xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up8x25-minmax-fp32-sse2-mul16-add16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <emmintrin.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16 
17 
xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16(
19     size_t channels,
20     size_t output_width,
21     const int8_t** input,
22     const void* weights,
23     int8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const int8_t* zero,
28     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   do {
34     const int8_t* i0 = input[0];
35     assert(i0 != NULL);
36     if XNN_UNPREDICTABLE(i0 != zero) {
37       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38     }
39     const int8_t* i1 = input[1];
40     assert(i1 != NULL);
41     if XNN_UNPREDICTABLE(i1 != zero) {
42       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43     }
44     const int8_t* i2 = input[2];
45     assert(i2 != NULL);
46     if XNN_UNPREDICTABLE(i2 != zero) {
47       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48     }
49     const int8_t* i3 = input[3];
50     assert(i3 != NULL);
51     if XNN_UNPREDICTABLE(i3 != zero) {
52       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53     }
54     const int8_t* i4 = input[4];
55     assert(i4 != NULL);
56     if XNN_UNPREDICTABLE(i4 != zero) {
57       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58     }
59     const int8_t* i5 = input[5];
60     assert(i5 != NULL);
61     if XNN_UNPREDICTABLE(i5 != zero) {
62       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63     }
64     const int8_t* i6 = input[6];
65     assert(i6 != NULL);
66     if XNN_UNPREDICTABLE(i6 != zero) {
67       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68     }
69     const int8_t* i7 = input[7];
70     assert(i7 != NULL);
71     if XNN_UNPREDICTABLE(i7 != zero) {
72       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73     }
74     const int8_t* i8 = input[8];
75     assert(i8 != NULL);
76     if XNN_UNPREDICTABLE(i8 != zero) {
77       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78     }
79     const int8_t* i9 = input[9];
80     assert(i9 != NULL);
81     if XNN_UNPREDICTABLE(i9 != zero) {
82       i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
83     }
84     const int8_t* i10 = input[10];
85     assert(i10 != NULL);
86     if XNN_UNPREDICTABLE(i10 != zero) {
87       i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
88     }
89     const int8_t* i11 = input[11];
90     assert(i11 != NULL);
91     if XNN_UNPREDICTABLE(i11 != zero) {
92       i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
93     }
94     const int8_t* i12 = input[12];
95     assert(i12 != NULL);
96     if XNN_UNPREDICTABLE(i12 != zero) {
97       i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
98     }
99     const int8_t* i13 = input[13];
100     assert(i13 != NULL);
101     if XNN_UNPREDICTABLE(i13 != zero) {
102       i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
103     }
104     const int8_t* i14 = input[14];
105     assert(i14 != NULL);
106     if XNN_UNPREDICTABLE(i14 != zero) {
107       i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
108     }
109     const int8_t* i15 = input[15];
110     assert(i15 != NULL);
111     if XNN_UNPREDICTABLE(i15 != zero) {
112       i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
113     }
114     const int8_t* i16 = input[16];
115     assert(i16 != NULL);
116     if XNN_UNPREDICTABLE(i16 != zero) {
117       i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
118     }
119     const int8_t* i17 = input[17];
120     assert(i17 != NULL);
121     if XNN_UNPREDICTABLE(i17 != zero) {
122       i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
123     }
124     const int8_t* i18 = input[18];
125     assert(i18 != NULL);
126     if XNN_UNPREDICTABLE(i18 != zero) {
127       i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
128     }
129     const int8_t* i19 = input[19];
130     assert(i19 != NULL);
131     if XNN_UNPREDICTABLE(i19 != zero) {
132       i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
133     }
134     const int8_t* i20 = input[20];
135     assert(i20 != NULL);
136     if XNN_UNPREDICTABLE(i20 != zero) {
137       i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
138     }
139     const int8_t* i21 = input[21];
140     assert(i21 != NULL);
141     if XNN_UNPREDICTABLE(i21 != zero) {
142       i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
143     }
144     const int8_t* i22 = input[22];
145     assert(i22 != NULL);
146     if XNN_UNPREDICTABLE(i22 != zero) {
147       i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
148     }
149     const int8_t* i23 = input[23];
150     assert(i23 != NULL);
151     if XNN_UNPREDICTABLE(i23 != zero) {
152       i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
153     }
154     const int8_t* i24 = input[24];
155     assert(i24 != NULL);
156     if XNN_UNPREDICTABLE(i24 != zero) {
157       i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
158     }
159     input = (const int8_t**) ((uintptr_t) input + input_stride);
160 
161     size_t c = channels;
162     const void* w = weights;
163     for (; c >= 8; c -= 8) {
164       __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
165       __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
166 
167 
168       const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
169       const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
170       i0 += 8;
171 
172       const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
173       const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
174 
175       __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
176 
177 
178       const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
179       const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
180       i1 += 8;
181 
182       const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
183       const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
184 
185       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
186 
187       const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
188       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
189       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
190 
191       const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
192       const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
193       i2 += 8;
194 
195       const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
196       const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
197 
198       vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
199 
200 
201       const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
202       const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
203       i3 += 8;
204 
205       const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
206       const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
207 
208       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
209 
210       const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
211       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
212       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
213 
214       const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
215       const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
216       i4 += 8;
217 
218       const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
219       const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
220 
221       vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
222 
223 
224       const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
225       const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
226       i5 += 8;
227 
228       const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
229       const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
230 
231       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
232 
233       const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
234       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
235       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
236 
237       const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
238       const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
239       i6 += 8;
240 
241       const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
242       const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
243 
244       vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
245 
246 
247       const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
248       const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
249       i7 += 8;
250 
251       const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
252       const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
253 
254       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
255 
256       const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
257       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
258       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
259 
260       const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
261       const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
262       i8 += 8;
263 
264       const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
265       const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
266 
267       vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
268 
269 
270       const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
271       const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
272       i9 += 8;
273 
274       const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
275       const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
276 
277       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
278 
279       const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
280       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
281       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
282 
283       const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
284       const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
285       i10 += 8;
286 
287       const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
288       const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
289 
290       vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
291 
292 
293       const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
294       const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
295       i11 += 8;
296 
297       const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
298       const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
299 
300       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
301 
302       const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
303       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
304       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
305 
306       const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
307       const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
308       i12 += 8;
309 
310       const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
311       const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
312 
313       vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
314 
315 
316       const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
317       const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
318       i13 += 8;
319 
320       const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
321       const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
322 
323       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
324 
325       const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
326       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
327       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
328 
329       const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
330       const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
331       i14 += 8;
332 
333       const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
334       const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
335 
336       vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
337 
338 
339       const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
340       const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
341       i15 += 8;
342 
343       const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
344       const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
345 
346       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
347 
348       const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
349       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
350       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
351 
352       const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
353       const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
354       i16 += 8;
355 
356       const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
357       const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
358 
359       vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
360 
361 
362       const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
363       const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
364       i17 += 8;
365 
366       const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
367       const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
368 
369       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
370 
371       const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
372       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
373       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
374 
375       const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
376       const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
377       i18 += 8;
378 
379       const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
380       const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
381 
382       vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
383 
384 
385       const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
386       const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
387       i19 += 8;
388 
389       const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
390       const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
391 
392       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
393 
394       const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
395       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
396       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
397 
398       const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
399       const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
400       i20 += 8;
401 
402       const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
403       const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
404 
405       vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
406 
407 
408       const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
409       const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
410       i21 += 8;
411 
412       const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
413       const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
414 
415       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
416 
417       const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
418       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
419       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
420 
421       const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
422       const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
423       i22 += 8;
424 
425       const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
426       const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
427 
428       vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
429 
430 
431       const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
432       const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
433       i23 += 8;
434 
435       const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
436       const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
437 
438       vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
439 
440       const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
441       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
442       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
443 
444       const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
445       const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
446       i24 += 8;
447 
448       const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
449       const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
450 
451       vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
452 
453       const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
454       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
455       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
456 
457       w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
458 
459       __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
460       __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
461 
462       const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
463       const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
464       w = (const void*) ((const float*) w + 8);
465       vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
466       vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
467 
468       const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
469       vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
470       vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
471 
472       vacc0123 = _mm_cvtps_epi32(vscaled0123);
473       vacc4567 = _mm_cvtps_epi32(vscaled4567);
474 
475       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
476       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
477 
478       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
479       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
480 
481       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
482 
483 
484       _mm_storel_epi64((__m128i*) output, vout0123456701234567);
485       output += 8;
486     }
487     if XNN_UNLIKELY(c != 0) {
488       {
489         __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
490         __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
491 
492 
493         const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
494         const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
495 
496         const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
497         const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
498 
499         __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
500 
501 
502         const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
503         const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
504 
505         const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
506         const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
507 
508         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
509 
510         const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
511         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
512         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
513 
514         const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
515         const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
516 
517         const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
518         const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
519 
520         vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
521 
522 
523         const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
524         const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
525 
526         const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
527         const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
528 
529         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
530 
531         const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
532         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
533         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
534 
535         const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
536         const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
537 
538         const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
539         const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
540 
541         vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
542 
543 
544         const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
545         const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
546 
547         const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
548         const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
549 
550         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
551 
552         const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
553         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
554         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
555 
556         const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
557         const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
558 
559         const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
560         const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
561 
562         vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
563 
564 
565         const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
566         const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
567 
568         const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
569         const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
570 
571         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
572 
573         const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
574         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
575         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
576 
577         const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
578         const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
579 
580         const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
581         const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
582 
583         vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
584 
585 
586         const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9);
587         const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
588 
589         const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8);
590         const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8);
591 
592         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567));
593 
594         const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
595         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567));
596         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567));
597 
598         const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10);
599         const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)));
600 
601         const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8);
602         const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8);
603 
604         vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567);
605 
606 
607         const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11);
608         const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)));
609 
610         const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8);
611         const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8);
612 
613         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567));
614 
615         const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
616         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567));
617         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567));
618 
619         const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12);
620         const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)));
621 
622         const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8);
623         const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8);
624 
625         vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567);
626 
627 
628         const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13);
629         const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)));
630 
631         const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8);
632         const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8);
633 
634         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567));
635 
636         const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
637         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567));
638         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567));
639 
640         const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14);
641         const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)));
642 
643         const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8);
644         const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8);
645 
646         vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567);
647 
648 
649         const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15);
650         const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)));
651 
652         const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8);
653         const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8);
654 
655         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567));
656 
657         const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
658         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567));
659         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567));
660 
661         const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16);
662         const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)));
663 
664         const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8);
665         const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8);
666 
667         vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567);
668 
669 
670         const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17);
671         const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)));
672 
673         const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8);
674         const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8);
675 
676         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567));
677 
678         const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
679         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567));
680         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567));
681 
682         const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18);
683         const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)));
684 
685         const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8);
686         const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8);
687 
688         vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567);
689 
690 
691         const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19);
692         const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)));
693 
694         const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8);
695         const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8);
696 
697         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567));
698 
699         const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
700         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567));
701         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567));
702 
703         const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20);
704         const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)));
705 
706         const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8);
707         const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8);
708 
709         vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567);
710 
711 
712         const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21);
713         const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)));
714 
715         const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8);
716         const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8);
717 
718         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567));
719 
720         const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
721         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567));
722         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567));
723 
724         const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22);
725         const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)));
726 
727         const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8);
728         const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8);
729 
730         vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567);
731 
732 
733         const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23);
734         const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)));
735 
736         const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8);
737         const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8);
738 
739         vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567));
740 
741         const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
742         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567));
743         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567));
744 
745         const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24);
746         const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)));
747 
748         const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8);
749         const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8);
750 
751         vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567);
752 
753         const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
754         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567));
755         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567));
756 
757 
758         __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
759         __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
760 
761         const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t)));
762         const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t) + 4 * sizeof(float)));
763         vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
764         vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
765 
766         const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
767         vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
768         vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
769 
770         vacc0123 = _mm_cvtps_epi32(vscaled0123);
771         vacc4567 = _mm_cvtps_epi32(vscaled4567);
772 
773 
774         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
775         __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
776 
777         vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
778 
779         __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
780 
781 
782         if (c & 4) {
783           unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
784           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
785           output += 4;
786         }
787         if (c & 2) {
788           unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
789           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
790           output += 2;
791         }
792         if (c & 1) {
793           *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
794           output += 1;
795         }
796       }
797     }
798 
799     output = (int8_t*) ((uintptr_t) output + output_increment);
800   } while (--output_width != 0);
801 }
802