warp_plane_sve.c (12077B)
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 <arm_neon.h> 13 14 #include "aom_dsp/arm/aom_neon_sve_bridge.h" 15 #include "warp_plane_neon.h" 16 17 DECLARE_ALIGNED(16, static const uint8_t, usdot_permute_idx[48]) = { 18 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, 19 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10, 20 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 21 }; 22 23 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in, 24 int sx, int alpha) { 25 // Only put the constant in every other lane to avoid double-counting when 26 // performing the pairwise add later. 27 const int32x4_t add_const = 28 vreinterpretq_s32_u64(vdupq_n_u64(1 << (8 + FILTER_BITS - 1))); 29 30 // Loading the 8 filter taps 31 int16x8_t f[4]; 32 load_filters_4(f, sx, alpha); 33 34 int8x16_t f01_u8 = vcombine_s8(vmovn_s16(f[0]), vmovn_s16(f[1])); 35 int8x16_t f23_u8 = vcombine_s8(vmovn_s16(f[2]), vmovn_s16(f[3])); 36 37 uint8x8_t in0 = vget_low_u8(in); 38 uint8x8_t in1 = vget_low_u8(vextq_u8(in, in, 1)); 39 uint8x8_t in2 = vget_low_u8(vextq_u8(in, in, 2)); 40 uint8x8_t in3 = vget_low_u8(vextq_u8(in, in, 3)); 41 42 int32x4_t m01 = vusdotq_s32(add_const, vcombine_u8(in0, in1), f01_u8); 43 int32x4_t m23 = vusdotq_s32(add_const, vcombine_u8(in2, in3), f23_u8); 44 45 int32x4_t m0123 = vpaddq_s32(m01, m23); 46 47 uint16x8_t res = 48 vcombine_u16(vqrshrun_n_s32(m0123, ROUND0_BITS), vdup_n_u16(0)); 49 return vreinterpretq_s16_u16(res); 50 } 51 52 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in, 53 int sx, int alpha) { 54 // Only put the constant in every other lane to avoid double-counting when 55 // performing the pairwise add later. 56 const int32x4_t add_const = 57 vreinterpretq_s32_u64(vdupq_n_u64(1 << (8 + FILTER_BITS - 1))); 58 59 // Loading the 8 filter taps 60 int16x8_t f[8]; 61 load_filters_8(f, sx, alpha); 62 63 int8x16_t f01_u8 = vcombine_s8(vmovn_s16(f[0]), vmovn_s16(f[1])); 64 int8x16_t f23_u8 = vcombine_s8(vmovn_s16(f[2]), vmovn_s16(f[3])); 65 int8x16_t f45_u8 = vcombine_s8(vmovn_s16(f[4]), vmovn_s16(f[5])); 66 int8x16_t f67_u8 = vcombine_s8(vmovn_s16(f[6]), vmovn_s16(f[7])); 67 68 uint8x8_t in0 = vget_low_u8(in); 69 uint8x8_t in1 = vget_low_u8(vextq_u8(in, in, 1)); 70 uint8x8_t in2 = vget_low_u8(vextq_u8(in, in, 2)); 71 uint8x8_t in3 = vget_low_u8(vextq_u8(in, in, 3)); 72 uint8x8_t in4 = vget_low_u8(vextq_u8(in, in, 4)); 73 uint8x8_t in5 = vget_low_u8(vextq_u8(in, in, 5)); 74 uint8x8_t in6 = vget_low_u8(vextq_u8(in, in, 6)); 75 uint8x8_t in7 = vget_low_u8(vextq_u8(in, in, 7)); 76 77 int32x4_t m01 = vusdotq_s32(add_const, vcombine_u8(in0, in1), f01_u8); 78 int32x4_t m23 = vusdotq_s32(add_const, vcombine_u8(in2, in3), f23_u8); 79 int32x4_t m45 = vusdotq_s32(add_const, vcombine_u8(in4, in5), f45_u8); 80 int32x4_t m67 = vusdotq_s32(add_const, vcombine_u8(in6, in7), f67_u8); 81 82 int32x4_t m0123 = vpaddq_s32(m01, m23); 83 int32x4_t m4567 = vpaddq_s32(m45, m67); 84 85 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(m0123, ROUND0_BITS), 86 vqrshrun_n_s32(m4567, ROUND0_BITS)); 87 return vreinterpretq_s16_u16(res); 88 } 89 90 static AOM_FORCE_INLINE int16x8_t 91 horizontal_filter_4x1_f1_beta0(const uint8x16_t in, int16x8_t f_s16) { 92 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); 93 94 int8x16_t f_s8 = vcombine_s8(vmovn_s16(f_s16), vmovn_s16(f_s16)); 95 96 uint8x16_t perm0 = vld1q_u8(&usdot_permute_idx[0]); 97 uint8x16_t perm1 = vld1q_u8(&usdot_permute_idx[16]); 98 99 // Permute samples ready for dot product. 100 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } 101 // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } 102 uint8x16_t in_0123 = vqtbl1q_u8(in, perm0); 103 uint8x16_t in_4567 = vqtbl1q_u8(in, perm1); 104 105 int32x4_t m0123 = vusdotq_laneq_s32(add_const, in_0123, f_s8, 0); 106 m0123 = vusdotq_laneq_s32(m0123, in_4567, f_s8, 1); 107 108 uint16x8_t res = 109 vcombine_u16(vqrshrun_n_s32(m0123, ROUND0_BITS), vdup_n_u16(0)); 110 return vreinterpretq_s16_u16(res); 111 } 112 113 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in, 114 int sx) { 115 int16x8_t f_s16 = vld1q_s16(av1_warped_filter[sx >> WARPEDDIFF_PREC_BITS]); 116 return horizontal_filter_4x1_f1_beta0(in, f_s16); 117 } 118 119 static AOM_FORCE_INLINE int16x8_t 120 horizontal_filter_8x1_f1_beta0(const uint8x16_t in, int16x8_t f_s16) { 121 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1)); 122 123 int8x16_t f_s8 = vcombine_s8(vmovn_s16(f_s16), vmovn_s16(f_s16)); 124 125 uint8x16_t perm0 = vld1q_u8(&usdot_permute_idx[0]); 126 uint8x16_t perm1 = vld1q_u8(&usdot_permute_idx[16]); 127 uint8x16_t perm2 = vld1q_u8(&usdot_permute_idx[32]); 128 129 // Permute samples ready for dot product. 130 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } 131 // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } 132 // { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 } 133 uint8x16_t in_0123 = vqtbl1q_u8(in, perm0); 134 uint8x16_t in_4567 = vqtbl1q_u8(in, perm1); 135 uint8x16_t in_89ab = vqtbl1q_u8(in, perm2); 136 137 int32x4_t m0123 = vusdotq_laneq_s32(add_const, in_0123, f_s8, 0); 138 m0123 = vusdotq_laneq_s32(m0123, in_4567, f_s8, 1); 139 140 int32x4_t m4567 = vusdotq_laneq_s32(add_const, in_4567, f_s8, 0); 141 m4567 = vusdotq_laneq_s32(m4567, in_89ab, f_s8, 1); 142 143 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(m0123, ROUND0_BITS), 144 vqrshrun_n_s32(m4567, ROUND0_BITS)); 145 return vreinterpretq_s16_u16(res); 146 } 147 148 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in, 149 int sx) { 150 int16x8_t f_s16 = vld1q_s16(av1_warped_filter[sx >> WARPEDDIFF_PREC_BITS]); 151 return horizontal_filter_8x1_f1_beta0(in, f_s16); 152 } 153 154 static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src, 155 int32x4_t *res, int sy) { 156 int16x4_t s0 = vget_low_s16(src[0]); 157 int16x4_t s1 = vget_low_s16(src[1]); 158 int16x4_t s2 = vget_low_s16(src[2]); 159 int16x4_t s3 = vget_low_s16(src[3]); 160 int16x4_t s4 = vget_low_s16(src[4]); 161 int16x4_t s5 = vget_low_s16(src[5]); 162 int16x4_t s6 = vget_low_s16(src[6]); 163 int16x4_t s7 = vget_low_s16(src[7]); 164 165 int16x8_t f = vld1q_s16(av1_warped_filter[sy >> WARPEDDIFF_PREC_BITS]); 166 167 int32x4_t m0123 = vmull_lane_s16(s0, vget_low_s16(f), 0); 168 m0123 = vmlal_lane_s16(m0123, s1, vget_low_s16(f), 1); 169 m0123 = vmlal_lane_s16(m0123, s2, vget_low_s16(f), 2); 170 m0123 = vmlal_lane_s16(m0123, s3, vget_low_s16(f), 3); 171 m0123 = vmlal_lane_s16(m0123, s4, vget_high_s16(f), 0); 172 m0123 = vmlal_lane_s16(m0123, s5, vget_high_s16(f), 1); 173 m0123 = vmlal_lane_s16(m0123, s6, vget_high_s16(f), 2); 174 m0123 = vmlal_lane_s16(m0123, s7, vget_high_s16(f), 3); 175 176 *res = m0123; 177 } 178 179 static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src, 180 int32x4_t *res, int sy, 181 int gamma) { 182 int16x8_t s0, s1, s2, s3; 183 transpose_elems_s16_4x8( 184 vget_low_s16(src[0]), vget_low_s16(src[1]), vget_low_s16(src[2]), 185 vget_low_s16(src[3]), vget_low_s16(src[4]), vget_low_s16(src[5]), 186 vget_low_s16(src[6]), vget_low_s16(src[7]), &s0, &s1, &s2, &s3); 187 188 int16x8_t f[4]; 189 load_filters_4(f, sy, gamma); 190 191 int64x2_t m0 = aom_sdotq_s16(vdupq_n_s64(0), s0, f[0]); 192 int64x2_t m1 = aom_sdotq_s16(vdupq_n_s64(0), s1, f[1]); 193 int64x2_t m2 = aom_sdotq_s16(vdupq_n_s64(0), s2, f[2]); 194 int64x2_t m3 = aom_sdotq_s16(vdupq_n_s64(0), s3, f[3]); 195 196 int64x2_t m01 = vpaddq_s64(m0, m1); 197 int64x2_t m23 = vpaddq_s64(m2, m3); 198 199 *res = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); 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 int64x2_t m0 = aom_sdotq_s16(vdupq_n_s64(0), s0, f[0]); 257 int64x2_t m1 = aom_sdotq_s16(vdupq_n_s64(0), s1, f[1]); 258 int64x2_t m2 = aom_sdotq_s16(vdupq_n_s64(0), s2, f[2]); 259 int64x2_t m3 = aom_sdotq_s16(vdupq_n_s64(0), s3, f[3]); 260 int64x2_t m4 = aom_sdotq_s16(vdupq_n_s64(0), s4, f[4]); 261 int64x2_t m5 = aom_sdotq_s16(vdupq_n_s64(0), s5, f[5]); 262 int64x2_t m6 = aom_sdotq_s16(vdupq_n_s64(0), s6, f[6]); 263 int64x2_t m7 = aom_sdotq_s16(vdupq_n_s64(0), s7, f[7]); 264 265 int64x2_t m01 = vpaddq_s64(m0, m1); 266 int64x2_t m23 = vpaddq_s64(m2, m3); 267 int64x2_t m45 = vpaddq_s64(m4, m5); 268 int64x2_t m67 = vpaddq_s64(m6, m7); 269 270 *res_low = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); 271 *res_high = vcombine_s32(vmovn_s64(m45), vmovn_s64(m67)); 272 } 273 274 void av1_warp_affine_sve(const int32_t *mat, const uint8_t *ref, int width, 275 int height, int stride, uint8_t *pred, int p_col, 276 int p_row, int p_width, int p_height, int p_stride, 277 int subsampling_x, int subsampling_y, 278 ConvolveParams *conv_params, int16_t alpha, 279 int16_t beta, int16_t gamma, int16_t delta) { 280 av1_warp_affine_common(mat, ref, width, height, stride, pred, p_col, p_row, 281 p_width, p_height, p_stride, subsampling_x, 282 subsampling_y, conv_params, alpha, beta, gamma, delta); 283 }