinterpretq_s8_u8(q6u8)); 224 225 q2s16 = vmulq_s16(q2s16, q11s16); 226 q13s16 = vmulq_s16(q13s16, q11s16); 227 228 q11u8 = vdupq_n_u8(3); 229 q12u8 = vdupq_n_u8(4); 230 231 q2s16 = vaddw_s8(q2s16, vget_low_s8(q14s8)); 232 q13s16 = vaddw_s8(q13s16, vget_high_s8(q14s8)); 233 234 d28s8 = vqmovn_s16(q2s16); 235 d29s8 = vqmovn_s16(q13s16); 236 q14s8 = vcombine_s8(d28s8, d29s8); 237 238 q14s8 = vandq_s8(q14s8, vreinterpretq_s8_u8(q15u8)); 239 240 q2s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q11u8)); 241 q3s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q12u8)); 242 q2s8 = vshrq_n_s8(q2s8, 3); 243 q14s8 = vshrq_n_s8(q3s8, 3); 244 245 q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q5u8), q2s8); 246 q10s8 = vqsubq_s8(vreinterpretq_s8_u8(q4u8), q14s8); 247 248 q6u8 = veorq_u8(vreinterpretq_u8_s8(q11s8), q0u8); 249 q7u8 = veorq_u8(vreinterpretq_u8_s8(q10s8), q0u8); 250 251 d2u8x2.val[0] = vget_low_u8(q6u8); // d12 252 d2u8x2.val[1] = vget_low_u8(q7u8); // d14 253 d3u8x2.val[0] = vget_high_u8(q6u8); // d13 254 d3u8x2.val[1] = vget_high_u8(q7u8); // d15 255 256 src1 = s - 1; 257 write_2x8(src1, p, d2u8x2, d3u8x2); 258 } 259 260 void vp8_loop_filter_bvs_neon( 261 unsigned char *y_ptr, 262 int y_stride, 263 const unsigned char *blimit) { 264 y_ptr += 4; 265 vp8_loop_filter_simple_vertical_edge_neon(y_ptr, y_stride, blimit); 266 y_ptr += 4; 267 vp8_loop_filter_simple_vertical_edge_neon(y_ptr, y_stride, blimit); 268 y_ptr += 4; 269 vp8_loop_filter_simple_vertical_edge_neon(y_ptr, y_stride, blimit); 270 return; 271 } 272 273 void vp8_loop_filter_mbvs_neon( 274 unsigned char *y_ptr, 275 int y_stride, 276 const unsigned char *blimit) { 277 vp8_loop_filter_simple_vertical_edge_neon(y_ptr, y_stride, blimit); 278 return; 279 } 280
