1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8
9 #include <assert.h>
10
11 #include <emmintrin.h>
12
13 #include <xnnpack/avgpool.h>
14
15
xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8(size_t output_pixels,size_t kernel_elements,size_t channels,const uint8_t ** input,size_t input_offset,const uint8_t * zero,int32_t * buffer,uint8_t * output,size_t input_increment,size_t output_increment,const union xnn_qu8_avgpool_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8(
17 size_t output_pixels,
18 size_t kernel_elements,
19 size_t channels,
20 const uint8_t** input,
21 size_t input_offset,
22 const uint8_t* zero,
23 int32_t* buffer,
24 uint8_t* output,
25 size_t input_increment,
26 size_t output_increment,
27 const union xnn_qu8_avgpool_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
28 {
29 assert(output_pixels != 0);
30 assert(kernel_elements > 9);
31 assert(channels != 0);
32
33 const __m128i vbias = _mm_load_si128((const __m128i*) ¶ms->sse2.bias);
34 const __m128i vzero = _mm_setzero_si128();
35 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
36 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
37 const __m128i vright_shift = _mm_loadl_epi64((const __m128i*) params->sse2.right_shift);
38
39 do {
40 {
41 const uint8_t* i0 = *input++;
42 assert(i0 != NULL);
43 if XNN_UNPREDICTABLE(i0 != zero) {
44 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
45 }
46 const uint8_t* i1 = *input++;
47 assert(i1 != NULL);
48 if XNN_UNPREDICTABLE(i1 != zero) {
49 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
50 }
51 const uint8_t* i2 = *input++;
52 assert(i2 != NULL);
53 if XNN_UNPREDICTABLE(i2 != zero) {
54 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
55 }
56 const uint8_t* i3 = *input++;
57 assert(i3 != NULL);
58 if XNN_UNPREDICTABLE(i3 != zero) {
59 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
60 }
61 const uint8_t* i4 = *input++;
62 assert(i4 != NULL);
63 if XNN_UNPREDICTABLE(i4 != zero) {
64 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
65 }
66 const uint8_t* i5 = *input++;
67 assert(i5 != NULL);
68 if XNN_UNPREDICTABLE(i5 != zero) {
69 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
70 }
71 const uint8_t* i6 = *input++;
72 assert(i6 != NULL);
73 if XNN_UNPREDICTABLE(i6 != zero) {
74 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
75 }
76 const uint8_t* i7 = *input++;
77 assert(i7 != NULL);
78 if XNN_UNPREDICTABLE(i7 != zero) {
79 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
80 }
81 const uint8_t* i8 = *input++;
82 assert(i8 != NULL);
83 if XNN_UNPREDICTABLE(i8 != zero) {
84 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
85 }
86
87 int32_t* b = buffer;
88 for (size_t c = 0; c < channels; c += 8) {
89 const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8;
90 const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8;
91 const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8;
92 const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8;
93 const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8;
94 const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8;
95 const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8;
96 const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8;
97 const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8); i8 += 8;
98
99 const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);
100 const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);
101 const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);
102 const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);
103 const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);
104 const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);
105 const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);
106 const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);
107 const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);
108
109 const __m128i vsum018 = _mm_add_epi16(_mm_add_epi16(vxi0, vxi1), vxi8);
110 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3);
111 const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5);
112 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7);
113
114 const __m128i vsum2345 = _mm_add_epi16(vsum23, vsum45);
115 const __m128i vsum01678 = _mm_add_epi16(vsum018, vsum67);
116 const __m128i vsum = _mm_add_epi16(vsum2345, vsum01678);
117
118 const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero));
119 const __m128i vacc_hi = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vsum, vzero));
120
121 _mm_store_si128((__m128i*) b, vacc_lo);
122 _mm_store_si128((__m128i*) b + 1, vacc_hi);
123 b += 8;
124 }
125 }
126
127 size_t k = kernel_elements;
128 for (k -= 9; k > 8; k -= 8) {
129 const uint8_t* i0 = *input++;
130 assert(i0 != NULL);
131 if XNN_UNPREDICTABLE(i0 != zero) {
132 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
133 }
134 const uint8_t* i1 = *input++;
135 assert(i1 != NULL);
136 if XNN_UNPREDICTABLE(i1 != zero) {
137 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
138 }
139 const uint8_t* i2 = *input++;
140 assert(i2 != NULL);
141 if XNN_UNPREDICTABLE(i2 != zero) {
142 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
143 }
144 const uint8_t* i3 = *input++;
145 assert(i3 != NULL);
146 if XNN_UNPREDICTABLE(i3 != zero) {
147 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
148 }
149 const uint8_t* i4 = *input++;
150 assert(i4 != NULL);
151 if XNN_UNPREDICTABLE(i4 != zero) {
152 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
153 }
154 const uint8_t* i5 = *input++;
155 assert(i5 != NULL);
156 if XNN_UNPREDICTABLE(i5 != zero) {
157 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
158 }
159 const uint8_t* i6 = *input++;
160 assert(i6 != NULL);
161 if XNN_UNPREDICTABLE(i6 != zero) {
162 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
163 }
164 const uint8_t* i7 = *input++;
165 assert(i7 != NULL);
166 if XNN_UNPREDICTABLE(i7 != zero) {
167 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
168 }
169
170 int32_t* b = buffer;
171 for (size_t c = 0; c < channels; c += 8) {
172 const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8;
173 const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8;
174 const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8;
175 const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8;
176 const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8;
177 const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8;
178 const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8;
179 const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8;
180 __m128i vacc_lo = _mm_load_si128((const __m128i*) b);
181 __m128i vacc_hi = _mm_load_si128((const __m128i*) b + 1);
182
183 const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);
184 const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);
185 const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);
186 const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);
187 const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);
188 const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);
189 const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);
190 const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);
191
192 const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1);
193 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3);
194 const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5);
195 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7);
196
197 const __m128i vsum0123 = _mm_add_epi16(vsum01, vsum23);
198 const __m128i vsum4567 = _mm_add_epi16(vsum45, vsum67);
199 const __m128i vsum = _mm_add_epi16(vsum0123, vsum4567);
200
201 vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vsum, vzero));
202 vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vsum, vzero));
203
204 _mm_store_si128((__m128i*) b, vacc_lo);
205 _mm_store_si128((__m128i*) b + 1, vacc_hi);
206 b += 8;
207 }
208 }
209
210 {
211 const uint8_t* i0 = input[0];
212 assert(i0 != NULL);
213 const uint8_t* i1 = input[1];
214 const uint8_t* i2 = input[2];
215 const uint8_t* i3 = input[3];
216 const uint8_t* i4 = input[4];
217 const uint8_t* i5 = input[5];
218 const uint8_t* i6 = input[6];
219 const uint8_t* i7 = input[7];
220 input = (const uint8_t**) ((uintptr_t) input + input_increment);
221 if (k < 2) {
222 i1 = zero;
223 }
224 assert(i1 != NULL);
225 if (k <= 2) {
226 i2 = zero;
227 }
228 assert(i2 != NULL);
229 if (k < 4) {
230 i3 = zero;
231 }
232 assert(i3 != NULL);
233 if (k <= 4) {
234 i4 = zero;
235 }
236 assert(i4 != NULL);
237 if (k < 6) {
238 i5 = zero;
239 }
240 assert(i5 != NULL);
241 if (k <= 6) {
242 i6 = zero;
243 }
244 assert(i6 != NULL);
245 if (k < 8) {
246 i7 = zero;
247 }
248 assert(i7 != NULL);
249 if XNN_UNPREDICTABLE(i0 != zero) {
250 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
251 }
252 if XNN_UNPREDICTABLE(i1 != zero) {
253 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
254 }
255 if XNN_UNPREDICTABLE(i2 != zero) {
256 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
257 }
258 if XNN_UNPREDICTABLE(i3 != zero) {
259 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
260 }
261 if XNN_UNPREDICTABLE(i4 != zero) {
262 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
263 }
264 if XNN_UNPREDICTABLE(i5 != zero) {
265 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
266 }
267 if XNN_UNPREDICTABLE(i6 != zero) {
268 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
269 }
270 if XNN_UNPREDICTABLE(i7 != zero) {
271 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
272 }
273
274 size_t c = channels;
275 int32_t* b = buffer;
276 while (c >= 8) {
277 const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8;
278 const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8;
279 const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8;
280 const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8;
281 const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8;
282 const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8;
283 const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8;
284 const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8;
285 __m128i vacc_lo = _mm_load_si128((const __m128i*) b);
286 __m128i vacc_hi = _mm_load_si128((const __m128i*) b + 1);
287 b += 8;
288
289 const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);
290 const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);
291 const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);
292 const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);
293 const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);
294 const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);
295 const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);
296 const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);
297
298 const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1);
299 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3);
300 const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5);
301 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7);
302
303 const __m128i vsum0123 = _mm_add_epi16(vsum01, vsum23);
304 const __m128i vsum4567 = _mm_add_epi16(vsum45, vsum67);
305 const __m128i vsum = _mm_add_epi16(vsum0123, vsum4567);
306
307 vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vsum, vzero));
308 vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vsum, vzero));
309
310 const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo);
311 const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi);
312
313 const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo);
314 const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi);
315
316 const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1));
317 const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1));
318
319 const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier);
320 const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier);
321
322 const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier);
323 const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier);
324
325 const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift);
326 const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift);
327 const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift);
328 const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift);
329
330 const __m128i vabs_scaled_lo0213 = _mm_castps_si128(
331 _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0)));
332 const __m128i vabs_scaled_hi0213 = _mm_castps_si128(
333 _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0)));
334
335 const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0));
336 const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0));
337
338 const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo);
339 const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi);
340
341 __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi);
342 vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_zero_point));
343 vout = _mm_packus_epi16(vout, vout);
344 vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_max));
345 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_min));
346
347 _mm_storel_epi64((__m128i*) output, vout);
348 output += 8;
349
350 c -= 8;
351 }
352 if (c != 0) {
353 const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0);
354 const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1);
355 const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2);
356 const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3);
357 const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4);
358 const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5);
359 const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6);
360 const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7);
361 __m128i vacc_lo = _mm_load_si128((const __m128i*) b);
362 __m128i vacc_hi = _mm_load_si128((const __m128i*) b + 1);
363
364 const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);
365 const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);
366 const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);
367 const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);
368 const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);
369 const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);
370 const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);
371 const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);
372
373 const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1);
374 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3);
375 const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5);
376 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7);
377
378 const __m128i vsum0123 = _mm_add_epi16(vsum01, vsum23);
379 const __m128i vsum4567 = _mm_add_epi16(vsum45, vsum67);
380 const __m128i vsum = _mm_add_epi16(vsum0123, vsum4567);
381
382 vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vsum, vzero));
383 vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vsum, vzero));
384
385 const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo);
386 const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi);
387
388 const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo);
389 const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi);
390
391 const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1));
392 const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1));
393
394 const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier);
395 const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier);
396
397 const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier);
398 const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier);
399
400 const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift);
401 const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift);
402 const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift);
403 const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift);
404
405 const __m128i vabs_scaled_lo0213 = _mm_castps_si128(
406 _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0)));
407 const __m128i vabs_scaled_hi0213 = _mm_castps_si128(
408 _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0)));
409
410 const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0));
411 const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0));
412
413 const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo);
414 const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi);
415
416 __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi);
417 vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_zero_point));
418 vout = _mm_packus_epi16(vout, vout);
419 vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_max));
420 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_min));
421
422 if (c & 4) {
423 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout);
424 output += 4;
425 vout = _mm_srli_epi64(vout, 32);
426 }
427 if (c & 2) {
428 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout, 0);
429 output += 2;
430 vout = _mm_srli_epi32(vout, 16);
431 }
432 if (c & 1) {
433 *((uint8_t*) output) = (uint8_t) _mm_cvtsi128_si32(vout);
434 output += 1;
435 }
436 }
437 }
438 output = (uint8_t*) ((uintptr_t) output + output_increment);
439 } while (--output_pixels != 0);
440 }
441