1 /*
2 * Copyright (c) 2016, 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 <emmintrin.h>
14 #include <stdio.h>
15
16 #include "aom_dsp/x86/synonyms.h"
17 #include "aom_dsp/x86/sum_squares_sse2.h"
18 #include "config/aom_dsp_rtcd.h"
19
xx_loadh_64(__m128i a,const void * b)20 static inline __m128i xx_loadh_64(__m128i a, const void *b) {
21 const __m128d ad = _mm_castsi128_pd(a);
22 return _mm_castpd_si128(_mm_loadh_pd(ad, (double *)b));
23 }
24
xx_cvtsi128_si64(__m128i a)25 static inline uint64_t xx_cvtsi128_si64(__m128i a) {
26 #if AOM_ARCH_X86_64
27 return (uint64_t)_mm_cvtsi128_si64(a);
28 #else
29 {
30 uint64_t tmp;
31 _mm_storel_epi64((__m128i *)&tmp, a);
32 return tmp;
33 }
34 #endif
35 }
36
sum_squares_i16_4x4_sse2(const int16_t * src,int stride)37 static inline __m128i sum_squares_i16_4x4_sse2(const int16_t *src, int stride) {
38 const __m128i v_val_0_w = xx_loadl_64(src + 0 * stride);
39 const __m128i v_val_2_w = xx_loadl_64(src + 2 * stride);
40 const __m128i v_val_01_w = xx_loadh_64(v_val_0_w, src + 1 * stride);
41 const __m128i v_val_23_w = xx_loadh_64(v_val_2_w, src + 3 * stride);
42 const __m128i v_sq_01_d = _mm_madd_epi16(v_val_01_w, v_val_01_w);
43 const __m128i v_sq_23_d = _mm_madd_epi16(v_val_23_w, v_val_23_w);
44
45 return _mm_add_epi32(v_sq_01_d, v_sq_23_d);
46 }
47
aom_sum_squares_2d_i16_4x4_sse2(const int16_t * src,int stride)48 uint64_t aom_sum_squares_2d_i16_4x4_sse2(const int16_t *src, int stride) {
49 const __m128i v_sum_0123_d = sum_squares_i16_4x4_sse2(src, stride);
50 __m128i v_sum_d =
51 _mm_add_epi32(v_sum_0123_d, _mm_srli_epi64(v_sum_0123_d, 32));
52 v_sum_d = _mm_add_epi32(v_sum_d, _mm_srli_si128(v_sum_d, 8));
53 return (uint64_t)_mm_cvtsi128_si32(v_sum_d);
54 }
55
aom_sum_sse_2d_i16_4x4_sse2(const int16_t * src,int stride,int * sum)56 uint64_t aom_sum_sse_2d_i16_4x4_sse2(const int16_t *src, int stride, int *sum) {
57 const __m128i one_reg = _mm_set1_epi16(1);
58 const __m128i v_val_0_w = xx_loadl_64(src + 0 * stride);
59 const __m128i v_val_2_w = xx_loadl_64(src + 2 * stride);
60 __m128i v_val_01_w = xx_loadh_64(v_val_0_w, src + 1 * stride);
61 __m128i v_val_23_w = xx_loadh_64(v_val_2_w, src + 3 * stride);
62
63 __m128i v_sum_0123_d = _mm_add_epi16(v_val_01_w, v_val_23_w);
64 v_sum_0123_d = _mm_madd_epi16(v_sum_0123_d, one_reg);
65 v_sum_0123_d = _mm_add_epi32(v_sum_0123_d, _mm_srli_si128(v_sum_0123_d, 8));
66 v_sum_0123_d = _mm_add_epi32(v_sum_0123_d, _mm_srli_si128(v_sum_0123_d, 4));
67 *sum = _mm_cvtsi128_si32(v_sum_0123_d);
68
69 const __m128i v_sq_01_d = _mm_madd_epi16(v_val_01_w, v_val_01_w);
70 const __m128i v_sq_23_d = _mm_madd_epi16(v_val_23_w, v_val_23_w);
71 __m128i v_sq_0123_d = _mm_add_epi32(v_sq_01_d, v_sq_23_d);
72 v_sq_0123_d = _mm_add_epi32(v_sq_0123_d, _mm_srli_si128(v_sq_0123_d, 8));
73 v_sq_0123_d = _mm_add_epi32(v_sq_0123_d, _mm_srli_si128(v_sq_0123_d, 4));
74 return (uint64_t)_mm_cvtsi128_si32(v_sq_0123_d);
75 }
76
aom_sum_squares_2d_i16_4xn_sse2(const int16_t * src,int stride,int height)77 uint64_t aom_sum_squares_2d_i16_4xn_sse2(const int16_t *src, int stride,
78 int height) {
79 int r = 0;
80 __m128i v_acc_q = _mm_setzero_si128();
81 do {
82 const __m128i v_acc_d = sum_squares_i16_4x4_sse2(src, stride);
83 v_acc_q = _mm_add_epi32(v_acc_q, v_acc_d);
84 src += stride << 2;
85 r += 4;
86 } while (r < height);
87 const __m128i v_zext_mask_q = _mm_set1_epi64x(~0u);
88 __m128i v_acc_64 = _mm_add_epi64(_mm_srli_epi64(v_acc_q, 32),
89 _mm_and_si128(v_acc_q, v_zext_mask_q));
90 v_acc_64 = _mm_add_epi64(v_acc_64, _mm_srli_si128(v_acc_64, 8));
91 return xx_cvtsi128_si64(v_acc_64);
92 }
93
aom_sum_sse_2d_i16_4xn_sse2(const int16_t * src,int stride,int height,int * sum)94 uint64_t aom_sum_sse_2d_i16_4xn_sse2(const int16_t *src, int stride, int height,
95 int *sum) {
96 int r = 0;
97 uint64_t sse = 0;
98 do {
99 int curr_sum = 0;
100 sse += aom_sum_sse_2d_i16_4x4_sse2(src, stride, &curr_sum);
101 *sum += curr_sum;
102 src += stride << 2;
103 r += 4;
104 } while (r < height);
105 return sse;
106 }
107
108 #ifdef __GNUC__
109 // This prevents GCC/Clang from inlining this function into
110 // aom_sum_squares_2d_i16_sse2, which in turn saves some stack
111 // maintenance instructions in the common case of 4x4.
112 __attribute__((noinline))
113 #endif
114 uint64_t
aom_sum_squares_2d_i16_nxn_sse2(const int16_t * src,int stride,int width,int height)115 aom_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride, int width,
116 int height) {
117 int r = 0;
118
119 const __m128i v_zext_mask_q = _mm_set1_epi64x(~0u);
120 __m128i v_acc_q = _mm_setzero_si128();
121
122 do {
123 __m128i v_acc_d = _mm_setzero_si128();
124 int c = 0;
125 do {
126 const int16_t *b = src + c;
127
128 const __m128i v_val_0_w = xx_load_128(b + 0 * stride);
129 const __m128i v_val_1_w = xx_load_128(b + 1 * stride);
130 const __m128i v_val_2_w = xx_load_128(b + 2 * stride);
131 const __m128i v_val_3_w = xx_load_128(b + 3 * stride);
132
133 const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
134 const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
135 const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
136 const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
137
138 const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
139 const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
140
141 const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
142
143 v_acc_d = _mm_add_epi32(v_acc_d, v_sum_0123_d);
144 c += 8;
145 } while (c < width);
146
147 v_acc_q = _mm_add_epi64(v_acc_q, _mm_and_si128(v_acc_d, v_zext_mask_q));
148 v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_epi64(v_acc_d, 32));
149
150 src += 4 * stride;
151 r += 4;
152 } while (r < height);
153
154 v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8));
155 return xx_cvtsi128_si64(v_acc_q);
156 }
157
158 #ifdef __GNUC__
159 // This prevents GCC/Clang from inlining this function into
160 // aom_sum_sse_2d_i16_nxn_sse2, which in turn saves some stack
161 // maintenance instructions in the common case of 4x4.
162 __attribute__((noinline))
163 #endif
164 uint64_t
aom_sum_sse_2d_i16_nxn_sse2(const int16_t * src,int stride,int width,int height,int * sum)165 aom_sum_sse_2d_i16_nxn_sse2(const int16_t *src, int stride, int width,
166 int height, int *sum) {
167 int r = 0;
168 uint64_t result;
169 const __m128i zero_reg = _mm_setzero_si128();
170 const __m128i one_reg = _mm_set1_epi16(1);
171
172 __m128i v_sse_total = zero_reg;
173 __m128i v_sum_total = zero_reg;
174
175 do {
176 int c = 0;
177 __m128i v_sse_row = zero_reg;
178 do {
179 const int16_t *b = src + c;
180
181 __m128i v_val_0_w = xx_load_128(b + 0 * stride);
182 __m128i v_val_1_w = xx_load_128(b + 1 * stride);
183 __m128i v_val_2_w = xx_load_128(b + 2 * stride);
184 __m128i v_val_3_w = xx_load_128(b + 3 * stride);
185
186 const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
187 const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
188 const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
189 const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
190 const __m128i v_sq_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
191 const __m128i v_sq_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
192 const __m128i v_sq_0123_d = _mm_add_epi32(v_sq_01_d, v_sq_23_d);
193 v_sse_row = _mm_add_epi32(v_sse_row, v_sq_0123_d);
194
195 const __m128i v_sum_01 = _mm_add_epi16(v_val_0_w, v_val_1_w);
196 const __m128i v_sum_23 = _mm_add_epi16(v_val_2_w, v_val_3_w);
197 __m128i v_sum_0123_d = _mm_add_epi16(v_sum_01, v_sum_23);
198 v_sum_0123_d = _mm_madd_epi16(v_sum_0123_d, one_reg);
199 v_sum_total = _mm_add_epi32(v_sum_total, v_sum_0123_d);
200
201 c += 8;
202 } while (c < width);
203
204 const __m128i v_sse_row_low = _mm_unpacklo_epi32(v_sse_row, zero_reg);
205 const __m128i v_sse_row_hi = _mm_unpackhi_epi32(v_sse_row, zero_reg);
206 v_sse_row = _mm_add_epi64(v_sse_row_low, v_sse_row_hi);
207 v_sse_total = _mm_add_epi64(v_sse_total, v_sse_row);
208 src += 4 * stride;
209 r += 4;
210 } while (r < height);
211
212 v_sum_total = _mm_add_epi32(v_sum_total, _mm_srli_si128(v_sum_total, 8));
213 v_sum_total = _mm_add_epi32(v_sum_total, _mm_srli_si128(v_sum_total, 4));
214 *sum += _mm_cvtsi128_si32(v_sum_total);
215
216 v_sse_total = _mm_add_epi64(v_sse_total, _mm_srli_si128(v_sse_total, 8));
217 xx_storel_64(&result, v_sse_total);
218 return result;
219 }
220
aom_sum_squares_2d_i16_sse2(const int16_t * src,int stride,int width,int height)221 uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width,
222 int height) {
223 // 4 elements per row only requires half an XMM register, so this
224 // must be a special case, but also note that over 75% of all calls
225 // are with size == 4, so it is also the common case.
226 if (LIKELY(width == 4 && height == 4)) {
227 return aom_sum_squares_2d_i16_4x4_sse2(src, stride);
228 } else if (LIKELY(width == 4 && (height & 3) == 0)) {
229 return aom_sum_squares_2d_i16_4xn_sse2(src, stride, height);
230 } else if (LIKELY((width & 7) == 0 && (height & 3) == 0)) {
231 // Generic case
232 return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height);
233 } else {
234 return aom_sum_squares_2d_i16_c(src, stride, width, height);
235 }
236 }
237
aom_sum_sse_2d_i16_sse2(const int16_t * src,int src_stride,int width,int height,int * sum)238 uint64_t aom_sum_sse_2d_i16_sse2(const int16_t *src, int src_stride, int width,
239 int height, int *sum) {
240 if (LIKELY(width == 4 && height == 4)) {
241 return aom_sum_sse_2d_i16_4x4_sse2(src, src_stride, sum);
242 } else if (LIKELY(width == 4 && (height & 3) == 0)) {
243 return aom_sum_sse_2d_i16_4xn_sse2(src, src_stride, height, sum);
244 } else if (LIKELY((width & 7) == 0 && (height & 3) == 0)) {
245 // Generic case
246 return aom_sum_sse_2d_i16_nxn_sse2(src, src_stride, width, height, sum);
247 } else {
248 return aom_sum_sse_2d_i16_c(src, src_stride, width, height, sum);
249 }
250 }
251
252 //////////////////////////////////////////////////////////////////////////////
253 // 1D version
254 //////////////////////////////////////////////////////////////////////////////
255
aom_sum_squares_i16_64n_sse2(const int16_t * src,uint32_t n)256 static uint64_t aom_sum_squares_i16_64n_sse2(const int16_t *src, uint32_t n) {
257 const __m128i v_zext_mask_q = _mm_set1_epi64x(~0u);
258 __m128i v_acc0_q = _mm_setzero_si128();
259 __m128i v_acc1_q = _mm_setzero_si128();
260
261 const int16_t *const end = src + n;
262
263 assert(n % 64 == 0);
264
265 while (src < end) {
266 const __m128i v_val_0_w = xx_load_128(src);
267 const __m128i v_val_1_w = xx_load_128(src + 8);
268 const __m128i v_val_2_w = xx_load_128(src + 16);
269 const __m128i v_val_3_w = xx_load_128(src + 24);
270 const __m128i v_val_4_w = xx_load_128(src + 32);
271 const __m128i v_val_5_w = xx_load_128(src + 40);
272 const __m128i v_val_6_w = xx_load_128(src + 48);
273 const __m128i v_val_7_w = xx_load_128(src + 56);
274
275 const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
276 const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
277 const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
278 const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
279 const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w);
280 const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w);
281 const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w);
282 const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w);
283
284 const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
285 const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
286 const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d);
287 const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d);
288
289 const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
290 const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d);
291
292 const __m128i v_sum_d = _mm_add_epi32(v_sum_0123_d, v_sum_4567_d);
293
294 v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_and_si128(v_sum_d, v_zext_mask_q));
295 v_acc1_q = _mm_add_epi64(v_acc1_q, _mm_srli_epi64(v_sum_d, 32));
296
297 src += 64;
298 }
299
300 v_acc0_q = _mm_add_epi64(v_acc0_q, v_acc1_q);
301 v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8));
302 return xx_cvtsi128_si64(v_acc0_q);
303 }
304
aom_sum_squares_i16_sse2(const int16_t * src,uint32_t n)305 uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t n) {
306 if (n % 64 == 0) {
307 return aom_sum_squares_i16_64n_sse2(src, n);
308 } else if (n > 64) {
309 const uint32_t k = n & ~63u;
310 return aom_sum_squares_i16_64n_sse2(src, k) +
311 aom_sum_squares_i16_c(src + k, n - k);
312 } else {
313 return aom_sum_squares_i16_c(src, n);
314 }
315 }
316
317 // Accumulate sum of 16-bit elements in the vector
mm_accumulate_epi16(__m128i vec_a)318 static inline int32_t mm_accumulate_epi16(__m128i vec_a) {
319 __m128i vtmp = _mm_srli_si128(vec_a, 8);
320 vec_a = _mm_add_epi16(vec_a, vtmp);
321 vtmp = _mm_srli_si128(vec_a, 4);
322 vec_a = _mm_add_epi16(vec_a, vtmp);
323 vtmp = _mm_srli_si128(vec_a, 2);
324 vec_a = _mm_add_epi16(vec_a, vtmp);
325 return _mm_extract_epi16(vec_a, 0);
326 }
327
328 // Accumulate sum of 32-bit elements in the vector
mm_accumulate_epi32(__m128i vec_a)329 static inline int32_t mm_accumulate_epi32(__m128i vec_a) {
330 __m128i vtmp = _mm_srli_si128(vec_a, 8);
331 vec_a = _mm_add_epi32(vec_a, vtmp);
332 vtmp = _mm_srli_si128(vec_a, 4);
333 vec_a = _mm_add_epi32(vec_a, vtmp);
334 return _mm_cvtsi128_si32(vec_a);
335 }
336
aom_var_2d_u8_sse2(uint8_t * src,int src_stride,int width,int height)337 uint64_t aom_var_2d_u8_sse2(uint8_t *src, int src_stride, int width,
338 int height) {
339 uint8_t *srcp;
340 uint64_t s = 0, ss = 0;
341 __m128i vzero = _mm_setzero_si128();
342 __m128i v_acc_sum = vzero;
343 __m128i v_acc_sqs = vzero;
344 int i, j;
345
346 // Process 16 elements in a row
347 for (i = 0; i < width - 15; i += 16) {
348 srcp = src + i;
349 // Process 8 columns at a time
350 for (j = 0; j < height - 7; j += 8) {
351 __m128i vsrc[8];
352 for (int k = 0; k < 8; k++) {
353 vsrc[k] = _mm_loadu_si128((__m128i *)srcp);
354 srcp += src_stride;
355 }
356 for (int k = 0; k < 8; k++) {
357 __m128i vsrc0 = _mm_unpacklo_epi8(vsrc[k], vzero);
358 __m128i vsrc1 = _mm_unpackhi_epi8(vsrc[k], vzero);
359 v_acc_sum = _mm_add_epi16(v_acc_sum, vsrc0);
360 v_acc_sum = _mm_add_epi16(v_acc_sum, vsrc1);
361
362 __m128i vsqs0 = _mm_madd_epi16(vsrc0, vsrc0);
363 __m128i vsqs1 = _mm_madd_epi16(vsrc1, vsrc1);
364 v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs0);
365 v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs1);
366 }
367
368 // Update total sum and clear the vectors
369 s += mm_accumulate_epi16(v_acc_sum);
370 ss += mm_accumulate_epi32(v_acc_sqs);
371 v_acc_sum = vzero;
372 v_acc_sqs = vzero;
373 }
374
375 // Process remaining rows (height not a multiple of 8)
376 for (; j < height; j++) {
377 __m128i vsrc = _mm_loadu_si128((__m128i *)srcp);
378 __m128i vsrc0 = _mm_unpacklo_epi8(vsrc, vzero);
379 __m128i vsrc1 = _mm_unpackhi_epi8(vsrc, vzero);
380 v_acc_sum = _mm_add_epi16(v_acc_sum, vsrc0);
381 v_acc_sum = _mm_add_epi16(v_acc_sum, vsrc1);
382
383 __m128i vsqs0 = _mm_madd_epi16(vsrc0, vsrc0);
384 __m128i vsqs1 = _mm_madd_epi16(vsrc1, vsrc1);
385 v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs0);
386 v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs1);
387
388 srcp += src_stride;
389 }
390
391 // Update total sum and clear the vectors
392 s += mm_accumulate_epi16(v_acc_sum);
393 ss += mm_accumulate_epi32(v_acc_sqs);
394 v_acc_sum = vzero;
395 v_acc_sqs = vzero;
396 }
397
398 // Process the remaining area using C
399 srcp = src;
400 for (int k = 0; k < height; k++) {
401 for (int m = i; m < width; m++) {
402 uint8_t val = srcp[m];
403 s += val;
404 ss += val * val;
405 }
406 srcp += src_stride;
407 }
408 return (ss - s * s / (width * height));
409 }
410
411 #if CONFIG_AV1_HIGHBITDEPTH
aom_var_2d_u16_sse2(uint8_t * src,int src_stride,int width,int height)412 uint64_t aom_var_2d_u16_sse2(uint8_t *src, int src_stride, int width,
413 int height) {
414 uint16_t *srcp1 = CONVERT_TO_SHORTPTR(src), *srcp;
415 uint64_t s = 0, ss = 0;
416 __m128i vzero = _mm_setzero_si128();
417 __m128i v_acc_sum = vzero;
418 __m128i v_acc_sqs = vzero;
419 int i, j;
420
421 // Process 8 elements in a row
422 for (i = 0; i < width - 8; i += 8) {
423 srcp = srcp1 + i;
424 // Process 8 columns at a time
425 for (j = 0; j < height - 8; j += 8) {
426 __m128i vsrc[8];
427 for (int k = 0; k < 8; k++) {
428 vsrc[k] = _mm_loadu_si128((__m128i *)srcp);
429 srcp += src_stride;
430 }
431 for (int k = 0; k < 8; k++) {
432 __m128i vsrc0 = _mm_unpacklo_epi16(vsrc[k], vzero);
433 __m128i vsrc1 = _mm_unpackhi_epi16(vsrc[k], vzero);
434 v_acc_sum = _mm_add_epi32(vsrc0, v_acc_sum);
435 v_acc_sum = _mm_add_epi32(vsrc1, v_acc_sum);
436
437 __m128i vsqs0 = _mm_madd_epi16(vsrc[k], vsrc[k]);
438 v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs0);
439 }
440
441 // Update total sum and clear the vectors
442 s += mm_accumulate_epi32(v_acc_sum);
443 ss += mm_accumulate_epi32(v_acc_sqs);
444 v_acc_sum = vzero;
445 v_acc_sqs = vzero;
446 }
447
448 // Process remaining rows (height not a multiple of 8)
449 for (; j < height; j++) {
450 __m128i vsrc = _mm_loadu_si128((__m128i *)srcp);
451 __m128i vsrc0 = _mm_unpacklo_epi16(vsrc, vzero);
452 __m128i vsrc1 = _mm_unpackhi_epi16(vsrc, vzero);
453 v_acc_sum = _mm_add_epi32(vsrc0, v_acc_sum);
454 v_acc_sum = _mm_add_epi32(vsrc1, v_acc_sum);
455
456 __m128i vsqs0 = _mm_madd_epi16(vsrc, vsrc);
457 v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs0);
458 srcp += src_stride;
459 }
460
461 // Update total sum and clear the vectors
462 s += mm_accumulate_epi32(v_acc_sum);
463 ss += mm_accumulate_epi32(v_acc_sqs);
464 v_acc_sum = vzero;
465 v_acc_sqs = vzero;
466 }
467
468 // Process the remaining area using C
469 srcp = srcp1;
470 for (int k = 0; k < height; k++) {
471 for (int m = i; m < width; m++) {
472 uint16_t val = srcp[m];
473 s += val;
474 ss += val * val;
475 }
476 srcp += src_stride;
477 }
478 return (ss - s * s / (width * height));
479 }
480 #endif // CONFIG_AV1_HIGHBITDEPTH
481