warp_plane_neon_i8mm.c (12808B)
1 /* 2 * Copyright (c) 2023, Alliance for Open Media. All rights reserved. 3 * 4 * This source code is subject to the terms of the BSD 2 Clause License and 5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 6 * was not distributed with this source code in the LICENSE file, you can 7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 8 * Media Patent License 1.0 was not distributed with this source code in the 9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 10 */ 11 12 #include "warp_plane_neon.h" 13 14 DECLARE_ALIGNED(16, static const uint8_t, usdot_permute_idx[48]) = { 15 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, 16 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10, 17 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 18 }; 19 20 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, 21 int sx, int alpha) { 22 // Only put the constant in every other lane to avoid double-counting when 23 // performing the pairwise add later. 24 const int32x4_t add_const = 25 vreinterpretq_s32_u64(vdupq_n_u64(1 << (8 + FILTER_BITS - 1))); 26 27 // Loading the 8 filter taps 28 int16x8_t f[4]; 29 load_filters_4(f, sx, alpha); 30 31 int8x16_t f01_u8 = vcombine_s8(vmovn_s16(f[0]), vmovn_s16(f[1])); 32 int8x16_t f23_u8 = vcombine_s8(vmovn_s16(f[2]), vmovn_s16(f[3])); 33 34 uint8x8_t in0 = vget_low_u8(in); 35 uint8x8_t in1 = vget_low_u8(vextq_u8(in, in, 1)); 36 uint8x8_t in2 = vget_low_u8(vextq_u8(in, in, 2)); 37 uint8x8_t in3 = vget_low_u8(vextq_u8(in, in, 3)); 38 39 int32x4_t m01 = vusdotq_s32(add_const, vcombine_u8(in0, in1), f01_u8); 40 int32x4_t m23 = vusdotq_s32(add_const, vcombine_u8(in2, in3), f23_u8); 41 42 int32x4_t m0123 = vpaddq_s32(m01, m23); 43 44 uint16x8_t res = 45 vcombine_u16(vqrshrun_n_s32(m0123, ROUND0_BITS), vdup_n_u16(0)); 46 return vreinterpretq_s16_u16(res); 47 } 48 49 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, 50 int sx, int alpha) { 51 // Only put the constant in every other lane to avoid double-counting when 52 // performing the pairwise add later. 53 const int32x4_t add_const = 54 vreinterpretq_s32_u64(vdupq_n_u64(1 << (8 + FILTER_BITS - 1))); 55 56 // Loading the 8 filter taps 57 int16x8_t f[8]; 58 load_filters_8(f, sx, alpha); 59 60 int8x16_t f01_u8 = vcombine_s8(vmovn_s16(f[0]), vmovn_s16(f[1])); 61 int8x16_t f23_u8 = vcombine_s8(vmovn_s16(f[2]), vmovn_s16(f[3])); 62 int8x16_t f45_u8 = vcombine_s8(vmovn_s16(f[4]), vmovn_s16(f[5])); 63 int8x16_t f67_u8 = vcombine_s8(vmovn_s16(f[6]), vmovn_s16(f[7])); 64 65 uint8x8_t in0 = vget_low_u8(in); 66 uint8x8_t in1 = vget_low_u8(vextq_u8(in, in, 1)); 67 uint8x8_t in2 = vget_low_u8(vextq_u8(in, in, 2)); 68 uint8x8_t in3 = vget_low_u8(vextq_u8(in, in, 3)); 69 uint8x8_t in4 = vget_low_u8(vextq_u8(in, in, 4)); 70 uint8x8_t in5 = vget_low_u8(vextq_u8(in, in, 5)); 71 uint8x8_t in6 = vget_low_u8(vextq_u8(in, in, 6)); 72 uint8x8_t in7 = vget_low_u8(vextq_u8(in, in, 7)); 73 74 int32x4_t m01 = vusdotq_s32(add_const, vcombine_u8(in0, in1), f01_u8); 75 int32x4_t m23 = vusdotq_s32(add_const, vcombine_u8(in2, in3), f23_u8); 76 int32x4_t m45 = vusdotq_s32(add_const, vcombine_u8(in4, in5), f45_u8); 77 int32x4_t m67 = vusdotq_s32(add_const, vcombine_u8(in6, in7), f67_u8); 78 79 int32x4_t m0123 = vpaddq_s32(m01, m23); 80 int32x4_t m4567 = vpaddq_s32(m45, m67); 81 82 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(m0123, ROUND0_BITS), 83 vqrshrun_n_s32(m4567, ROUND0_BITS)); 84 return vreinterpretq_s16_u16(res); 85 } 86 87 static AOM_FORCE_INLINE int16x8_t 88 horizontal_filter_4x1_f1_beta0(const uint8x16_t in, int16x8_t f_s16) { 89 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); 90 91 int8x16_t f_s8 = vcombine_s8(vmovn_s16(f_s16), vmovn_s16(f_s16)); 92 93 uint8x16_t perm0 = vld1q_u8(&usdot_permute_idx[0]); 94 uint8x16_t perm1 = vld1q_u8(&usdot_permute_idx[16]); 95 96 // Permute samples ready for dot product. 97 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } 98 // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } 99 uint8x16_t in_0123 = vqtbl1q_u8(in, perm0); 100 uint8x16_t in_4567 = vqtbl1q_u8(in, perm1); 101 102 int32x4_t m0123 = vusdotq_laneq_s32(add_const, in_0123, f_s8, 0); 103 m0123 = vusdotq_laneq_s32(m0123, in_4567, f_s8, 1); 104 105 uint16x8_t res = 106 vcombine_u16(vqrshrun_n_s32(m0123, ROUND0_BITS), vdup_n_u16(0)); 107 return vreinterpretq_s16_u16(res); 108 } 109 110 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, 111 int sx) { 112 int16x8_t f_s16 = vld1q_s16(av1_warped_filter[sx >> WARPEDDIFF_PREC_BITS]); 113 return horizontal_filter_4x1_f1_beta0(in, f_s16); 114 } 115 116 static AOM_FORCE_INLINE int16x8_t 117 horizontal_filter_8x1_f1_beta0(const uint8x16_t in, int16x8_t f_s16) { 118 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); 119 120 int8x16_t f_s8 = vcombine_s8(vmovn_s16(f_s16), vmovn_s16(f_s16)); 121 122 uint8x16_t perm0 = vld1q_u8(&usdot_permute_idx[0]); 123 uint8x16_t perm1 = vld1q_u8(&usdot_permute_idx[16]); 124 uint8x16_t perm2 = vld1q_u8(&usdot_permute_idx[32]); 125 126 // Permute samples ready for dot product. 127 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } 128 // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } 129 // { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 } 130 uint8x16_t in_0123 = vqtbl1q_u8(in, perm0); 131 uint8x16_t in_4567 = vqtbl1q_u8(in, perm1); 132 uint8x16_t in_89ab = vqtbl1q_u8(in, perm2); 133 134 int32x4_t m0123 = vusdotq_laneq_s32(add_const, in_0123, f_s8, 0); 135 m0123 = vusdotq_laneq_s32(m0123, in_4567, f_s8, 1); 136 137 int32x4_t m4567 = vusdotq_laneq_s32(add_const, in_4567, f_s8, 0); 138 m4567 = vusdotq_laneq_s32(m4567, in_89ab, f_s8, 1); 139 140 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(m0123, ROUND0_BITS), 141 vqrshrun_n_s32(m4567, ROUND0_BITS)); 142 return vreinterpretq_s16_u16(res); 143 } 144 145 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, 146 int sx) { 147 int16x8_t f_s16 = vld1q_s16(av1_warped_filter[sx >> WARPEDDIFF_PREC_BITS]); 148 return horizontal_filter_8x1_f1_beta0(in, f_s16); 149 } 150 151 static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src, 152 int32x4_t *res, int sy) { 153 int16x4_t s0 = vget_low_s16(src[0]); 154 int16x4_t s1 = vget_low_s16(src[1]); 155 int16x4_t s2 = vget_low_s16(src[2]); 156 int16x4_t s3 = vget_low_s16(src[3]); 157 int16x4_t s4 = vget_low_s16(src[4]); 158 int16x4_t s5 = vget_low_s16(src[5]); 159 int16x4_t s6 = vget_low_s16(src[6]); 160 int16x4_t s7 = vget_low_s16(src[7]); 161 162 int16x8_t f = vld1q_s16(av1_warped_filter[sy >> WARPEDDIFF_PREC_BITS]); 163 164 int32x4_t m0123 = vmull_lane_s16(s0, vget_low_s16(f), 0); 165 m0123 = vmlal_lane_s16(m0123, s1, vget_low_s16(f), 1); 166 m0123 = vmlal_lane_s16(m0123, s2, vget_low_s16(f), 2); 167 m0123 = vmlal_lane_s16(m0123, s3, vget_low_s16(f), 3); 168 m0123 = vmlal_lane_s16(m0123, s4, vget_high_s16(f), 0); 169 m0123 = vmlal_lane_s16(m0123, s5, vget_high_s16(f), 1); 170 m0123 = vmlal_lane_s16(m0123, s6, vget_high_s16(f), 2); 171 m0123 = vmlal_lane_s16(m0123, s7, vget_high_s16(f), 3); 172 173 *res = m0123; 174 } 175 176 static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src, 177 int32x4_t *res, int sy, 178 int gamma) { 179 int16x8_t s0, s1, s2, s3; 180 transpose_elems_s16_4x8( 181 vget_low_s16(src[0]), vget_low_s16(src[1]), vget_low_s16(src[2]), 182 vget_low_s16(src[3]), vget_low_s16(src[4]), vget_low_s16(src[5]), 183 vget_low_s16(src[6]), vget_low_s16(src[7]), &s0, &s1, &s2, &s3); 184 185 int16x8_t f[4]; 186 load_filters_4(f, sy, gamma); 187 188 int32x4_t m0 = vmull_s16(vget_low_s16(s0), vget_low_s16(f[0])); 189 m0 = vmlal_s16(m0, vget_high_s16(s0), vget_high_s16(f[0])); 190 int32x4_t m1 = vmull_s16(vget_low_s16(s1), vget_low_s16(f[1])); 191 m1 = vmlal_s16(m1, vget_high_s16(s1), vget_high_s16(f[1])); 192 int32x4_t m2 = vmull_s16(vget_low_s16(s2), vget_low_s16(f[2])); 193 m2 = vmlal_s16(m2, vget_high_s16(s2), vget_high_s16(f[2])); 194 int32x4_t m3 = vmull_s16(vget_low_s16(s3), vget_low_s16(f[3])); 195 m3 = vmlal_s16(m3, vget_high_s16(s3), vget_high_s16(f[3])); 196 197 int32x4_t m0123_pairs[] = { m0, m1, m2, m3 }; 198 199 *res = horizontal_add_4d_s32x4(m0123_pairs); 200 } 201 202 static AOM_FORCE_INLINE void vertical_filter_8x1_f1(const int16x8_t *src, 203 int32x4_t *res_low, 204 int32x4_t *res_high, 205 int sy) { 206 int16x8_t s0 = src[0]; 207 int16x8_t s1 = src[1]; 208 int16x8_t s2 = src[2]; 209 int16x8_t s3 = src[3]; 210 int16x8_t s4 = src[4]; 211 int16x8_t s5 = src[5]; 212 int16x8_t s6 = src[6]; 213 int16x8_t s7 = src[7]; 214 215 int16x8_t f = vld1q_s16(av1_warped_filter[sy >> WARPEDDIFF_PREC_BITS]); 216 217 int32x4_t m0123 = vmull_lane_s16(vget_low_s16(s0), vget_low_s16(f), 0); 218 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s1), vget_low_s16(f), 1); 219 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s2), vget_low_s16(f), 2); 220 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s3), vget_low_s16(f), 3); 221 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s4), vget_high_s16(f), 0); 222 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s5), vget_high_s16(f), 1); 223 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s6), vget_high_s16(f), 2); 224 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s7), vget_high_s16(f), 3); 225 226 int32x4_t m4567 = vmull_lane_s16(vget_high_s16(s0), vget_low_s16(f), 0); 227 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s1), vget_low_s16(f), 1); 228 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s2), vget_low_s16(f), 2); 229 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s3), vget_low_s16(f), 3); 230 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s4), vget_high_s16(f), 0); 231 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s5), vget_high_s16(f), 1); 232 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s6), vget_high_s16(f), 2); 233 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s7), vget_high_s16(f), 3); 234 235 *res_low = m0123; 236 *res_high = m4567; 237 } 238 239 static AOM_FORCE_INLINE void vertical_filter_8x1_f8(const int16x8_t *src, 240 int32x4_t *res_low, 241 int32x4_t *res_high, int sy, 242 int gamma) { 243 int16x8_t s0 = src[0]; 244 int16x8_t s1 = src[1]; 245 int16x8_t s2 = src[2]; 246 int16x8_t s3 = src[3]; 247 int16x8_t s4 = src[4]; 248 int16x8_t s5 = src[5]; 249 int16x8_t s6 = src[6]; 250 int16x8_t s7 = src[7]; 251 transpose_elems_inplace_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7); 252 253 int16x8_t f[8]; 254 load_filters_8(f, sy, gamma); 255 256 int32x4_t m0 = vmull_s16(vget_low_s16(s0), vget_low_s16(f[0])); 257 m0 = vmlal_s16(m0, vget_high_s16(s0), vget_high_s16(f[0])); 258 int32x4_t m1 = vmull_s16(vget_low_s16(s1), vget_low_s16(f[1])); 259 m1 = vmlal_s16(m1, vget_high_s16(s1), vget_high_s16(f[1])); 260 int32x4_t m2 = vmull_s16(vget_low_s16(s2), vget_low_s16(f[2])); 261 m2 = vmlal_s16(m2, vget_high_s16(s2), vget_high_s16(f[2])); 262 int32x4_t m3 = vmull_s16(vget_low_s16(s3), vget_low_s16(f[3])); 263 m3 = vmlal_s16(m3, vget_high_s16(s3), vget_high_s16(f[3])); 264 int32x4_t m4 = vmull_s16(vget_low_s16(s4), vget_low_s16(f[4])); 265 m4 = vmlal_s16(m4, vget_high_s16(s4), vget_high_s16(f[4])); 266 int32x4_t m5 = vmull_s16(vget_low_s16(s5), vget_low_s16(f[5])); 267 m5 = vmlal_s16(m5, vget_high_s16(s5), vget_high_s16(f[5])); 268 int32x4_t m6 = vmull_s16(vget_low_s16(s6), vget_low_s16(f[6])); 269 m6 = vmlal_s16(m6, vget_high_s16(s6), vget_high_s16(f[6])); 270 int32x4_t m7 = vmull_s16(vget_low_s16(s7), vget_low_s16(f[7])); 271 m7 = vmlal_s16(m7, vget_high_s16(s7), vget_high_s16(f[7])); 272 273 int32x4_t m0123_pairs[] = { m0, m1, m2, m3 }; 274 int32x4_t m4567_pairs[] = { m4, m5, m6, m7 }; 275 276 *res_low = horizontal_add_4d_s32x4(m0123_pairs); 277 *res_high = horizontal_add_4d_s32x4(m4567_pairs); 278 } 279 280 void av1_warp_affine_neon_i8mm(const int32_t *mat, const uint8_t *ref, 281 int width, int height, int stride, uint8_t *pred, 282 int p_col, int p_row, int p_width, int p_height, 283 int p_stride, int subsampling_x, 284 int subsampling_y, ConvolveParams *conv_params, 285 int16_t alpha, int16_t beta, int16_t gamma, 286 int16_t delta) { 287 av1_warp_affine_common(mat, ref, width, height, stride, pred, p_col, p_row, 288 p_width, p_height, p_stride, subsampling_x, 289 subsampling_y, conv_params, alpha, beta, gamma, delta); 290 }