• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #ifndef AOM_AV1_COMMON_ARM_WARP_PLANE_NEON_H_
12 #define AOM_AV1_COMMON_ARM_WARP_PLANE_NEON_H_
13 
14 #include <assert.h>
15 #include <arm_neon.h>
16 #include <memory.h>
17 #include <math.h>
18 
19 #include "aom_dsp/aom_dsp_common.h"
20 #include "aom_dsp/arm/sum_neon.h"
21 #include "aom_dsp/arm/transpose_neon.h"
22 #include "aom_ports/mem.h"
23 #include "config/av1_rtcd.h"
24 #include "av1/common/warped_motion.h"
25 #include "av1/common/scale.h"
26 
27 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in,
28                                                            int sx, int alpha);
29 
30 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in,
31                                                            int sx, int alpha);
32 
33 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in,
34                                                            int sx);
35 
36 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in,
37                                                            int sx);
38 
39 static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src,
40                                                     int32x4_t *res, int sy);
41 
42 static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src,
43                                                     int32x4_t *res, int sy,
44                                                     int gamma);
45 
46 static AOM_FORCE_INLINE void vertical_filter_8x1_f1(const int16x8_t *src,
47                                                     int32x4_t *res_low,
48                                                     int32x4_t *res_high,
49                                                     int sy);
50 
51 static AOM_FORCE_INLINE void vertical_filter_8x1_f8(const int16x8_t *src,
52                                                     int32x4_t *res_low,
53                                                     int32x4_t *res_high, int sy,
54                                                     int gamma);
55 
load_filters_4(int16x8_t out[],int offset,int stride)56 static AOM_FORCE_INLINE void load_filters_4(int16x8_t out[], int offset,
57                                             int stride) {
58   out[0] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 0 * stride) >>
59                                                       WARPEDDIFF_PREC_BITS)));
60   out[1] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 1 * stride) >>
61                                                       WARPEDDIFF_PREC_BITS)));
62   out[2] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 2 * stride) >>
63                                                       WARPEDDIFF_PREC_BITS)));
64   out[3] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 3 * stride) >>
65                                                       WARPEDDIFF_PREC_BITS)));
66 }
67 
load_filters_8(int16x8_t out[],int offset,int stride)68 static AOM_FORCE_INLINE void load_filters_8(int16x8_t out[], int offset,
69                                             int stride) {
70   out[0] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 0 * stride) >>
71                                                       WARPEDDIFF_PREC_BITS)));
72   out[1] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 1 * stride) >>
73                                                       WARPEDDIFF_PREC_BITS)));
74   out[2] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 2 * stride) >>
75                                                       WARPEDDIFF_PREC_BITS)));
76   out[3] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 3 * stride) >>
77                                                       WARPEDDIFF_PREC_BITS)));
78   out[4] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 4 * stride) >>
79                                                       WARPEDDIFF_PREC_BITS)));
80   out[5] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 5 * stride) >>
81                                                       WARPEDDIFF_PREC_BITS)));
82   out[6] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 6 * stride) >>
83                                                       WARPEDDIFF_PREC_BITS)));
84   out[7] = vld1q_s16((int16_t *)(av1_warped_filter + ((offset + 7 * stride) >>
85                                                       WARPEDDIFF_PREC_BITS)));
86 }
87 
clamp_iy(int iy,int height)88 static AOM_FORCE_INLINE int clamp_iy(int iy, int height) {
89   return clamp(iy, 0, height - 1);
90 }
91 
warp_affine_horizontal(const uint8_t * ref,int width,int height,int stride,int p_width,int p_height,int16_t alpha,int16_t beta,const int64_t x4,const int64_t y4,const int i,int16x8_t tmp[])92 static AOM_FORCE_INLINE void warp_affine_horizontal(
93     const uint8_t *ref, int width, int height, int stride, int p_width,
94     int p_height, int16_t alpha, int16_t beta, const int64_t x4,
95     const int64_t y4, const int i, int16x8_t tmp[]) {
96   const int bd = 8;
97   const int reduce_bits_horiz = ROUND0_BITS;
98   const int height_limit = AOMMIN(8, p_height - i) + 7;
99 
100   int32_t ix4 = (int32_t)(x4 >> WARPEDMODEL_PREC_BITS);
101   int32_t iy4 = (int32_t)(y4 >> WARPEDMODEL_PREC_BITS);
102 
103   int32_t sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
104   sx4 += alpha * (-4) + beta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
105          (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
106   sx4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
107 
108   if (ix4 <= -7) {
109     for (int k = 0; k < height_limit; ++k) {
110       int iy = clamp_iy(iy4 + k - 7, height);
111       int16_t dup_val =
112           (1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) +
113           ref[iy * stride] * (1 << (FILTER_BITS - reduce_bits_horiz));
114       tmp[k] = vdupq_n_s16(dup_val);
115     }
116     return;
117   } else if (ix4 >= width + 6) {
118     for (int k = 0; k < height_limit; ++k) {
119       int iy = clamp_iy(iy4 + k - 7, height);
120       int16_t dup_val = (1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) +
121                         ref[iy * stride + (width - 1)] *
122                             (1 << (FILTER_BITS - reduce_bits_horiz));
123       tmp[k] = vdupq_n_s16(dup_val);
124     }
125     return;
126   }
127 
128   static const uint8_t kIotaArr[] = { 0, 1, 2,  3,  4,  5,  6,  7,
129                                       8, 9, 10, 11, 12, 13, 14, 15 };
130   const uint8x16_t indx = vld1q_u8(kIotaArr);
131 
132   const int out_of_boundary_left = -(ix4 - 6);
133   const int out_of_boundary_right = (ix4 + 8) - width;
134 
135 #define APPLY_HORIZONTAL_SHIFT(fn, ...)                                \
136   do {                                                                 \
137     if (out_of_boundary_left >= 0 || out_of_boundary_right >= 0) {     \
138       for (int k = 0; k < height_limit; ++k) {                         \
139         const int iy = clamp_iy(iy4 + k - 7, height);                  \
140         const uint8_t *src = ref + iy * stride + ix4 - 7;              \
141         uint8x16_t src_1 = vld1q_u8(src);                              \
142                                                                        \
143         if (out_of_boundary_left >= 0) {                               \
144           int limit = out_of_boundary_left + 1;                        \
145           uint8x16_t cmp_vec = vdupq_n_u8(out_of_boundary_left);       \
146           uint8x16_t vec_dup = vdupq_n_u8(*(src + limit));             \
147           uint8x16_t mask_val = vcleq_u8(indx, cmp_vec);               \
148           src_1 = vbslq_u8(mask_val, vec_dup, src_1);                  \
149         }                                                              \
150         if (out_of_boundary_right >= 0) {                              \
151           int limit = 15 - (out_of_boundary_right + 1);                \
152           uint8x16_t cmp_vec = vdupq_n_u8(15 - out_of_boundary_right); \
153           uint8x16_t vec_dup = vdupq_n_u8(*(src + limit));             \
154           uint8x16_t mask_val = vcgeq_u8(indx, cmp_vec);               \
155           src_1 = vbslq_u8(mask_val, vec_dup, src_1);                  \
156         }                                                              \
157         tmp[k] = (fn)(src_1, __VA_ARGS__);                             \
158       }                                                                \
159     } else {                                                           \
160       for (int k = 0; k < height_limit; ++k) {                         \
161         const int iy = clamp_iy(iy4 + k - 7, height);                  \
162         const uint8_t *src = ref + iy * stride + ix4 - 7;              \
163         uint8x16_t src_1 = vld1q_u8(src);                              \
164         tmp[k] = (fn)(src_1, __VA_ARGS__);                             \
165       }                                                                \
166     }                                                                  \
167   } while (0)
168 
169   if (p_width == 4) {
170     if (beta == 0) {
171       if (alpha == 0) {
172         APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f1, sx4);
173       } else {
174         APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f4, sx4, alpha);
175       }
176     } else {
177       if (alpha == 0) {
178         APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f1,
179                                (sx4 + beta * (k - 3)));
180       } else {
181         APPLY_HORIZONTAL_SHIFT(horizontal_filter_4x1_f4, (sx4 + beta * (k - 3)),
182                                alpha);
183       }
184     }
185   } else {
186     if (beta == 0) {
187       if (alpha == 0) {
188         APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f1, sx4);
189       } else {
190         APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f8, sx4, alpha);
191       }
192     } else {
193       if (alpha == 0) {
194         APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f1,
195                                (sx4 + beta * (k - 3)));
196       } else {
197         APPLY_HORIZONTAL_SHIFT(horizontal_filter_8x1_f8, (sx4 + beta * (k - 3)),
198                                alpha);
199       }
200     }
201   }
202 }
203 
warp_affine_vertical(uint8_t * pred,int p_width,int p_height,int p_stride,int is_compound,uint16_t * dst,int dst_stride,int do_average,int use_dist_wtd_comp_avg,int16_t gamma,int16_t delta,const int64_t y4,const int i,const int j,int16x8_t tmp[],const int fwd,const int bwd)204 static AOM_FORCE_INLINE void warp_affine_vertical(
205     uint8_t *pred, int p_width, int p_height, int p_stride, int is_compound,
206     uint16_t *dst, int dst_stride, int do_average, int use_dist_wtd_comp_avg,
207     int16_t gamma, int16_t delta, const int64_t y4, const int i, const int j,
208     int16x8_t tmp[], const int fwd, const int bwd) {
209   const int bd = 8;
210   const int reduce_bits_horiz = ROUND0_BITS;
211   const int offset_bits_vert = bd + 2 * FILTER_BITS - reduce_bits_horiz;
212   int add_const_vert;
213   if (is_compound) {
214     add_const_vert =
215         (1 << offset_bits_vert) + (1 << (COMPOUND_ROUND1_BITS - 1));
216   } else {
217     add_const_vert =
218         (1 << offset_bits_vert) + (1 << (2 * FILTER_BITS - ROUND0_BITS - 1));
219   }
220   const int sub_constant = (1 << (bd - 1)) + (1 << bd);
221 
222   const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS;
223   const int res_sub_const =
224       (1 << (2 * FILTER_BITS - ROUND0_BITS - COMPOUND_ROUND1_BITS - 1)) -
225       (1 << (offset_bits - COMPOUND_ROUND1_BITS)) -
226       (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1));
227 
228   int32_t sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
229   sy4 += gamma * (-4) + delta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
230          (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
231   sy4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
232 
233   if (p_width > 4) {
234     for (int k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
235       int sy = sy4 + delta * (k + 4);
236       const int16x8_t *v_src = tmp + (k + 4);
237 
238       int32x4_t res_lo, res_hi;
239       if (gamma == 0) {
240         vertical_filter_8x1_f1(v_src, &res_lo, &res_hi, sy);
241       } else {
242         vertical_filter_8x1_f8(v_src, &res_lo, &res_hi, sy, gamma);
243       }
244 
245       res_lo = vaddq_s32(res_lo, vdupq_n_s32(add_const_vert));
246       res_hi = vaddq_s32(res_hi, vdupq_n_s32(add_const_vert));
247 
248       if (is_compound) {
249         uint16_t *const p = (uint16_t *)&dst[(i + k + 4) * dst_stride + j];
250         int16x8_t res_s16 =
251             vcombine_s16(vshrn_n_s32(res_lo, COMPOUND_ROUND1_BITS),
252                          vshrn_n_s32(res_hi, COMPOUND_ROUND1_BITS));
253         if (do_average) {
254           int16x8_t tmp16 = vreinterpretq_s16_u16(vld1q_u16(p));
255           if (use_dist_wtd_comp_avg) {
256             int32x4_t tmp32_lo = vmull_n_s16(vget_low_s16(tmp16), fwd);
257             int32x4_t tmp32_hi = vmull_n_s16(vget_high_s16(tmp16), fwd);
258             tmp32_lo = vmlal_n_s16(tmp32_lo, vget_low_s16(res_s16), bwd);
259             tmp32_hi = vmlal_n_s16(tmp32_hi, vget_high_s16(res_s16), bwd);
260             tmp16 = vcombine_s16(vshrn_n_s32(tmp32_lo, DIST_PRECISION_BITS),
261                                  vshrn_n_s32(tmp32_hi, DIST_PRECISION_BITS));
262           } else {
263             tmp16 = vhaddq_s16(tmp16, res_s16);
264           }
265           int16x8_t res = vaddq_s16(tmp16, vdupq_n_s16(res_sub_const));
266           uint8x8_t res8 = vqshrun_n_s16(
267               res, 2 * FILTER_BITS - ROUND0_BITS - COMPOUND_ROUND1_BITS);
268           vst1_u8(&pred[(i + k + 4) * p_stride + j], res8);
269         } else {
270           vst1q_u16(p, vreinterpretq_u16_s16(res_s16));
271         }
272       } else {
273         int16x8_t res16 =
274             vcombine_s16(vshrn_n_s32(res_lo, 2 * FILTER_BITS - ROUND0_BITS),
275                          vshrn_n_s32(res_hi, 2 * FILTER_BITS - ROUND0_BITS));
276         res16 = vsubq_s16(res16, vdupq_n_s16(sub_constant));
277 
278         uint8_t *const p = (uint8_t *)&pred[(i + k + 4) * p_stride + j];
279         vst1_u8(p, vqmovun_s16(res16));
280       }
281     }
282   } else {
283     // p_width == 4
284     for (int k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
285       int sy = sy4 + delta * (k + 4);
286       const int16x8_t *v_src = tmp + (k + 4);
287 
288       int32x4_t res_lo;
289       if (gamma == 0) {
290         vertical_filter_4x1_f1(v_src, &res_lo, sy);
291       } else {
292         vertical_filter_4x1_f4(v_src, &res_lo, sy, gamma);
293       }
294 
295       res_lo = vaddq_s32(res_lo, vdupq_n_s32(add_const_vert));
296 
297       if (is_compound) {
298         uint16_t *const p = (uint16_t *)&dst[(i + k + 4) * dst_stride + j];
299 
300         int16x4_t res_lo_s16 = vshrn_n_s32(res_lo, COMPOUND_ROUND1_BITS);
301         if (do_average) {
302           uint8_t *const dst8 = &pred[(i + k + 4) * p_stride + j];
303           int16x4_t tmp16_lo = vreinterpret_s16_u16(vld1_u16(p));
304           if (use_dist_wtd_comp_avg) {
305             int32x4_t tmp32_lo = vmull_n_s16(tmp16_lo, fwd);
306             tmp32_lo = vmlal_n_s16(tmp32_lo, res_lo_s16, bwd);
307             tmp16_lo = vshrn_n_s32(tmp32_lo, DIST_PRECISION_BITS);
308           } else {
309             tmp16_lo = vhadd_s16(tmp16_lo, res_lo_s16);
310           }
311           int16x4_t res = vadd_s16(tmp16_lo, vdup_n_s16(res_sub_const));
312           uint8x8_t res8 = vqshrun_n_s16(
313               vcombine_s16(res, vdup_n_s16(0)),
314               2 * FILTER_BITS - ROUND0_BITS - COMPOUND_ROUND1_BITS);
315           vst1_lane_u32((uint32_t *)dst8, vreinterpret_u32_u8(res8), 0);
316         } else {
317           uint16x4_t res_u16_low = vreinterpret_u16_s16(res_lo_s16);
318           vst1_u16(p, res_u16_low);
319         }
320       } else {
321         int16x4_t res16 = vshrn_n_s32(res_lo, 2 * FILTER_BITS - ROUND0_BITS);
322         res16 = vsub_s16(res16, vdup_n_s16(sub_constant));
323 
324         uint8_t *const p = (uint8_t *)&pred[(i + k + 4) * p_stride + j];
325         uint8x8_t val = vqmovun_s16(vcombine_s16(res16, vdup_n_s16(0)));
326         vst1_lane_u32((uint32_t *)p, vreinterpret_u32_u8(val), 0);
327       }
328     }
329   }
330 }
331 
av1_warp_affine_common(const int32_t * mat,const uint8_t * ref,int width,int height,int stride,uint8_t * pred,int p_col,int p_row,int p_width,int p_height,int p_stride,int subsampling_x,int subsampling_y,ConvolveParams * conv_params,int16_t alpha,int16_t beta,int16_t gamma,int16_t delta)332 static AOM_FORCE_INLINE void av1_warp_affine_common(
333     const int32_t *mat, const uint8_t *ref, int width, int height, int stride,
334     uint8_t *pred, int p_col, int p_row, int p_width, int p_height,
335     int p_stride, int subsampling_x, int subsampling_y,
336     ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma,
337     int16_t delta) {
338   const int w0 = conv_params->fwd_offset;
339   const int w1 = conv_params->bck_offset;
340   const int is_compound = conv_params->is_compound;
341   uint16_t *const dst = conv_params->dst;
342   const int dst_stride = conv_params->dst_stride;
343   const int do_average = conv_params->do_average;
344   const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg;
345 
346   assert(IMPLIES(is_compound, dst != NULL));
347   assert(IMPLIES(do_average, is_compound));
348 
349   for (int i = 0; i < p_height; i += 8) {
350     for (int j = 0; j < p_width; j += 8) {
351       const int32_t src_x = (p_col + j + 4) << subsampling_x;
352       const int32_t src_y = (p_row + i + 4) << subsampling_y;
353       const int64_t dst_x =
354           (int64_t)mat[2] * src_x + (int64_t)mat[3] * src_y + (int64_t)mat[0];
355       const int64_t dst_y =
356           (int64_t)mat[4] * src_x + (int64_t)mat[5] * src_y + (int64_t)mat[1];
357 
358       const int64_t x4 = dst_x >> subsampling_x;
359       const int64_t y4 = dst_y >> subsampling_y;
360 
361       int16x8_t tmp[15];
362       warp_affine_horizontal(ref, width, height, stride, p_width, p_height,
363                              alpha, beta, x4, y4, i, tmp);
364       warp_affine_vertical(pred, p_width, p_height, p_stride, is_compound, dst,
365                            dst_stride, do_average, use_dist_wtd_comp_avg, gamma,
366                            delta, y4, i, j, tmp, w0, w1);
367     }
368   }
369 }
370 
371 #endif  // AOM_AV1_COMMON_ARM_WARP_PLANE_NEON_H_
372