xref: /aosp_15_r20/external/libaom/aom_dsp/arm/aom_scaled_convolve8_neon_dotprod.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   // Transform sample range to [-128, 127] for 8-bit signed dot product.
28   int8x16_t s01_128 = vreinterpretq_s8_u8(vsubq_u8(s01, vdupq_n_u8(128)));
29   int8x16_t s23_128 = vreinterpretq_s8_u8(vsubq_u8(s23, vdupq_n_u8(128)));
30 
31   // Accumulate into 128 << (FILTER_BITS - 1) / 2 to account for range
32   // transform.
33   const int32x4_t acc = vdupq_n_s32((128 << (FILTER_BITS - 1)) / 2);
34   int32x4_t sum01 = vdotq_s32(acc, s01_128, filter_x2);
35   int32x4_t sum23 = vdotq_s32(acc, s23_128, filter_x2);
36 
37   int32x4_t sum0123 = vpaddq_s32(sum01, sum23);
38   int16x8_t sum = vcombine_s16(vmovn_s32(sum0123), vdup_n_s16(0));
39 
40   // We halved the filter values so -1 from right shift.
41   return vqrshrun_n_s16(sum, FILTER_BITS - 1);
42 }
43 
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)44 static inline uint8x8_t convolve8_8_h(uint8x8_t s0, uint8x8_t s1, uint8x8_t s2,
45                                       uint8x8_t s3, uint8x8_t s4, uint8x8_t s5,
46                                       uint8x8_t s6, uint8x8_t s7,
47                                       int8x8_t filter) {
48   int8x16_t filter_x2 = vcombine_s8(filter, filter);
49 
50   uint8x16_t s01 = vcombine_u8(s0, s1);
51   uint8x16_t s23 = vcombine_u8(s2, s3);
52   uint8x16_t s45 = vcombine_u8(s4, s5);
53   uint8x16_t s67 = vcombine_u8(s6, s7);
54 
55   // Transform sample range to [-128, 127] for 8-bit signed dot product.
56   int8x16_t s01_128 = vreinterpretq_s8_u8(vsubq_u8(s01, vdupq_n_u8(128)));
57   int8x16_t s23_128 = vreinterpretq_s8_u8(vsubq_u8(s23, vdupq_n_u8(128)));
58   int8x16_t s45_128 = vreinterpretq_s8_u8(vsubq_u8(s45, vdupq_n_u8(128)));
59   int8x16_t s67_128 = vreinterpretq_s8_u8(vsubq_u8(s67, vdupq_n_u8(128)));
60 
61   // Accumulate into 128 << (FILTER_BITS - 1) / 2 to account for range
62   // transform.
63   const int32x4_t acc = vdupq_n_s32((128 << (FILTER_BITS - 1)) / 2);
64   int32x4_t sum01 = vdotq_s32(acc, s01_128, filter_x2);
65   int32x4_t sum23 = vdotq_s32(acc, s23_128, filter_x2);
66   int32x4_t sum45 = vdotq_s32(acc, s45_128, filter_x2);
67   int32x4_t sum67 = vdotq_s32(acc, s67_128, filter_x2);
68 
69   int32x4_t sum0123 = vpaddq_s32(sum01, sum23);
70   int32x4_t sum4567 = vpaddq_s32(sum45, sum67);
71   int16x8_t sum = vcombine_s16(vmovn_s32(sum0123), vmovn_s32(sum4567));
72 
73   // We halved the filter values so -1 from right shift.
74   return vqrshrun_n_s16(sum, FILTER_BITS - 1);
75 }
76 
scaled_convolve_horiz_neon_dotprod(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)77 static inline void scaled_convolve_horiz_neon_dotprod(
78     const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
79     const ptrdiff_t dst_stride, const InterpKernel *const x_filter,
80     const int x0_q4, const int x_step_q4, int w, int h) {
81   DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]);
82 
83   if (w == 4) {
84     do {
85       int x_q4 = x0_q4;
86 
87       // Process a 4x4 tile.
88       for (int r = 0; r < 4; ++r) {
89         // Halve filter values (all even) to avoid the need for saturating
90         // arithmetic in convolution kernels.
91         const int8x8_t filter =
92             vshrn_n_s16(vld1q_s16(x_filter[x_q4 & SUBPEL_MASK]), 1);
93 
94         const uint8_t *s = &src[x_q4 >> SUBPEL_BITS];
95         uint8x8_t s0, s1, s2, s3;
96         load_u8_8x4(s, src_stride, &s0, &s1, &s2, &s3);
97 
98         uint8x8_t d0 = convolve8_4_h(s0, s1, s2, s3, filter);
99 
100         store_u8_4x1(&temp[4 * r], d0);
101 
102         x_q4 += x_step_q4;
103       }
104 
105       // Transpose the 4x4 result tile and store.
106       uint8x8_t d01 = vld1_u8(temp + 0);
107       uint8x8_t d23 = vld1_u8(temp + 8);
108 
109       transpose_elems_inplace_u8_4x4(&d01, &d23);
110 
111       store_u8x4_strided_x2(dst + 0 * dst_stride, 2 * dst_stride, d01);
112       store_u8x4_strided_x2(dst + 1 * dst_stride, 2 * dst_stride, d23);
113 
114       src += 4 * src_stride;
115       dst += 4 * dst_stride;
116       h -= 4;
117     } while (h > 0);
118     return;
119   }
120 
121   // w >= 8
122   do {
123     int x_q4 = x0_q4;
124     uint8_t *d = dst;
125     int width = w;
126 
127     do {
128       // Process an 8x8 tile.
129       for (int r = 0; r < 8; ++r) {
130         // Halve filter values (all even) to avoid the need for saturating
131         // arithmetic in convolution kernels.
132         const int8x8_t filter =
133             vshrn_n_s16(vld1q_s16(x_filter[x_q4 & SUBPEL_MASK]), 1);
134 
135         const uint8_t *s = &src[x_q4 >> SUBPEL_BITS];
136         uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7;
137         load_u8_8x8(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
138 
139         uint8x8_t d0 = convolve8_8_h(s0, s1, s2, s3, s4, s5, s6, s7, filter);
140 
141         vst1_u8(&temp[r * 8], d0);
142 
143         x_q4 += x_step_q4;
144       }
145 
146       // Transpose the 8x8 result tile and store.
147       uint8x8_t d0, d1, d2, d3, d4, d5, d6, d7;
148       load_u8_8x8(temp, 8, &d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
149 
150       transpose_elems_inplace_u8_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
151 
152       store_u8_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
153 
154       d += 8;
155       width -= 8;
156     } while (width != 0);
157 
158     src += 8 * src_stride;
159     dst += 8 * dst_stride;
160     h -= 8;
161   } while (h > 0);
162 }
163 
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)164 static inline uint8x8_t convolve8_4_v(uint8x8_t s0, uint8x8_t s1, uint8x8_t s2,
165                                       uint8x8_t s3, uint8x8_t s4, uint8x8_t s5,
166                                       uint8x8_t s6, uint8x8_t s7,
167                                       int8x8_t filter) {
168   uint8x16_t s01 = vcombine_u8(vzip1_u8(s0, s1), vdup_n_u8(0));
169   uint8x16_t s23 = vcombine_u8(vzip1_u8(s2, s3), vdup_n_u8(0));
170   uint8x16_t s45 = vcombine_u8(vzip1_u8(s4, s5), vdup_n_u8(0));
171   uint8x16_t s67 = vcombine_u8(vzip1_u8(s6, s7), vdup_n_u8(0));
172 
173   uint8x16_t s0123 = vreinterpretq_u8_u16(
174       vzip1q_u16(vreinterpretq_u16_u8(s01), vreinterpretq_u16_u8(s23)));
175   uint8x16_t s4567 = vreinterpretq_u8_u16(
176       vzip1q_u16(vreinterpretq_u16_u8(s45), vreinterpretq_u16_u8(s67)));
177 
178   // Transform sample range to [-128, 127] for 8-bit signed dot product.
179   int8x16_t s0123_128 = vreinterpretq_s8_u8(vsubq_u8(s0123, vdupq_n_u8(128)));
180   int8x16_t s4567_128 = vreinterpretq_s8_u8(vsubq_u8(s4567, vdupq_n_u8(128)));
181 
182   // Accumulate into 128 << (FILTER_BITS - 1) to account for range transform.
183   int32x4_t sum = vdupq_n_s32(128 << (FILTER_BITS - 1));
184   sum = vdotq_lane_s32(sum, s0123_128, filter, 0);
185   sum = vdotq_lane_s32(sum, s4567_128, filter, 1);
186 
187   // We halved the filter values so -1 from right shift.
188   return vqrshrun_n_s16(vcombine_s16(vmovn_s32(sum), vdup_n_s16(0)),
189                         FILTER_BITS - 1);
190 }
191 
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)192 static inline uint8x8_t convolve8_8_v(uint8x8_t s0, uint8x8_t s1, uint8x8_t s2,
193                                       uint8x8_t s3, uint8x8_t s4, uint8x8_t s5,
194                                       uint8x8_t s6, uint8x8_t s7,
195                                       int8x8_t filter) {
196   uint8x16_t s01 =
197       vzip1q_u8(vcombine_u8(s0, vdup_n_u8(0)), vcombine_u8(s1, vdup_n_u8(0)));
198   uint8x16_t s23 =
199       vzip1q_u8(vcombine_u8(s2, vdup_n_u8(0)), vcombine_u8(s3, vdup_n_u8(0)));
200   uint8x16_t s45 =
201       vzip1q_u8(vcombine_u8(s4, vdup_n_u8(0)), vcombine_u8(s5, vdup_n_u8(0)));
202   uint8x16_t s67 =
203       vzip1q_u8(vcombine_u8(s6, vdup_n_u8(0)), vcombine_u8(s7, vdup_n_u8(0)));
204 
205   uint8x16_t s0123[2] = {
206     vreinterpretq_u8_u16(
207         vzip1q_u16(vreinterpretq_u16_u8(s01), vreinterpretq_u16_u8(s23))),
208     vreinterpretq_u8_u16(
209         vzip2q_u16(vreinterpretq_u16_u8(s01), vreinterpretq_u16_u8(s23)))
210   };
211   uint8x16_t s4567[2] = {
212     vreinterpretq_u8_u16(
213         vzip1q_u16(vreinterpretq_u16_u8(s45), vreinterpretq_u16_u8(s67))),
214     vreinterpretq_u8_u16(
215         vzip2q_u16(vreinterpretq_u16_u8(s45), vreinterpretq_u16_u8(s67)))
216   };
217 
218   // Transform sample range to [-128, 127] for 8-bit signed dot product.
219   int8x16_t s0123_128[2] = {
220     vreinterpretq_s8_u8(vsubq_u8(s0123[0], vdupq_n_u8(128))),
221     vreinterpretq_s8_u8(vsubq_u8(s0123[1], vdupq_n_u8(128)))
222   };
223   int8x16_t s4567_128[2] = {
224     vreinterpretq_s8_u8(vsubq_u8(s4567[0], vdupq_n_u8(128))),
225     vreinterpretq_s8_u8(vsubq_u8(s4567[1], vdupq_n_u8(128)))
226   };
227 
228   // Accumulate into 128 << (FILTER_BITS - 1) to account for range transform.
229   const int32x4_t acc = vdupq_n_s32(128 << (FILTER_BITS - 1));
230 
231   int32x4_t sum0123 = vdotq_lane_s32(acc, s0123_128[0], filter, 0);
232   sum0123 = vdotq_lane_s32(sum0123, s4567_128[0], filter, 1);
233 
234   int32x4_t sum4567 = vdotq_lane_s32(acc, s0123_128[1], filter, 0);
235   sum4567 = vdotq_lane_s32(sum4567, s4567_128[1], filter, 1);
236 
237   int16x8_t sum = vcombine_s16(vmovn_s32(sum0123), vmovn_s32(sum4567));
238   // We halved the filter values so -1 from right shift.
239   return vqrshrun_n_s16(sum, FILTER_BITS - 1);
240 }
241 
scaled_convolve_vert_neon_dotprod(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)242 static inline void scaled_convolve_vert_neon_dotprod(
243     const uint8_t *src, const ptrdiff_t src_stride, uint8_t *dst,
244     const ptrdiff_t dst_stride, const InterpKernel *const y_filter,
245     const int y0_q4, const int y_step_q4, int w, int h) {
246   int y_q4 = y0_q4;
247 
248   if (w == 4) {
249     do {
250       const uint8_t *s = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
251 
252       if (y_q4 & SUBPEL_MASK) {
253         // Halve filter values (all even) to avoid the need for saturating
254         // arithmetic in convolution kernels.
255         const int8x8_t filter =
256             vshrn_n_s16(vld1q_s16(y_filter[y_q4 & SUBPEL_MASK]), 1);
257 
258         uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7;
259         load_u8_8x8(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
260 
261         uint8x8_t d0 = convolve8_4_v(s0, s1, s2, s3, s4, s5, s6, s7, filter);
262 
263         store_u8_4x1(dst, d0);
264       } else {
265         // Memcpy for non-subpel locations.
266         memcpy(dst, &s[(SUBPEL_TAPS / 2 - 1) * src_stride], 4);
267       }
268 
269       y_q4 += y_step_q4;
270       dst += dst_stride;
271     } while (--h != 0);
272     return;
273   }
274 
275   // w >= 8
276   do {
277     const uint8_t *s = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
278     uint8_t *d = dst;
279     int width = w;
280 
281     if (y_q4 & SUBPEL_MASK) {
282       // Halve filter values (all even) to avoid the need for saturating
283       // arithmetic in convolution kernels.
284       const int8x8_t filter =
285           vshrn_n_s16(vld1q_s16(y_filter[y_q4 & SUBPEL_MASK]), 1);
286 
287       do {
288         uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7;
289         load_u8_8x8(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
290 
291         uint8x8_t d0 = convolve8_8_v(s0, s1, s2, s3, s4, s5, s6, s7, filter);
292 
293         vst1_u8(d, d0);
294 
295         s += 8;
296         d += 8;
297         width -= 8;
298       } while (width != 0);
299     } else {
300       // Memcpy for non-subpel locations.
301       s += (SUBPEL_TAPS / 2 - 1) * src_stride;
302 
303       do {
304         uint8x8_t s0 = vld1_u8(s);
305         vst1_u8(d, s0);
306         s += 8;
307         d += 8;
308         width -= 8;
309       } while (width != 0);
310     }
311 
312     y_q4 += y_step_q4;
313     dst += dst_stride;
314   } while (--h != 0);
315 }
316 
aom_scaled_2d_neon_dotprod(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)317 void aom_scaled_2d_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride,
318                                 uint8_t *dst, ptrdiff_t dst_stride,
319                                 const InterpKernel *filter, int x0_q4,
320                                 int x_step_q4, int y0_q4, int y_step_q4, int w,
321                                 int h) {
322   // Fixed size intermediate buffer, im_block, places limits on parameters.
323   // 2d filtering proceeds in 2 steps:
324   //   (1) Interpolate horizontally into an intermediate buffer, temp.
325   //   (2) Interpolate temp vertically to derive the sub-pixel result.
326   // Deriving the maximum number of rows in the im_block buffer (135):
327   // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
328   // --Largest block size is 64x64 pixels.
329   // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
330   //   original frame (in 1/16th pixel units).
331   // --Must round-up because block may be located at sub-pixel position.
332   // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
333   // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
334   // --Require an additional 8 rows for the horiz_w8 transpose tail.
335   // When calling in frame scaling function, the smallest scaling factor is x1/4
336   // ==> y_step_q4 = 64. Since w and h are at most 16, the temp buffer is still
337   // big enough.
338   DECLARE_ALIGNED(16, uint8_t, im_block[(135 + 8) * 64]);
339   const int im_height =
340       (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
341   const ptrdiff_t im_stride = 64;
342 
343   assert(w <= 64);
344   assert(h <= 64);
345   assert(y_step_q4 <= 32 || (y_step_q4 <= 64 && h <= 32));
346   assert(x_step_q4 <= 64);
347 
348   // Account for needing SUBPEL_TAPS / 2 - 1 lines prior and SUBPEL_TAPS / 2
349   // lines post both horizontally and vertically.
350   const ptrdiff_t horiz_offset = SUBPEL_TAPS / 2 - 1;
351   const ptrdiff_t vert_offset = (SUBPEL_TAPS / 2 - 1) * src_stride;
352 
353   scaled_convolve_horiz_neon_dotprod(src - horiz_offset - vert_offset,
354                                      src_stride, im_block, im_stride, filter,
355                                      x0_q4, x_step_q4, w, im_height);
356 
357   scaled_convolve_vert_neon_dotprod(im_block, im_stride, dst, dst_stride,
358                                     filter, y0_q4, y_step_q4, w, h);
359 }
360