xref: /aosp_15_r20/external/libaom/aom_dsp/arm/aom_scaled_convolve8_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 <arm_neon.h>
13 #include <assert.h>
14 
15 #include "aom_dsp/arm/aom_convolve8_neon.h"
16 #include "aom_dsp/arm/mem_neon.h"
17 #include "aom_dsp/arm/transpose_neon.h"
18 #include "config/aom_dsp_rtcd.h"
19 
convolve8_4_h(uint8x8_t s0,uint8x8_t s1,uint8x8_t s2,uint8x8_t s3,int8x8_t filter)20 static inline uint8x8_t convolve8_4_h(uint8x8_t s0, uint8x8_t s1, uint8x8_t s2,
21                                       uint8x8_t s3, int8x8_t filter) {
22   int8x16_t filter_x2 = vcombine_s8(filter, filter);
23 
24   uint8x16_t s01 = vcombine_u8(s0, s1);
25   uint8x16_t s23 = vcombine_u8(s2, s3);
26 
27   int32x4_t sum01 = vusdotq_s32(vdupq_n_s32(0), s01, filter_x2);
28   int32x4_t sum23 = vusdotq_s32(vdupq_n_s32(0), s23, filter_x2);
29 
30   int32x4_t sum0123 = vpaddq_s32(sum01, sum23);
31   int16x8_t sum = vcombine_s16(vmovn_s32(sum0123), vdup_n_s16(0));
32 
33   // We halved the filter values so -1 from right shift.
34   return vqrshrun_n_s16(sum, FILTER_BITS - 1);
35 }
36 
convolve8_8_h(uint8x8_t s0,uint8x8_t s1,uint8x8_t s2,uint8x8_t s3,uint8x8_t s4,uint8x8_t s5,uint8x8_t s6,uint8x8_t s7,int8x8_t filter)37 static inline uint8x8_t convolve8_8_h(uint8x8_t s0, uint8x8_t s1, uint8x8_t s2,
38                                       uint8x8_t s3, uint8x8_t s4, uint8x8_t s5,
39                                       uint8x8_t s6, uint8x8_t s7,
40                                       int8x8_t filter) {
41   int8x16_t filter_x2 = vcombine_s8(filter, filter);
42 
43   uint8x16_t s01 = vcombine_u8(s0, s1);
44   uint8x16_t s23 = vcombine_u8(s2, s3);
45   uint8x16_t s45 = vcombine_u8(s4, s5);
46   uint8x16_t s67 = vcombine_u8(s6, s7);
47 
48   int32x4_t sum01 = vusdotq_s32(vdupq_n_s32(0), s01, filter_x2);
49   int32x4_t sum23 = vusdotq_s32(vdupq_n_s32(0), s23, filter_x2);
50   int32x4_t sum45 = vusdotq_s32(vdupq_n_s32(0), s45, filter_x2);
51   int32x4_t sum67 = vusdotq_s32(vdupq_n_s32(0), s67, filter_x2);
52 
53   int32x4_t sum0123 = vpaddq_s32(sum01, sum23);
54   int32x4_t sum4567 = vpaddq_s32(sum45, sum67);
55   int16x8_t sum = vcombine_s16(vmovn_s32(sum0123), vmovn_s32(sum4567));
56 
57   // We halved the filter values so -1 from right shift.
58   return vqrshrun_n_s16(sum, FILTER_BITS - 1);
59 }
60 
scaled_convolve_horiz_neon_i8mm(const uint8_t * src,const ptrdiff_t src_stride,uint8_t * dst,const ptrdiff_t dst_stride,const InterpKernel * const x_filter,const int x0_q4,const int x_step_q4,int w,int h)61 static inline void scaled_convolve_horiz_neon_i8mm(
62     const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
63     const ptrdiff_t dst_stride, const InterpKernel *const x_filter,
64     const int x0_q4, const int x_step_q4, int w, int h) {
65   DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]);
66 
67   if (w == 4) {
68     do {
69       int x_q4 = x0_q4;
70 
71       // Process a 4x4 tile.
72       for (int r = 0; r < 4; ++r) {
73         // Halve filter values (all even) to avoid the need for saturating
74         // arithmetic in convolution kernels.
75         const int8x8_t filter =
76             vshrn_n_s16(vld1q_s16(x_filter[x_q4 & SUBPEL_MASK]), 1);
77 
78         const uint8_t *s = &src[x_q4 >> SUBPEL_BITS];
79         uint8x8_t s0, s1, s2, s3;
80         load_u8_8x4(s, src_stride, &s0, &s1, &s2, &s3);
81 
82         uint8x8_t d0 = convolve8_4_h(s0, s1, s2, s3, filter);
83 
84         store_u8_4x1(&temp[4 * r], d0);
85 
86         x_q4 += x_step_q4;
87       }
88 
89       // Transpose the 4x4 result tile and store.
90       uint8x8_t d01 = vld1_u8(temp + 0);
91       uint8x8_t d23 = vld1_u8(temp + 8);
92 
93       transpose_elems_inplace_u8_4x4(&d01, &d23);
94 
95       store_u8x4_strided_x2(dst + 0 * dst_stride, 2 * dst_stride, d01);
96       store_u8x4_strided_x2(dst + 1 * dst_stride, 2 * dst_stride, d23);
97 
98       src += 4 * src_stride;
99       dst += 4 * dst_stride;
100       h -= 4;
101     } while (h > 0);
102     return;
103   }
104 
105   // w >= 8
106   do {
107     int x_q4 = x0_q4;
108     uint8_t *d = dst;
109     int width = w;
110 
111     do {
112       // Process an 8x8 tile.
113       for (int r = 0; r < 8; ++r) {
114         // Halve filter values (all even) to avoid the need for saturating
115         // arithmetic in convolution kernels.
116         const int8x8_t filter =
117             vshrn_n_s16(vld1q_s16(x_filter[x_q4 & SUBPEL_MASK]), 1);
118 
119         const uint8_t *s = &src[x_q4 >> SUBPEL_BITS];
120         uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7;
121         load_u8_8x8(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
122 
123         uint8x8_t d0 = convolve8_8_h(s0, s1, s2, s3, s4, s5, s6, s7, filter);
124 
125         vst1_u8(&temp[r * 8], d0);
126 
127         x_q4 += x_step_q4;
128       }
129 
130       // Transpose the 8x8 result tile and store.
131       uint8x8_t d0, d1, d2, d3, d4, d5, d6, d7;
132       load_u8_8x8(temp, 8, &d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
133 
134       transpose_elems_inplace_u8_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
135 
136       store_u8_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
137 
138       d += 8;
139       width -= 8;
140     } while (width != 0);
141 
142     src += 8 * src_stride;
143     dst += 8 * dst_stride;
144     h -= 8;
145   } while (h > 0);
146 }
147 
convolve8_4_v(uint8x8_t s0,uint8x8_t s1,uint8x8_t s2,uint8x8_t s3,uint8x8_t s4,uint8x8_t s5,uint8x8_t s6,uint8x8_t s7,int8x8_t filter)148 static inline uint8x8_t convolve8_4_v(uint8x8_t s0, uint8x8_t s1, uint8x8_t s2,
149                                       uint8x8_t s3, uint8x8_t s4, uint8x8_t s5,
150                                       uint8x8_t s6, uint8x8_t s7,
151                                       int8x8_t filter) {
152   uint8x16_t s01 = vcombine_u8(vzip1_u8(s0, s1), vdup_n_u8(0));
153   uint8x16_t s23 = vcombine_u8(vzip1_u8(s2, s3), vdup_n_u8(0));
154   uint8x16_t s45 = vcombine_u8(vzip1_u8(s4, s5), vdup_n_u8(0));
155   uint8x16_t s67 = vcombine_u8(vzip1_u8(s6, s7), vdup_n_u8(0));
156 
157   uint8x16_t s0123 = vreinterpretq_u8_u16(
158       vzip1q_u16(vreinterpretq_u16_u8(s01), vreinterpretq_u16_u8(s23)));
159   uint8x16_t s4567 = vreinterpretq_u8_u16(
160       vzip1q_u16(vreinterpretq_u16_u8(s45), vreinterpretq_u16_u8(s67)));
161 
162   int32x4_t sum = vusdotq_lane_s32(vdupq_n_s32(0), s0123, filter, 0);
163   sum = vusdotq_lane_s32(sum, s4567, filter, 1);
164 
165   // We halved the filter values so -1 from right shift.
166   return vqrshrun_n_s16(vcombine_s16(vmovn_s32(sum), vdup_n_s16(0)),
167                         FILTER_BITS - 1);
168 }
169 
convolve8_8_v(uint8x8_t s0,uint8x8_t s1,uint8x8_t s2,uint8x8_t s3,uint8x8_t s4,uint8x8_t s5,uint8x8_t s6,uint8x8_t s7,int8x8_t filter)170 static inline uint8x8_t convolve8_8_v(uint8x8_t s0, uint8x8_t s1, uint8x8_t s2,
171                                       uint8x8_t s3, uint8x8_t s4, uint8x8_t s5,
172                                       uint8x8_t s6, uint8x8_t s7,
173                                       int8x8_t filter) {
174   uint8x16_t s01 =
175       vzip1q_u8(vcombine_u8(s0, vdup_n_u8(0)), vcombine_u8(s1, vdup_n_u8(0)));
176   uint8x16_t s23 =
177       vzip1q_u8(vcombine_u8(s2, vdup_n_u8(0)), vcombine_u8(s3, vdup_n_u8(0)));
178   uint8x16_t s45 =
179       vzip1q_u8(vcombine_u8(s4, vdup_n_u8(0)), vcombine_u8(s5, vdup_n_u8(0)));
180   uint8x16_t s67 =
181       vzip1q_u8(vcombine_u8(s6, vdup_n_u8(0)), vcombine_u8(s7, vdup_n_u8(0)));
182 
183   uint8x16_t s0123[2] = {
184     vreinterpretq_u8_u16(
185         vzip1q_u16(vreinterpretq_u16_u8(s01), vreinterpretq_u16_u8(s23))),
186     vreinterpretq_u8_u16(
187         vzip2q_u16(vreinterpretq_u16_u8(s01), vreinterpretq_u16_u8(s23)))
188   };
189   uint8x16_t s4567[2] = {
190     vreinterpretq_u8_u16(
191         vzip1q_u16(vreinterpretq_u16_u8(s45), vreinterpretq_u16_u8(s67))),
192     vreinterpretq_u8_u16(
193         vzip2q_u16(vreinterpretq_u16_u8(s45), vreinterpretq_u16_u8(s67)))
194   };
195 
196   int32x4_t sum0123 = vusdotq_lane_s32(vdupq_n_s32(0), s0123[0], filter, 0);
197   sum0123 = vusdotq_lane_s32(sum0123, s4567[0], filter, 1);
198 
199   int32x4_t sum4567 = vusdotq_lane_s32(vdupq_n_s32(0), s0123[1], filter, 0);
200   sum4567 = vusdotq_lane_s32(sum4567, s4567[1], filter, 1);
201 
202   int16x8_t sum = vcombine_s16(vmovn_s32(sum0123), vmovn_s32(sum4567));
203   // We halved the filter values so -1 from right shift.
204   return vqrshrun_n_s16(sum, FILTER_BITS - 1);
205 }
206 
scaled_convolve_vert_neon_i8mm(const uint8_t * src,const ptrdiff_t src_stride,uint8_t * dst,const ptrdiff_t dst_stride,const InterpKernel * const y_filter,const int y0_q4,const int y_step_q4,int w,int h)207 static inline void scaled_convolve_vert_neon_i8mm(
208     const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
209     const ptrdiff_t dst_stride, const InterpKernel *const y_filter,
210     const int y0_q4, const int y_step_q4, int w, int h) {
211   int y_q4 = y0_q4;
212 
213   if (w == 4) {
214     do {
215       const uint8_t *s = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
216 
217       if (y_q4 & SUBPEL_MASK) {
218         // Halve filter values (all even) to avoid the need for saturating
219         // arithmetic in convolution kernels.
220         const int8x8_t filter =
221             vshrn_n_s16(vld1q_s16(y_filter[y_q4 & SUBPEL_MASK]), 1);
222 
223         uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7;
224         load_u8_8x8(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
225 
226         uint8x8_t d0 = convolve8_4_v(s0, s1, s2, s3, s4, s5, s6, s7, filter);
227 
228         store_u8_4x1(dst, d0);
229       } else {
230         // Memcpy for non-subpel locations.
231         memcpy(dst, &s[(SUBPEL_TAPS / 2 - 1) * src_stride], 4);
232       }
233 
234       y_q4 += y_step_q4;
235       dst += dst_stride;
236     } while (--h != 0);
237     return;
238   }
239 
240   // w >= 8
241   do {
242     const uint8_t *s = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
243     uint8_t *d = dst;
244     int width = w;
245 
246     if (y_q4 & SUBPEL_MASK) {
247       // Halve filter values (all even) to avoid the need for saturating
248       // arithmetic in convolution kernels.
249       const int8x8_t filter =
250           vshrn_n_s16(vld1q_s16(y_filter[y_q4 & SUBPEL_MASK]), 1);
251 
252       do {
253         uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7;
254         load_u8_8x8(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
255 
256         uint8x8_t d0 = convolve8_8_v(s0, s1, s2, s3, s4, s5, s6, s7, filter);
257 
258         vst1_u8(d, d0);
259 
260         s += 8;
261         d += 8;
262         width -= 8;
263       } while (width != 0);
264     } else {
265       // Memcpy for non-subpel locations.
266       s += (SUBPEL_TAPS / 2 - 1) * src_stride;
267 
268       do {
269         uint8x8_t s0 = vld1_u8(s);
270         vst1_u8(d, s0);
271         s += 8;
272         d += 8;
273         width -= 8;
274       } while (width != 0);
275     }
276 
277     y_q4 += y_step_q4;
278     dst += dst_stride;
279   } while (--h != 0);
280 }
281 
aom_scaled_2d_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,const InterpKernel * filter,int x0_q4,int x_step_q4,int y0_q4,int y_step_q4,int w,int h)282 void aom_scaled_2d_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
283                              uint8_t *dst, ptrdiff_t dst_stride,
284                              const InterpKernel *filter, int x0_q4,
285                              int x_step_q4, int y0_q4, int y_step_q4, int w,
286                              int h) {
287   // Fixed size intermediate buffer, im_block, places limits on parameters.
288   // 2d filtering proceeds in 2 steps:
289   //   (1) Interpolate horizontally into an intermediate buffer, temp.
290   //   (2) Interpolate temp vertically to derive the sub-pixel result.
291   // Deriving the maximum number of rows in the im_block buffer (135):
292   // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
293   // --Largest block size is 64x64 pixels.
294   // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
295   //   original frame (in 1/16th pixel units).
296   // --Must round-up because block may be located at sub-pixel position.
297   // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
298   // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
299   // --Require an additional 8 rows for the horiz_w8 transpose tail.
300   // When calling in frame scaling function, the smallest scaling factor is x1/4
301   // ==> y_step_q4 = 64. Since w and h are at most 16, the temp buffer is still
302   // big enough.
303   DECLARE_ALIGNED(16, uint8_t, im_block[(135 + 8) * 64]);
304   const int im_height =
305       (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
306   const ptrdiff_t im_stride = 64;
307 
308   assert(w <= 64);
309   assert(h <= 64);
310   assert(y_step_q4 <= 32 || (y_step_q4 <= 64 && h <= 32));
311   assert(x_step_q4 <= 64);
312 
313   // Account for needing SUBPEL_TAPS / 2 - 1 lines prior and SUBPEL_TAPS / 2
314   // lines post both horizontally and vertically.
315   const ptrdiff_t horiz_offset = SUBPEL_TAPS / 2 - 1;
316   const ptrdiff_t vert_offset = (SUBPEL_TAPS / 2 - 1) * src_stride;
317 
318   scaled_convolve_horiz_neon_i8mm(src - horiz_offset - vert_offset, src_stride,
319                                   im_block, im_stride, filter, x0_q4, x_step_q4,
320                                   w, im_height);
321 
322   scaled_convolve_vert_neon_i8mm(im_block, im_stride, dst, dst_stride, filter,
323                                  y0_q4, y_step_q4, w, h);
324 }
325