• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2020, 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 
14 #include <arm_neon.h>
15 
16 #include "av1/encoder/rdopt.h"
17 #include "config/av1_rtcd.h"
18 
19 // Process horizontal and vertical correlations in a 4x4 block of pixels.
20 // We actually use the 4x4 pixels to calculate correlations corresponding to
21 // the top-left 3x3 pixels, so this function must be called with 1x1 overlap,
22 // moving the window along/down by 3 pixels at a time.
horver_correlation_4x4(const int16_t * diff,int stride,int32x4_t * xy_sum_32,int32x4_t * xz_sum_32,int32x4_t * x_sum_32,int32x4_t * x2_sum_32)23 INLINE static void horver_correlation_4x4(const int16_t *diff, int stride,
24                                           int32x4_t *xy_sum_32,
25                                           int32x4_t *xz_sum_32,
26                                           int32x4_t *x_sum_32,
27                                           int32x4_t *x2_sum_32) {
28   // Pixels in this 4x4   [ a b c d ]
29   // are referred to as:  [ e f g h ]
30   //                      [ i j k l ]
31   //                      [ m n o p ]
32 
33   const int16x4_t pixelsa_2_lo = vld1_s16(diff + (0 * stride));
34   const int16x4_t pixelsa_2_sli =
35       vreinterpret_s16_s64(vshl_n_s64(vreinterpret_s64_s16(pixelsa_2_lo), 16));
36   const int16x4_t pixelsb_2_lo = vld1_s16(diff + (1 * stride));
37   const int16x4_t pixelsb_2_sli =
38       vreinterpret_s16_s64(vshl_n_s64(vreinterpret_s64_s16(pixelsb_2_lo), 16));
39   const int16x4_t pixelsa_1_lo = vld1_s16(diff + (2 * stride));
40   const int16x4_t pixelsa_1_sli =
41       vreinterpret_s16_s64(vshl_n_s64(vreinterpret_s64_s16(pixelsa_1_lo), 16));
42   const int16x4_t pixelsb_1_lo = vld1_s16(diff + (3 * stride));
43   const int16x4_t pixelsb_1_sli =
44       vreinterpret_s16_s64(vshl_n_s64(vreinterpret_s64_s16(pixelsb_1_lo), 16));
45 
46   const int16x8_t slli_a = vcombine_s16(pixelsa_1_sli, pixelsa_2_sli);
47 
48   *xy_sum_32 = vmlal_s16(*xy_sum_32, pixelsa_1_lo, pixelsa_1_sli);
49   *xy_sum_32 = vmlal_s16(*xy_sum_32, pixelsa_2_lo, pixelsa_2_sli);
50   *xy_sum_32 = vmlal_s16(*xy_sum_32, pixelsb_2_lo, pixelsb_2_sli);
51 
52   *xz_sum_32 = vmlal_s16(*xz_sum_32, pixelsa_1_sli, pixelsb_1_sli);
53   *xz_sum_32 = vmlal_s16(*xz_sum_32, pixelsa_2_sli, pixelsb_2_sli);
54   *xz_sum_32 = vmlal_s16(*xz_sum_32, pixelsa_1_sli, pixelsb_2_sli);
55 
56   // Now calculate the straight sums, x_sum += a+b+c+e+f+g+i+j+k
57   // (sum up every element in slli_a and swap_b)
58   *x_sum_32 = vpadalq_s16(*x_sum_32, slli_a);
59   *x_sum_32 = vaddw_s16(*x_sum_32, pixelsb_2_sli);
60 
61   // Also sum their squares
62   *x2_sum_32 = vmlal_s16(*x2_sum_32, pixelsa_1_sli, pixelsa_1_sli);
63   *x2_sum_32 = vmlal_s16(*x2_sum_32, pixelsa_2_sli, pixelsa_2_sli);
64   *x2_sum_32 = vmlal_s16(*x2_sum_32, pixelsb_2_sli, pixelsb_2_sli);
65 }
66 
av1_get_horver_correlation_full_neon(const int16_t * diff,int stride,int width,int height,float * hcorr,float * vcorr)67 void av1_get_horver_correlation_full_neon(const int16_t *diff, int stride,
68                                           int width, int height, float *hcorr,
69                                           float *vcorr) {
70   // The following notation is used:
71   // x - current pixel
72   // y - right neighbour pixel
73   // z - below neighbour pixel
74   // w - down-right neighbour pixel
75   int64_t xy_sum = 0, xz_sum = 0;
76   int64_t x_sum = 0, x2_sum = 0;
77   int32x4_t zero = vdupq_n_s32(0);
78   int64x2_t v_x_sum = vreinterpretq_s64_s32(zero);
79   int64x2_t v_xy_sum = vreinterpretq_s64_s32(zero);
80   int64x2_t v_xz_sum = vreinterpretq_s64_s32(zero);
81   int64x2_t v_x2_sum = vreinterpretq_s64_s32(zero);
82   // Process horizontal and vertical correlations through the body in 4x4
83   // blocks.  This excludes the final row and column and possibly one extra
84   // column depending how 3 divides into width and height
85 
86   for (int i = 0; i <= height - 4; i += 3) {
87     int32x4_t xy_sum_32 = zero;
88     int32x4_t xz_sum_32 = zero;
89     int32x4_t x_sum_32 = zero;
90     int32x4_t x2_sum_32 = zero;
91     for (int j = 0; j <= width - 4; j += 3) {
92       horver_correlation_4x4(&diff[i * stride + j], stride, &xy_sum_32,
93                              &xz_sum_32, &x_sum_32, &x2_sum_32);
94     }
95     v_xy_sum = vpadalq_s32(v_xy_sum, xy_sum_32);
96     v_xz_sum = vpadalq_s32(v_xz_sum, xz_sum_32);
97     v_x_sum = vpadalq_s32(v_x_sum, x_sum_32);
98     v_x2_sum = vpadalq_s32(v_x2_sum, x2_sum_32);
99   }
100 #if defined(__aarch64__)
101   xy_sum = vaddvq_s64(v_xy_sum);
102   xz_sum = vaddvq_s64(v_xz_sum);
103   x2_sum = vaddvq_s64(v_x2_sum);
104   x_sum = vaddvq_s64(v_x_sum);
105 #else
106   xy_sum = vget_lane_s64(
107       vadd_s64(vget_low_s64(v_xy_sum), vget_high_s64(v_xy_sum)), 0);
108   xz_sum = vget_lane_s64(
109       vadd_s64(vget_low_s64(v_xz_sum), vget_high_s64(v_xz_sum)), 0);
110   x2_sum = vget_lane_s64(
111       vadd_s64(vget_low_s64(v_x2_sum), vget_high_s64(v_x2_sum)), 0);
112   x_sum =
113       vget_lane_s64(vadd_s64(vget_low_s64(v_x_sum), vget_high_s64(v_x_sum)), 0);
114 #endif
115   // x_sum now covers every pixel except the final 1-2 rows and 1-2 cols
116   int64_t x_finalrow = 0, x_finalcol = 0, x2_finalrow = 0, x2_finalcol = 0;
117 
118   // Do we have 2 rows remaining or just the one?  Note that width and height
119   // are powers of 2, so each modulo 3 must be 1 or 2.
120   if (height % 3 == 1) {  // Just horiz corrs on the final row
121     const int16_t x0 = diff[(height - 1) * stride];
122     x_sum += x0;
123     x_finalrow += x0;
124     x2_sum += x0 * x0;
125     x2_finalrow += x0 * x0;
126     if (width >= 8) {
127       int32x4_t v_y_sum = zero;
128       int32x4_t v_y2_sum = zero;
129       int32x4_t v_xy_sum_a = zero;
130       int k = width - 1;
131       int j = 0;
132       while ((k - 8) > 0) {
133         const int16x8_t v_x = vld1q_s16(&diff[(height - 1) * stride + j]);
134         const int16x8_t v_y = vld1q_s16(&diff[(height - 1) * stride + j + 1]);
135         const int16x4_t v_x_lo = vget_low_s16(v_x);
136         const int16x4_t v_x_hi = vget_high_s16(v_x);
137         const int16x4_t v_y_lo = vget_low_s16(v_y);
138         const int16x4_t v_y_hi = vget_high_s16(v_y);
139         v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_x_lo, v_y_lo);
140         v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_x_hi, v_y_hi);
141         v_y2_sum = vmlal_s16(v_y2_sum, v_y_lo, v_y_lo);
142         v_y2_sum = vmlal_s16(v_y2_sum, v_y_hi, v_y_hi);
143         v_y_sum = vpadalq_s16(v_y_sum, v_y);
144         k -= 8;
145         j += 8;
146       }
147 
148       const int16x8_t v_l = vld1q_s16(&diff[(height - 1) * stride] + j);
149       const int16x8_t v_x =
150           vextq_s16(vextq_s16(vreinterpretq_s16_s32(zero), v_l, 7),
151                     vreinterpretq_s16_s32(zero), 1);
152       const int16x8_t v_y = vextq_s16(v_l, vreinterpretq_s16_s32(zero), 1);
153       const int16x4_t v_x_lo = vget_low_s16(v_x);
154       const int16x4_t v_x_hi = vget_high_s16(v_x);
155       const int16x4_t v_y_lo = vget_low_s16(v_y);
156       const int16x4_t v_y_hi = vget_high_s16(v_y);
157       v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_x_lo, v_y_lo);
158       v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_x_hi, v_y_hi);
159       v_y2_sum = vmlal_s16(v_y2_sum, v_y_lo, v_y_lo);
160       v_y2_sum = vmlal_s16(v_y2_sum, v_y_hi, v_y_hi);
161       const int32x4_t v_y_sum_a = vpadalq_s16(v_y_sum, v_y);
162       const int64x2_t v_xy_sum2 = vpaddlq_s32(v_xy_sum_a);
163 #if defined(__aarch64__)
164       const int64x2_t v_y2_sum_a = vpaddlq_s32(v_y2_sum);
165       xy_sum += vaddvq_s64(v_xy_sum2);
166       const int32_t y = vaddvq_s32(v_y_sum_a);
167       const int64_t y2 = vaddvq_s64(v_y2_sum_a);
168 #else
169       xy_sum += vget_lane_s64(
170           vadd_s64(vget_low_s64(v_xy_sum2), vget_high_s64(v_xy_sum2)), 0);
171       const int64x2_t v_y_a = vpaddlq_s32(v_y_sum_a);
172       const int64_t y =
173           vget_lane_s64(vadd_s64(vget_low_s64(v_y_a), vget_high_s64(v_y_a)), 0);
174       const int64x2_t v_y2_sum_b = vpaddlq_s32(v_y2_sum);
175       int64_t y2 = vget_lane_s64(
176           vadd_s64(vget_low_s64(v_y2_sum_b), vget_high_s64(v_y2_sum_b)), 0);
177 #endif
178       x_sum += y;
179       x2_sum += y2;
180       x_finalrow += y;
181       x2_finalrow += y2;
182     } else {
183       for (int j = 0; j < width - 1; ++j) {
184         const int16_t x = diff[(height - 1) * stride + j];
185         const int16_t y = diff[(height - 1) * stride + j + 1];
186         xy_sum += x * y;
187         x_sum += y;
188         x2_sum += y * y;
189         x_finalrow += y;
190         x2_finalrow += y * y;
191       }
192     }
193   } else {  // Two rows remaining to do
194     const int16_t x0 = diff[(height - 2) * stride];
195     const int16_t z0 = diff[(height - 1) * stride];
196     x_sum += x0 + z0;
197     x2_sum += x0 * x0 + z0 * z0;
198     x_finalrow += z0;
199     x2_finalrow += z0 * z0;
200     if (width >= 8) {
201       int32x4_t v_y2_sum = zero;
202       int32x4_t v_w2_sum = zero;
203       int32x4_t v_xy_sum_a = zero;
204       int32x4_t v_xz_sum_a = zero;
205       int32x4_t v_x_sum_a = zero;
206       int32x4_t v_w_sum = zero;
207       int k = width - 1;
208       int j = 0;
209       while ((k - 8) > 0) {
210         const int16x8_t v_x = vld1q_s16(&diff[(height - 2) * stride + j]);
211         const int16x8_t v_y = vld1q_s16(&diff[(height - 2) * stride + j + 1]);
212         const int16x8_t v_z = vld1q_s16(&diff[(height - 1) * stride + j]);
213         const int16x8_t v_w = vld1q_s16(&diff[(height - 1) * stride + j + 1]);
214 
215         const int16x4_t v_x_lo = vget_low_s16(v_x);
216         const int16x4_t v_y_lo = vget_low_s16(v_y);
217         const int16x4_t v_z_lo = vget_low_s16(v_z);
218         const int16x4_t v_w_lo = vget_low_s16(v_w);
219         const int16x4_t v_x_hi = vget_high_s16(v_x);
220         const int16x4_t v_y_hi = vget_high_s16(v_y);
221         const int16x4_t v_z_hi = vget_high_s16(v_z);
222         const int16x4_t v_w_hi = vget_high_s16(v_w);
223 
224         v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_x_lo, v_y_lo);
225         v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_x_hi, v_y_hi);
226         v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_z_lo, v_w_lo);
227         v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_z_hi, v_w_hi);
228 
229         v_xz_sum_a = vmlal_s16(v_xz_sum_a, v_x_lo, v_z_lo);
230         v_xz_sum_a = vmlal_s16(v_xz_sum_a, v_x_hi, v_z_hi);
231 
232         v_w2_sum = vmlal_s16(v_w2_sum, v_w_lo, v_w_lo);
233         v_w2_sum = vmlal_s16(v_w2_sum, v_w_hi, v_w_hi);
234         v_y2_sum = vmlal_s16(v_y2_sum, v_y_lo, v_y_lo);
235         v_y2_sum = vmlal_s16(v_y2_sum, v_y_hi, v_y_hi);
236 
237         v_w_sum = vpadalq_s16(v_w_sum, v_w);
238         v_x_sum_a = vpadalq_s16(v_x_sum_a, v_y);
239         v_x_sum_a = vpadalq_s16(v_x_sum_a, v_w);
240 
241         k -= 8;
242         j += 8;
243       }
244       const int16x8_t v_l = vld1q_s16(&diff[(height - 2) * stride] + j);
245       const int16x8_t v_x =
246           vextq_s16(vextq_s16(vreinterpretq_s16_s32(zero), v_l, 7),
247                     vreinterpretq_s16_s32(zero), 1);
248       const int16x8_t v_y = vextq_s16(v_l, vreinterpretq_s16_s32(zero), 1);
249       const int16x8_t v_l_2 = vld1q_s16(&diff[(height - 1) * stride] + j);
250       const int16x8_t v_z =
251           vextq_s16(vextq_s16(vreinterpretq_s16_s32(zero), v_l_2, 7),
252                     vreinterpretq_s16_s32(zero), 1);
253       const int16x8_t v_w = vextq_s16(v_l_2, vreinterpretq_s16_s32(zero), 1);
254 
255       const int16x4_t v_x_lo = vget_low_s16(v_x);
256       const int16x4_t v_y_lo = vget_low_s16(v_y);
257       const int16x4_t v_z_lo = vget_low_s16(v_z);
258       const int16x4_t v_w_lo = vget_low_s16(v_w);
259       const int16x4_t v_x_hi = vget_high_s16(v_x);
260       const int16x4_t v_y_hi = vget_high_s16(v_y);
261       const int16x4_t v_z_hi = vget_high_s16(v_z);
262       const int16x4_t v_w_hi = vget_high_s16(v_w);
263 
264       v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_x_lo, v_y_lo);
265       v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_x_hi, v_y_hi);
266       v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_z_lo, v_w_lo);
267       v_xy_sum_a = vmlal_s16(v_xy_sum_a, v_z_hi, v_w_hi);
268 
269       v_xz_sum_a = vmlal_s16(v_xz_sum_a, v_x_lo, v_z_lo);
270       v_xz_sum_a = vmlal_s16(v_xz_sum_a, v_x_hi, v_z_hi);
271 
272       v_w2_sum = vmlal_s16(v_w2_sum, v_w_lo, v_w_lo);
273       v_w2_sum = vmlal_s16(v_w2_sum, v_w_hi, v_w_hi);
274       v_y2_sum = vmlal_s16(v_y2_sum, v_y_lo, v_y_lo);
275       v_y2_sum = vmlal_s16(v_y2_sum, v_y_hi, v_y_hi);
276 
277       v_w_sum = vpadalq_s16(v_w_sum, v_w);
278       v_x_sum_a = vpadalq_s16(v_x_sum_a, v_y);
279       v_x_sum_a = vpadalq_s16(v_x_sum_a, v_w);
280 
281 #if defined(__aarch64__)
282       xy_sum += vaddvq_s64(vpaddlq_s32(v_xy_sum_a));
283       xz_sum += vaddvq_s64(vpaddlq_s32(v_xz_sum_a));
284       x_sum += vaddvq_s32(v_x_sum_a);
285       x_finalrow += vaddvq_s32(v_w_sum);
286       int64_t y2 = vaddvq_s64(vpaddlq_s32(v_y2_sum));
287       int64_t w2 = vaddvq_s64(vpaddlq_s32(v_w2_sum));
288 #else
289       const int64x2_t v_xy_sum2 = vpaddlq_s32(v_xy_sum_a);
290       xy_sum += vget_lane_s64(
291           vadd_s64(vget_low_s64(v_xy_sum2), vget_high_s64(v_xy_sum2)), 0);
292       const int64x2_t v_xz_sum2 = vpaddlq_s32(v_xz_sum_a);
293       xz_sum += vget_lane_s64(
294           vadd_s64(vget_low_s64(v_xz_sum2), vget_high_s64(v_xz_sum2)), 0);
295       const int64x2_t v_x_sum2 = vpaddlq_s32(v_x_sum_a);
296       x_sum += vget_lane_s64(
297           vadd_s64(vget_low_s64(v_x_sum2), vget_high_s64(v_x_sum2)), 0);
298       const int64x2_t v_w_sum_a = vpaddlq_s32(v_w_sum);
299       x_finalrow += vget_lane_s64(
300           vadd_s64(vget_low_s64(v_w_sum_a), vget_high_s64(v_w_sum_a)), 0);
301       const int64x2_t v_y2_sum_a = vpaddlq_s32(v_y2_sum);
302       int64_t y2 = vget_lane_s64(
303           vadd_s64(vget_low_s64(v_y2_sum_a), vget_high_s64(v_y2_sum_a)), 0);
304       const int64x2_t v_w2_sum_a = vpaddlq_s32(v_w2_sum);
305       int64_t w2 = vget_lane_s64(
306           vadd_s64(vget_low_s64(v_w2_sum_a), vget_high_s64(v_w2_sum_a)), 0);
307 #endif
308       x2_sum += y2 + w2;
309       x2_finalrow += w2;
310     } else {
311       for (int j = 0; j < width - 1; ++j) {
312         const int16_t x = diff[(height - 2) * stride + j];
313         const int16_t y = diff[(height - 2) * stride + j + 1];
314         const int16_t z = diff[(height - 1) * stride + j];
315         const int16_t w = diff[(height - 1) * stride + j + 1];
316 
317         // Horizontal and vertical correlations for the penultimate row:
318         xy_sum += x * y;
319         xz_sum += x * z;
320 
321         // Now just horizontal correlations for the final row:
322         xy_sum += z * w;
323 
324         x_sum += y + w;
325         x2_sum += y * y + w * w;
326         x_finalrow += w;
327         x2_finalrow += w * w;
328       }
329     }
330   }
331 
332   // Do we have 2 columns remaining or just the one?
333   if (width % 3 == 1) {  // Just vert corrs on the final col
334     const int16_t x0 = diff[width - 1];
335     x_sum += x0;
336     x_finalcol += x0;
337     x2_sum += x0 * x0;
338     x2_finalcol += x0 * x0;
339     for (int i = 0; i < height - 1; ++i) {
340       const int16_t x = diff[i * stride + width - 1];
341       const int16_t z = diff[(i + 1) * stride + width - 1];
342       xz_sum += x * z;
343       x_finalcol += z;
344       x2_finalcol += z * z;
345       // So the bottom-right elements don't get counted twice:
346       if (i < height - (height % 3 == 1 ? 2 : 3)) {
347         x_sum += z;
348         x2_sum += z * z;
349       }
350     }
351   } else {  // Two cols remaining
352     const int16_t x0 = diff[width - 2];
353     const int16_t y0 = diff[width - 1];
354     x_sum += x0 + y0;
355     x2_sum += x0 * x0 + y0 * y0;
356     x_finalcol += y0;
357     x2_finalcol += y0 * y0;
358     for (int i = 0; i < height - 1; ++i) {
359       const int16_t x = diff[i * stride + width - 2];
360       const int16_t y = diff[i * stride + width - 1];
361       const int16_t z = diff[(i + 1) * stride + width - 2];
362       const int16_t w = diff[(i + 1) * stride + width - 1];
363 
364       // Horizontal and vertical correlations for the penultimate col:
365       // Skip these on the last iteration of this loop if we also had two
366       // rows remaining, otherwise the final horizontal and vertical correlation
367       // get erroneously processed twice
368       if (i < height - 2 || height % 3 == 1) {
369         xy_sum += x * y;
370         xz_sum += x * z;
371       }
372 
373       x_finalcol += w;
374       x2_finalcol += w * w;
375       // So the bottom-right elements don't get counted twice:
376       if (i < height - (height % 3 == 1 ? 2 : 3)) {
377         x_sum += z + w;
378         x2_sum += z * z + w * w;
379       }
380 
381       // Now just vertical correlations for the final column:
382       xz_sum += y * w;
383     }
384   }
385 
386   // Calculate the simple sums and squared-sums
387   int64_t x_firstrow = 0, x_firstcol = 0;
388   int64_t x2_firstrow = 0, x2_firstcol = 0;
389 
390   if (width >= 8) {
391     int32x4_t v_x_firstrow = zero;
392     int32x4_t v_x2_firstrow = zero;
393     for (int j = 0; j < width; j += 8) {
394       const int16x8_t v_diff = vld1q_s16(diff + j);
395       const int16x4_t v_diff_lo = vget_low_s16(v_diff);
396       const int16x4_t v_diff_hi = vget_high_s16(v_diff);
397       v_x_firstrow = vpadalq_s16(v_x_firstrow, v_diff);
398       v_x2_firstrow = vmlal_s16(v_x2_firstrow, v_diff_lo, v_diff_lo);
399       v_x2_firstrow = vmlal_s16(v_x2_firstrow, v_diff_hi, v_diff_hi);
400     }
401 #if defined(__aarch64__)
402     x_firstrow += vaddvq_s32(v_x_firstrow);
403     x2_firstrow += vaddvq_s32(v_x2_firstrow);
404 #else
405     const int64x2_t v_x_firstrow_64 = vpaddlq_s32(v_x_firstrow);
406     x_firstrow += vget_lane_s64(
407         vadd_s64(vget_low_s64(v_x_firstrow_64), vget_high_s64(v_x_firstrow_64)),
408         0);
409     const int64x2_t v_x2_firstrow_64 = vpaddlq_s32(v_x2_firstrow);
410     x2_firstrow += vget_lane_s64(vadd_s64(vget_low_s64(v_x2_firstrow_64),
411                                           vget_high_s64(v_x2_firstrow_64)),
412                                  0);
413 #endif
414   } else {
415     for (int j = 0; j < width; ++j) {
416       x_firstrow += diff[j];
417       x2_firstrow += diff[j] * diff[j];
418     }
419   }
420   for (int i = 0; i < height; ++i) {
421     x_firstcol += diff[i * stride];
422     x2_firstcol += diff[i * stride] * diff[i * stride];
423   }
424 
425   int64_t xhor_sum = x_sum - x_finalcol;
426   int64_t xver_sum = x_sum - x_finalrow;
427   int64_t y_sum = x_sum - x_firstcol;
428   int64_t z_sum = x_sum - x_firstrow;
429   int64_t x2hor_sum = x2_sum - x2_finalcol;
430   int64_t x2ver_sum = x2_sum - x2_finalrow;
431   int64_t y2_sum = x2_sum - x2_firstcol;
432   int64_t z2_sum = x2_sum - x2_firstrow;
433 
434   const float num_hor = (float)(height * (width - 1));
435   const float num_ver = (float)((height - 1) * width);
436 
437   const float xhor_var_n = x2hor_sum - (xhor_sum * xhor_sum) / num_hor;
438   const float xver_var_n = x2ver_sum - (xver_sum * xver_sum) / num_ver;
439 
440   const float y_var_n = y2_sum - (y_sum * y_sum) / num_hor;
441   const float z_var_n = z2_sum - (z_sum * z_sum) / num_ver;
442 
443   const float xy_var_n = xy_sum - (xhor_sum * y_sum) / num_hor;
444   const float xz_var_n = xz_sum - (xver_sum * z_sum) / num_ver;
445 
446   if (xhor_var_n > 0 && y_var_n > 0) {
447     *hcorr = xy_var_n / sqrtf(xhor_var_n * y_var_n);
448     *hcorr = *hcorr < 0 ? 0 : *hcorr;
449   } else {
450     *hcorr = 1.0;
451   }
452   if (xver_var_n > 0 && z_var_n > 0) {
453     *vcorr = xz_var_n / sqrtf(xver_var_n * z_var_n);
454     *vcorr = *vcorr < 0 ? 0 : *vcorr;
455   } else {
456     *vcorr = 1.0;
457   }
458 }
459