xref: /aosp_15_r20/external/libaom/av1/common/arm/av1_convolve_scale_neon_i8mm.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
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 <assert.h>
13 #include <arm_neon.h>
14 #include <stddef.h>
15 #include <stdint.h>
16 
17 #include "config/aom_config.h"
18 #include "config/av1_rtcd.h"
19 
20 #include "aom_dsp/aom_dsp_common.h"
21 #include "aom_dsp/aom_filter.h"
22 #include "aom_dsp/arm/mem_neon.h"
23 #include "aom_dsp/arm/transpose_neon.h"
24 #include "aom_ports/mem.h"
25 #include "av1/common/arm/convolve_scale_neon.h"
26 #include "av1/common/convolve.h"
27 #include "av1/common/enums.h"
28 #include "av1/common/filter.h"
29 
30 // clang-format off
31 DECLARE_ALIGNED(16, static const uint8_t, kScale2DotProdPermuteTbl[32]) = {
32   0, 1, 2, 3, 2, 3, 4, 5, 4, 5,  6,  7,  6,  7,  8,  9,
33   4, 5, 6, 7, 6, 7, 8, 9, 8, 9, 10, 11, 10, 11, 12, 13
34 };
35 // clang-format on
36 
convolve8_4_h(const uint8x8_t s0,const uint8x8_t s1,const uint8x8_t s2,const uint8x8_t s3,const int8x8_t filter,const int32x4_t horiz_const)37 static inline int16x4_t convolve8_4_h(const uint8x8_t s0, const uint8x8_t s1,
38                                       const uint8x8_t s2, const uint8x8_t s3,
39                                       const int8x8_t filter,
40                                       const int32x4_t horiz_const) {
41   const int8x16_t filters = vcombine_s8(filter, filter);
42 
43   uint8x16_t s01 = vcombine_u8(s0, s1);
44   uint8x16_t s23 = vcombine_u8(s2, s3);
45 
46   int32x4_t sum01 = vusdotq_s32(horiz_const, s01, filters);
47   int32x4_t sum23 = vusdotq_s32(horiz_const, s23, filters);
48 
49   int32x4_t sum = vpaddq_s32(sum01, sum23);
50 
51   // We halved the filter values so -1 from right shift.
52   return vshrn_n_s32(sum, ROUND0_BITS - 1);
53 }
54 
convolve8_8_h(const uint8x8_t s0,const uint8x8_t s1,const uint8x8_t s2,const uint8x8_t s3,const uint8x8_t s4,const uint8x8_t s5,const uint8x8_t s6,const uint8x8_t s7,const int8x8_t filter,const int32x4_t horiz_const)55 static inline int16x8_t convolve8_8_h(const uint8x8_t s0, const uint8x8_t s1,
56                                       const uint8x8_t s2, const uint8x8_t s3,
57                                       const uint8x8_t s4, const uint8x8_t s5,
58                                       const uint8x8_t s6, const uint8x8_t s7,
59                                       const int8x8_t filter,
60                                       const int32x4_t horiz_const) {
61   const int8x16_t filters = vcombine_s8(filter, filter);
62 
63   uint8x16_t s01 = vcombine_u8(s0, s1);
64   uint8x16_t s23 = vcombine_u8(s2, s3);
65   uint8x16_t s45 = vcombine_u8(s4, s5);
66   uint8x16_t s67 = vcombine_u8(s6, s7);
67 
68   int32x4_t sum01 = vusdotq_s32(horiz_const, s01, filters);
69   int32x4_t sum23 = vusdotq_s32(horiz_const, s23, filters);
70   int32x4_t sum45 = vusdotq_s32(horiz_const, s45, filters);
71   int32x4_t sum67 = vusdotq_s32(horiz_const, s67, filters);
72 
73   int32x4_t sum0123 = vpaddq_s32(sum01, sum23);
74   int32x4_t sum4567 = vpaddq_s32(sum45, sum67);
75 
76   // We halved the filter values so -1 from right shift.
77   return vcombine_s16(vshrn_n_s32(sum0123, ROUND0_BITS - 1),
78                       vshrn_n_s32(sum4567, ROUND0_BITS - 1));
79 }
80 
convolve_horiz_scale_neon_i8mm(const uint8_t * src,int src_stride,int16_t * dst,int dst_stride,int w,int h,const int16_t * x_filter,const int subpel_x_qn,const int x_step_qn)81 static inline void convolve_horiz_scale_neon_i8mm(const uint8_t *src,
82                                                   int src_stride, int16_t *dst,
83                                                   int dst_stride, int w, int h,
84                                                   const int16_t *x_filter,
85                                                   const int subpel_x_qn,
86                                                   const int x_step_qn) {
87   DECLARE_ALIGNED(16, int16_t, temp[8 * 8]);
88   const int bd = 8;
89   // A shim of 1 << (ROUND0_BITS - 1) enables us to use non-rounding
90   // shifts - which are generally faster than rounding shifts on modern CPUs.
91   // Divide the total by 4: we halved the filter values and will use a pairwise
92   // add in the convolution kernel.
93   const int32x4_t horiz_offset = vdupq_n_s32(
94       ((1 << (bd + FILTER_BITS - 1)) + (1 << (ROUND0_BITS - 1))) >> 2);
95 
96   if (w == 4) {
97     do {
98       int x_qn = subpel_x_qn;
99 
100       // Process a 4x4 tile.
101       for (int r = 0; r < 4; r++) {
102         const uint8_t *const s = &src[x_qn >> SCALE_SUBPEL_BITS];
103 
104         const ptrdiff_t filter_offset =
105             SUBPEL_TAPS * ((x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS);
106         // Filter values are all even so halve them to fit in int8_t.
107         const int8x8_t filter =
108             vshrn_n_s16(vld1q_s16(x_filter + filter_offset), 1);
109 
110         uint8x8_t t0, t1, t2, t3;
111         load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
112 
113         int16x4_t d0 = convolve8_4_h(t0, t1, t2, t3, filter, horiz_offset);
114 
115         vst1_s16(&temp[r * 4], d0);
116         x_qn += x_step_qn;
117       }
118 
119       // Transpose the 4x4 result tile and store.
120       int16x4_t d0, d1, d2, d3;
121       load_s16_4x4(temp, 4, &d0, &d1, &d2, &d3);
122 
123       transpose_elems_inplace_s16_4x4(&d0, &d1, &d2, &d3);
124 
125       store_s16_4x4(dst, dst_stride, d0, d1, d2, d3);
126 
127       dst += 4 * dst_stride;
128       src += 4 * src_stride;
129       h -= 4;
130     } while (h > 0);
131   } else {
132     do {
133       int x_qn = subpel_x_qn;
134       int16_t *d = dst;
135       int width = w;
136 
137       do {
138         // Process an 8x8 tile.
139         for (int r = 0; r < 8; r++) {
140           const uint8_t *const s = &src[(x_qn >> SCALE_SUBPEL_BITS)];
141 
142           const ptrdiff_t filter_offset =
143               SUBPEL_TAPS * ((x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS);
144           // Filter values are all even so halve them to fit in int8_t.
145           const int8x8_t filter =
146               vshrn_n_s16(vld1q_s16(x_filter + filter_offset), 1);
147 
148           uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7;
149           load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
150 
151           int16x8_t d0 = convolve8_8_h(t0, t1, t2, t3, t4, t5, t6, t7, filter,
152                                        horiz_offset);
153 
154           vst1q_s16(&temp[r * 8], d0);
155 
156           x_qn += x_step_qn;
157         }
158 
159         // Transpose the 8x8 result tile and store.
160         int16x8_t d0, d1, d2, d3, d4, d5, d6, d7;
161         load_s16_8x8(temp, 8, &d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
162 
163         transpose_elems_inplace_s16_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
164 
165         store_s16_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
166 
167         d += 8;
168         width -= 8;
169       } while (width != 0);
170 
171       dst += 8 * dst_stride;
172       src += 8 * src_stride;
173       h -= 8;
174     } while (h > 0);
175   }
176 }
177 
convolve8_4_h_scale_2(uint8x16_t samples,const int8x8_t filters,const int32x4_t horiz_const,const uint8x16x2_t permute_tbl)178 static inline int16x4_t convolve8_4_h_scale_2(uint8x16_t samples,
179                                               const int8x8_t filters,
180                                               const int32x4_t horiz_const,
181                                               const uint8x16x2_t permute_tbl) {
182   // Permute samples ready for dot product.
183   // { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5,  6,  7,  6,  7,  8,  9 }
184   // { 4, 5, 6, 7, 6, 7, 8, 9, 8, 9, 10, 11, 10, 11, 12, 13 }
185   uint8x16_t perm_samples[2] = { vqtbl1q_u8(samples, permute_tbl.val[0]),
186                                  vqtbl1q_u8(samples, permute_tbl.val[1]) };
187 
188   int32x4_t sum = vusdotq_lane_s32(horiz_const, perm_samples[0], filters, 0);
189   sum = vusdotq_lane_s32(sum, perm_samples[1], filters, 1);
190 
191   // We halved the filter values so -1 from right shift.
192   return vshrn_n_s32(sum, ROUND0_BITS - 1);
193 }
194 
convolve8_8_h_scale_2(uint8x16_t samples[2],const int8x8_t filters,const int32x4_t horiz_const,const uint8x16x2_t permute_tbl)195 static inline int16x8_t convolve8_8_h_scale_2(uint8x16_t samples[2],
196                                               const int8x8_t filters,
197                                               const int32x4_t horiz_const,
198                                               const uint8x16x2_t permute_tbl) {
199   // Permute samples ready for dot product.
200   // { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5,  6,  7,  6,  7,  8,  9 }
201   // { 4, 5, 6, 7, 6, 7, 8, 9, 8, 9, 10, 11, 10, 11, 12, 13 }
202   uint8x16_t perm_samples[4] = { vqtbl1q_u8(samples[0], permute_tbl.val[0]),
203                                  vqtbl1q_u8(samples[0], permute_tbl.val[1]),
204                                  vqtbl1q_u8(samples[1], permute_tbl.val[0]),
205                                  vqtbl1q_u8(samples[1], permute_tbl.val[1]) };
206 
207   // First 4 output values.
208   int32x4_t sum0123 =
209       vusdotq_lane_s32(horiz_const, perm_samples[0], filters, 0);
210   sum0123 = vusdotq_lane_s32(sum0123, perm_samples[1], filters, 1);
211 
212   // Second 4 output values.
213   int32x4_t sum4567 =
214       vusdotq_lane_s32(horiz_const, perm_samples[2], filters, 0);
215   sum4567 = vusdotq_lane_s32(sum4567, perm_samples[3], filters, 1);
216 
217   // We halved the filter values so -1 from right shift.
218   return vcombine_s16(vshrn_n_s32(sum0123, ROUND0_BITS - 1),
219                       vshrn_n_s32(sum4567, ROUND0_BITS - 1));
220 }
221 
convolve_horiz_scale_2_neon_i8mm(const uint8_t * src,int src_stride,int16_t * dst,int dst_stride,int w,int h,const int16_t * x_filter)222 static inline void convolve_horiz_scale_2_neon_i8mm(
223     const uint8_t *src, int src_stride, int16_t *dst, int dst_stride, int w,
224     int h, const int16_t *x_filter) {
225   const int bd = 8;
226   // A shim of 1 << (ROUND0_BITS - 1) enables us to use non-rounding
227   // shifts - which are generally faster than rounding shifts on modern CPUs.
228   // The additional -1 is needed because we are halving the filter values.
229   const int32x4_t horiz_offset =
230       vdupq_n_s32((1 << (bd + FILTER_BITS - 2)) + (1 << (ROUND0_BITS - 2)));
231 
232   const uint8x16x2_t permute_tbl = vld1q_u8_x2(kScale2DotProdPermuteTbl);
233   // Filter values are all even so halve them to fit in int8_t.
234   const int8x8_t filter = vshrn_n_s16(vld1q_s16(x_filter), 1);
235 
236   if (w == 4) {
237     do {
238       const uint8_t *s = src;
239       int16_t *d = dst;
240       int width = w;
241 
242       do {
243         uint8x16_t s0, s1, s2, s3;
244         load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3);
245 
246         int16x4_t d0 =
247             convolve8_4_h_scale_2(s0, filter, horiz_offset, permute_tbl);
248         int16x4_t d1 =
249             convolve8_4_h_scale_2(s1, filter, horiz_offset, permute_tbl);
250         int16x4_t d2 =
251             convolve8_4_h_scale_2(s2, filter, horiz_offset, permute_tbl);
252         int16x4_t d3 =
253             convolve8_4_h_scale_2(s3, filter, horiz_offset, permute_tbl);
254 
255         store_s16_4x4(d, dst_stride, d0, d1, d2, d3);
256 
257         s += 8;
258         d += 4;
259         width -= 4;
260       } while (width != 0);
261 
262       dst += 4 * dst_stride;
263       src += 4 * src_stride;
264       h -= 4;
265     } while (h > 0);
266   } else {
267     do {
268       const uint8_t *s = src;
269       int16_t *d = dst;
270       int width = w;
271 
272       do {
273         uint8x16_t s0[2], s1[2], s2[2], s3[2];
274         load_u8_16x4(s, src_stride, &s0[0], &s1[0], &s2[0], &s3[0]);
275         load_u8_16x4(s + 8, src_stride, &s0[1], &s1[1], &s2[1], &s3[1]);
276 
277         int16x8_t d0 =
278             convolve8_8_h_scale_2(s0, filter, horiz_offset, permute_tbl);
279         int16x8_t d1 =
280             convolve8_8_h_scale_2(s1, filter, horiz_offset, permute_tbl);
281         int16x8_t d2 =
282             convolve8_8_h_scale_2(s2, filter, horiz_offset, permute_tbl);
283         int16x8_t d3 =
284             convolve8_8_h_scale_2(s3, filter, horiz_offset, permute_tbl);
285 
286         store_s16_8x4(d, dst_stride, d0, d1, d2, d3);
287 
288         s += 16;
289         d += 8;
290         width -= 8;
291       } while (width != 0);
292 
293       dst += 4 * dst_stride;
294       src += 4 * src_stride;
295       h -= 4;
296     } while (h > 0);
297   }
298 }
299 
av1_convolve_2d_scale_neon_i8mm(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_qn,const int x_step_qn,const int subpel_y_qn,const int y_step_qn,ConvolveParams * conv_params)300 void av1_convolve_2d_scale_neon_i8mm(const uint8_t *src, int src_stride,
301                                      uint8_t *dst, int dst_stride, int w, int h,
302                                      const InterpFilterParams *filter_params_x,
303                                      const InterpFilterParams *filter_params_y,
304                                      const int subpel_x_qn, const int x_step_qn,
305                                      const int subpel_y_qn, const int y_step_qn,
306                                      ConvolveParams *conv_params) {
307   if (w < 4 || h < 4) {
308     av1_convolve_2d_scale_c(src, src_stride, dst, dst_stride, w, h,
309                             filter_params_x, filter_params_y, subpel_x_qn,
310                             x_step_qn, subpel_y_qn, y_step_qn, conv_params);
311     return;
312   }
313 
314   // For the interpolation 8-tap filters are used.
315   assert(filter_params_y->taps <= 8 && filter_params_x->taps <= 8);
316 
317   DECLARE_ALIGNED(32, int16_t,
318                   im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]);
319   int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
320              filter_params_y->taps;
321   int im_stride = MAX_SB_SIZE;
322   CONV_BUF_TYPE *dst16 = conv_params->dst;
323   const int dst16_stride = conv_params->dst_stride;
324 
325   // Account for needing filter_taps / 2 - 1 lines prior and filter_taps / 2
326   // lines post both horizontally and vertically.
327   const ptrdiff_t horiz_offset = filter_params_x->taps / 2 - 1;
328   const ptrdiff_t vert_offset = (filter_params_y->taps / 2 - 1) * src_stride;
329 
330   // Horizontal filter
331   if (x_step_qn != 2 * (1 << SCALE_SUBPEL_BITS)) {
332     convolve_horiz_scale_neon_i8mm(
333         src - horiz_offset - vert_offset, src_stride, im_block, im_stride, w,
334         im_h, filter_params_x->filter_ptr, subpel_x_qn, x_step_qn);
335   } else {
336     assert(subpel_x_qn < (1 << SCALE_SUBPEL_BITS));
337     // The filter index is calculated using the
338     // ((subpel_x_qn + x * x_step_qn) & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS
339     // equation, where the values of x are from 0 to w. If x_step_qn is a
340     // multiple of SCALE_SUBPEL_MASK we can leave it out of the equation.
341     const ptrdiff_t filter_offset =
342         SUBPEL_TAPS * ((subpel_x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS);
343     const int16_t *x_filter = filter_params_x->filter_ptr + filter_offset;
344 
345     // The source index is calculated using the (subpel_x_qn + x * x_step_qn) >>
346     // SCALE_SUBPEL_BITS, where the values of x are from 0 to w. If subpel_x_qn
347     // < (1 << SCALE_SUBPEL_BITS) and x_step_qn % (1 << SCALE_SUBPEL_BITS) == 0,
348     // the source index can be determined using the value x * (x_step_qn /
349     // (1 << SCALE_SUBPEL_BITS)).
350     convolve_horiz_scale_2_neon_i8mm(src - horiz_offset - vert_offset,
351                                      src_stride, im_block, im_stride, w, im_h,
352                                      x_filter);
353   }
354 
355   // Vertical filter
356   if (filter_params_y->interp_filter == MULTITAP_SHARP) {
357     if (UNLIKELY(conv_params->is_compound)) {
358       if (conv_params->do_average) {
359         if (conv_params->use_dist_wtd_comp_avg) {
360           compound_dist_wtd_convolve_vert_scale_8tap_neon(
361               im_block, im_stride, dst, dst_stride, dst16, dst16_stride, w, h,
362               filter_params_y->filter_ptr, conv_params, subpel_y_qn, y_step_qn);
363         } else {
364           compound_avg_convolve_vert_scale_8tap_neon(
365               im_block, im_stride, dst, dst_stride, dst16, dst16_stride, w, h,
366               filter_params_y->filter_ptr, subpel_y_qn, y_step_qn);
367         }
368       } else {
369         compound_convolve_vert_scale_8tap_neon(
370             im_block, im_stride, dst16, dst16_stride, w, h,
371             filter_params_y->filter_ptr, subpel_y_qn, y_step_qn);
372       }
373     } else {
374       convolve_vert_scale_8tap_neon(im_block, im_stride, dst, dst_stride, w, h,
375                                     filter_params_y->filter_ptr, subpel_y_qn,
376                                     y_step_qn);
377     }
378   } else {
379     if (UNLIKELY(conv_params->is_compound)) {
380       if (conv_params->do_average) {
381         if (conv_params->use_dist_wtd_comp_avg) {
382           compound_dist_wtd_convolve_vert_scale_6tap_neon(
383               im_block + im_stride, im_stride, dst, dst_stride, dst16,
384               dst16_stride, w, h, filter_params_y->filter_ptr, conv_params,
385               subpel_y_qn, y_step_qn);
386         } else {
387           compound_avg_convolve_vert_scale_6tap_neon(
388               im_block + im_stride, im_stride, dst, dst_stride, dst16,
389               dst16_stride, w, h, filter_params_y->filter_ptr, subpel_y_qn,
390               y_step_qn);
391         }
392       } else {
393         compound_convolve_vert_scale_6tap_neon(
394             im_block + im_stride, im_stride, dst16, dst16_stride, w, h,
395             filter_params_y->filter_ptr, subpel_y_qn, y_step_qn);
396       }
397     } else {
398       convolve_vert_scale_6tap_neon(
399           im_block + im_stride, im_stride, dst, dst_stride, w, h,
400           filter_params_y->filter_ptr, subpel_y_qn, y_step_qn);
401     }
402   }
403 }
404