highbd_warp_plane_sve.c (10768B)
1 /* 2 * Copyright (c) 2024, 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 #include <assert.h> 14 #include <stdbool.h> 15 #include <arm_neon_sve_bridge.h> 16 17 #include "aom_dsp/aom_dsp_common.h" 18 #include "aom_dsp/arm/aom_neon_sve_bridge.h" 19 #include "aom_dsp/arm/mem_neon.h" 20 #include "aom_dsp/arm/transpose_neon.h" 21 #include "aom_ports/mem.h" 22 #include "av1/common/scale.h" 23 #include "av1/common/warped_motion.h" 24 #include "config/av1_rtcd.h" 25 #include "highbd_warp_plane_neon.h" 26 27 static AOM_FORCE_INLINE int16x8_t 28 highbd_horizontal_filter_4x1_f4(int16x8_t rv0, int16x8_t rv1, int16x8_t rv2, 29 int16x8_t rv3, int bd, int sx, int alpha) { 30 int16x8_t f[4]; 31 load_filters_4(f, sx, alpha); 32 33 int64x2_t m0 = aom_sdotq_s16(vdupq_n_s64(0), rv0, f[0]); 34 int64x2_t m1 = aom_sdotq_s16(vdupq_n_s64(0), rv1, f[1]); 35 int64x2_t m2 = aom_sdotq_s16(vdupq_n_s64(0), rv2, f[2]); 36 int64x2_t m3 = aom_sdotq_s16(vdupq_n_s64(0), rv3, f[3]); 37 38 int64x2_t m01 = vpaddq_s64(m0, m1); 39 int64x2_t m23 = vpaddq_s64(m2, m3); 40 41 const int round0 = bd == 12 ? ROUND0_BITS + 2 : ROUND0_BITS; 42 const int offset_bits_horiz = bd + FILTER_BITS - 1; 43 44 int32x4_t res = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); 45 res = vaddq_s32(res, vdupq_n_s32(1 << offset_bits_horiz)); 46 res = vrshlq_s32(res, vdupq_n_s32(-round0)); 47 return vcombine_s16(vmovn_s32(res), vdup_n_s16(0)); 48 } 49 50 static AOM_FORCE_INLINE int16x8_t highbd_horizontal_filter_8x1_f8( 51 int16x8_t rv0, int16x8_t rv1, int16x8_t rv2, int16x8_t rv3, int16x8_t rv4, 52 int16x8_t rv5, int16x8_t rv6, int16x8_t rv7, int bd, int sx, int alpha) { 53 int16x8_t f[8]; 54 load_filters_8(f, sx, alpha); 55 56 int64x2_t m0 = aom_sdotq_s16(vdupq_n_s64(0), rv0, f[0]); 57 int64x2_t m1 = aom_sdotq_s16(vdupq_n_s64(0), rv1, f[1]); 58 int64x2_t m2 = aom_sdotq_s16(vdupq_n_s64(0), rv2, f[2]); 59 int64x2_t m3 = aom_sdotq_s16(vdupq_n_s64(0), rv3, f[3]); 60 int64x2_t m4 = aom_sdotq_s16(vdupq_n_s64(0), rv4, f[4]); 61 int64x2_t m5 = aom_sdotq_s16(vdupq_n_s64(0), rv5, f[5]); 62 int64x2_t m6 = aom_sdotq_s16(vdupq_n_s64(0), rv6, f[6]); 63 int64x2_t m7 = aom_sdotq_s16(vdupq_n_s64(0), rv7, f[7]); 64 65 int64x2_t m01 = vpaddq_s64(m0, m1); 66 int64x2_t m23 = vpaddq_s64(m2, m3); 67 int64x2_t m45 = vpaddq_s64(m4, m5); 68 int64x2_t m67 = vpaddq_s64(m6, m7); 69 70 const int round0 = bd == 12 ? ROUND0_BITS + 2 : ROUND0_BITS; 71 const int offset_bits_horiz = bd + FILTER_BITS - 1; 72 73 int32x4_t res0 = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); 74 int32x4_t res1 = vcombine_s32(vmovn_s64(m45), vmovn_s64(m67)); 75 res0 = vaddq_s32(res0, vdupq_n_s32(1 << offset_bits_horiz)); 76 res1 = vaddq_s32(res1, vdupq_n_s32(1 << offset_bits_horiz)); 77 res0 = vrshlq_s32(res0, vdupq_n_s32(-round0)); 78 res1 = vrshlq_s32(res1, vdupq_n_s32(-round0)); 79 return vcombine_s16(vmovn_s32(res0), vmovn_s32(res1)); 80 } 81 82 static AOM_FORCE_INLINE int16x8_t 83 highbd_horizontal_filter_4x1_f1(int16x8_t rv0, int16x8_t rv1, int16x8_t rv2, 84 int16x8_t rv3, int bd, int sx) { 85 int16x8_t f = load_filters_1(sx); 86 87 int64x2_t m0 = aom_sdotq_s16(vdupq_n_s64(0), rv0, f); 88 int64x2_t m1 = aom_sdotq_s16(vdupq_n_s64(0), rv1, f); 89 int64x2_t m2 = aom_sdotq_s16(vdupq_n_s64(0), rv2, f); 90 int64x2_t m3 = aom_sdotq_s16(vdupq_n_s64(0), rv3, f); 91 92 int64x2_t m01 = vpaddq_s64(m0, m1); 93 int64x2_t m23 = vpaddq_s64(m2, m3); 94 95 const int round0 = bd == 12 ? ROUND0_BITS + 2 : ROUND0_BITS; 96 const int offset_bits_horiz = bd + FILTER_BITS - 1; 97 98 int32x4_t res = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); 99 res = vaddq_s32(res, vdupq_n_s32(1 << offset_bits_horiz)); 100 res = vrshlq_s32(res, vdupq_n_s32(-round0)); 101 return vcombine_s16(vmovn_s32(res), vdup_n_s16(0)); 102 } 103 104 static AOM_FORCE_INLINE int16x8_t highbd_horizontal_filter_8x1_f1( 105 int16x8_t rv0, int16x8_t rv1, int16x8_t rv2, int16x8_t rv3, int16x8_t rv4, 106 int16x8_t rv5, int16x8_t rv6, int16x8_t rv7, int bd, int sx) { 107 int16x8_t f = load_filters_1(sx); 108 109 int64x2_t m0 = aom_sdotq_s16(vdupq_n_s64(0), rv0, f); 110 int64x2_t m1 = aom_sdotq_s16(vdupq_n_s64(0), rv1, f); 111 int64x2_t m2 = aom_sdotq_s16(vdupq_n_s64(0), rv2, f); 112 int64x2_t m3 = aom_sdotq_s16(vdupq_n_s64(0), rv3, f); 113 int64x2_t m4 = aom_sdotq_s16(vdupq_n_s64(0), rv4, f); 114 int64x2_t m5 = aom_sdotq_s16(vdupq_n_s64(0), rv5, f); 115 int64x2_t m6 = aom_sdotq_s16(vdupq_n_s64(0), rv6, f); 116 int64x2_t m7 = aom_sdotq_s16(vdupq_n_s64(0), rv7, f); 117 118 int64x2_t m01 = vpaddq_s64(m0, m1); 119 int64x2_t m23 = vpaddq_s64(m2, m3); 120 int64x2_t m45 = vpaddq_s64(m4, m5); 121 int64x2_t m67 = vpaddq_s64(m6, m7); 122 123 const int round0 = bd == 12 ? ROUND0_BITS + 2 : ROUND0_BITS; 124 const int offset_bits_horiz = bd + FILTER_BITS - 1; 125 126 int32x4_t res0 = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); 127 int32x4_t res1 = vcombine_s32(vmovn_s64(m45), vmovn_s64(m67)); 128 res0 = vaddq_s32(res0, vdupq_n_s32(1 << offset_bits_horiz)); 129 res1 = vaddq_s32(res1, vdupq_n_s32(1 << offset_bits_horiz)); 130 res0 = vrshlq_s32(res0, vdupq_n_s32(-round0)); 131 res1 = vrshlq_s32(res1, vdupq_n_s32(-round0)); 132 return vcombine_s16(vmovn_s32(res0), vmovn_s32(res1)); 133 } 134 135 static AOM_FORCE_INLINE int32x4_t vertical_filter_4x1_f1(const int16x8_t *tmp, 136 int sy) { 137 const int16x8_t f = load_filters_1(sy); 138 const int16x4_t f0123 = vget_low_s16(f); 139 const int16x4_t f4567 = vget_high_s16(f); 140 141 // No benefit to using SDOT here, the cost of rearrangement is too high. 142 int32x4_t m0123 = vmull_lane_s16(vget_low_s16(tmp[0]), f0123, 0); 143 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[1]), f0123, 1); 144 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[2]), f0123, 2); 145 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[3]), f0123, 3); 146 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[4]), f4567, 0); 147 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[5]), f4567, 1); 148 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[6]), f4567, 2); 149 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[7]), f4567, 3); 150 return m0123; 151 } 152 153 static AOM_FORCE_INLINE int32x4x2_t vertical_filter_8x1_f1(const int16x8_t *tmp, 154 int sy) { 155 const int16x8_t f = load_filters_1(sy); 156 const int16x4_t f0123 = vget_low_s16(f); 157 const int16x4_t f4567 = vget_high_s16(f); 158 159 // No benefit to using SDOT here, the cost of rearrangement is too high. 160 int32x4_t m0123 = vmull_lane_s16(vget_low_s16(tmp[0]), f0123, 0); 161 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[1]), f0123, 1); 162 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[2]), f0123, 2); 163 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[3]), f0123, 3); 164 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[4]), f4567, 0); 165 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[5]), f4567, 1); 166 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[6]), f4567, 2); 167 m0123 = vmlal_lane_s16(m0123, vget_low_s16(tmp[7]), f4567, 3); 168 169 int32x4_t m4567 = vmull_lane_s16(vget_high_s16(tmp[0]), f0123, 0); 170 m4567 = vmlal_lane_s16(m4567, vget_high_s16(tmp[1]), f0123, 1); 171 m4567 = vmlal_lane_s16(m4567, vget_high_s16(tmp[2]), f0123, 2); 172 m4567 = vmlal_lane_s16(m4567, vget_high_s16(tmp[3]), f0123, 3); 173 m4567 = vmlal_lane_s16(m4567, vget_high_s16(tmp[4]), f4567, 0); 174 m4567 = vmlal_lane_s16(m4567, vget_high_s16(tmp[5]), f4567, 1); 175 m4567 = vmlal_lane_s16(m4567, vget_high_s16(tmp[6]), f4567, 2); 176 m4567 = vmlal_lane_s16(m4567, vget_high_s16(tmp[7]), f4567, 3); 177 return (int32x4x2_t){ { m0123, m4567 } }; 178 } 179 180 static AOM_FORCE_INLINE int32x4_t vertical_filter_4x1_f4(const int16x8_t *tmp, 181 int sy, int gamma) { 182 int16x8_t s0, s1, s2, s3; 183 transpose_elems_s16_4x8( 184 vget_low_s16(tmp[0]), vget_low_s16(tmp[1]), vget_low_s16(tmp[2]), 185 vget_low_s16(tmp[3]), vget_low_s16(tmp[4]), vget_low_s16(tmp[5]), 186 vget_low_s16(tmp[6]), vget_low_s16(tmp[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 return vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); 199 } 200 201 static AOM_FORCE_INLINE int32x4x2_t vertical_filter_8x1_f8(const int16x8_t *tmp, 202 int sy, int gamma) { 203 int16x8_t s0 = tmp[0]; 204 int16x8_t s1 = tmp[1]; 205 int16x8_t s2 = tmp[2]; 206 int16x8_t s3 = tmp[3]; 207 int16x8_t s4 = tmp[4]; 208 int16x8_t s5 = tmp[5]; 209 int16x8_t s6 = tmp[6]; 210 int16x8_t s7 = tmp[7]; 211 transpose_elems_inplace_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7); 212 213 int16x8_t f[8]; 214 load_filters_8(f, sy, gamma); 215 216 int64x2_t m0 = aom_sdotq_s16(vdupq_n_s64(0), s0, f[0]); 217 int64x2_t m1 = aom_sdotq_s16(vdupq_n_s64(0), s1, f[1]); 218 int64x2_t m2 = aom_sdotq_s16(vdupq_n_s64(0), s2, f[2]); 219 int64x2_t m3 = aom_sdotq_s16(vdupq_n_s64(0), s3, f[3]); 220 int64x2_t m4 = aom_sdotq_s16(vdupq_n_s64(0), s4, f[4]); 221 int64x2_t m5 = aom_sdotq_s16(vdupq_n_s64(0), s5, f[5]); 222 int64x2_t m6 = aom_sdotq_s16(vdupq_n_s64(0), s6, f[6]); 223 int64x2_t m7 = aom_sdotq_s16(vdupq_n_s64(0), s7, f[7]); 224 225 int64x2_t m01 = vpaddq_s64(m0, m1); 226 int64x2_t m23 = vpaddq_s64(m2, m3); 227 int64x2_t m45 = vpaddq_s64(m4, m5); 228 int64x2_t m67 = vpaddq_s64(m6, m7); 229 230 int32x4x2_t ret; 231 ret.val[0] = vcombine_s32(vmovn_s64(m01), vmovn_s64(m23)); 232 ret.val[1] = vcombine_s32(vmovn_s64(m45), vmovn_s64(m67)); 233 return ret; 234 } 235 236 void av1_highbd_warp_affine_sve(const int32_t *mat, const uint16_t *ref, 237 int width, int height, int stride, 238 uint16_t *pred, int p_col, int p_row, 239 int p_width, int p_height, int p_stride, 240 int subsampling_x, int subsampling_y, int bd, 241 ConvolveParams *conv_params, int16_t alpha, 242 int16_t beta, int16_t gamma, int16_t delta) { 243 highbd_warp_affine_common(mat, ref, width, height, stride, pred, p_col, p_row, 244 p_width, p_height, p_stride, subsampling_x, 245 subsampling_y, bd, conv_params, alpha, beta, gamma, 246 delta); 247 }