xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up16x9-minmax-fp32-avx-mul16.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 <smmintrin.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16 
17 
xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16(
19     size_t channels,
20     size_t output_width,
21     const uint8_t** input,
22     const void* weights,
23     uint8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const uint8_t* zero,
28     const union xnn_qu8_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 uint8_t* i0 = input[0];
35     assert(i0 != NULL);
36     if XNN_UNPREDICTABLE(i0 != zero) {
37       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
38     }
39     const uint8_t* i1 = input[1];
40     assert(i1 != NULL);
41     if XNN_UNPREDICTABLE(i1 != zero) {
42       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
43     }
44     const uint8_t* i2 = input[2];
45     assert(i2 != NULL);
46     if XNN_UNPREDICTABLE(i2 != zero) {
47       i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
48     }
49     const uint8_t* i3 = input[3];
50     assert(i3 != NULL);
51     if XNN_UNPREDICTABLE(i3 != zero) {
52       i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
53     }
54     const uint8_t* i4 = input[4];
55     assert(i4 != NULL);
56     if XNN_UNPREDICTABLE(i4 != zero) {
57       i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
58     }
59     const uint8_t* i5 = input[5];
60     assert(i5 != NULL);
61     if XNN_UNPREDICTABLE(i5 != zero) {
62       i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
63     }
64     const uint8_t* i6 = input[6];
65     assert(i6 != NULL);
66     if XNN_UNPREDICTABLE(i6 != zero) {
67       i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
68     }
69     const uint8_t* i7 = input[7];
70     assert(i7 != NULL);
71     if XNN_UNPREDICTABLE(i7 != zero) {
72       i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
73     }
74     const uint8_t* i8 = input[8];
75     assert(i8 != NULL);
76     if XNN_UNPREDICTABLE(i8 != zero) {
77       i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
78     }
79     input = (const uint8_t**) ((uintptr_t) input + input_stride);
80 
81     size_t c = channels;
82     const void* w = weights;
83     const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
84     for (; c >= 16; c -= 16) {
85       __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
86       __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
87       __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
88       __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
89 
90 
91       const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
92       const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
93       const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
94       const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
95       const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
96       const __m128i vxi0x89ABCDEF = _mm_cvtepu8_epi16(vi0x89ABCDEF);
97       const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
98       const __m128i vxk0x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x89ABCDEF), vk_zero_point);
99       i0 += 16;
100 
101 
102       const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
103       const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
104       const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
105       const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
106 
107       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
108       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
109       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
110       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
111 
112       const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
113       const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
114       const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
115       const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
116       const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
117       const __m128i vxi1x89ABCDEF = _mm_cvtepu8_epi16(vi1x89ABCDEF);
118       const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
119       const __m128i vxk1x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x89ABCDEF), vk_zero_point);
120       i1 += 16;
121 
122 
123       const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
124       const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
125       const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
126       const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
127 
128       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
129       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
130       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
131       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
132 
133       const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
134       const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
135       const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
136       const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
137       const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
138       const __m128i vxi2x89ABCDEF = _mm_cvtepu8_epi16(vi2x89ABCDEF);
139       const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
140       const __m128i vxk2x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x89ABCDEF), vk_zero_point);
141       i2 += 16;
142 
143 
144       const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
145       const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
146       const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
147       const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
148 
149       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
150       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
151       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
152       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
153 
154       const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
155       const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
156       const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
157       const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
158       const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
159       const __m128i vxi3x89ABCDEF = _mm_cvtepu8_epi16(vi3x89ABCDEF);
160       const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
161       const __m128i vxk3x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x89ABCDEF), vk_zero_point);
162       i3 += 16;
163 
164 
165       const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
166       const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
167       const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
168       const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
169 
170       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
171       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
172       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
173       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
174 
175       const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
176       const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
177       const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
178       const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
179       const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
180       const __m128i vxi4x89ABCDEF = _mm_cvtepu8_epi16(vi4x89ABCDEF);
181       const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
182       const __m128i vxk4x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x89ABCDEF), vk_zero_point);
183       i4 += 16;
184 
185 
186       const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
187       const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
188       const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
189       const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
190 
191       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
192       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
193       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
194       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
195 
196       const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
197       const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
198       const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
199       const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
200       const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
201       const __m128i vxi5x89ABCDEF = _mm_cvtepu8_epi16(vi5x89ABCDEF);
202       const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
203       const __m128i vxk5x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x89ABCDEF), vk_zero_point);
204       i5 += 16;
205 
206 
207       const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
208       const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
209       const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
210       const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
211 
212       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
213       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
214       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
215       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
216 
217       const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
218       const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
219       const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
220       const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
221       const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
222       const __m128i vxi6x89ABCDEF = _mm_cvtepu8_epi16(vi6x89ABCDEF);
223       const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
224       const __m128i vxk6x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x89ABCDEF), vk_zero_point);
225       i6 += 16;
226 
227 
228       const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
229       const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
230       const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
231       const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
232 
233       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
234       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
235       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
236       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
237 
238       const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
239       const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
240       const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
241       const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
242       const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
243       const __m128i vxi7x89ABCDEF = _mm_cvtepu8_epi16(vi7x89ABCDEF);
244       const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
245       const __m128i vxk7x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x89ABCDEF), vk_zero_point);
246       i7 += 16;
247 
248 
249       const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
250       const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
251       const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
252       const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
253 
254       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
255       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
256       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
257       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
258 
259       const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
260       const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
261       const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
262       const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
263       const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
264       const __m128i vxi8x89ABCDEF = _mm_cvtepu8_epi16(vi8x89ABCDEF);
265       const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
266       const __m128i vxk8x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x89ABCDEF), vk_zero_point);
267       i8 += 16;
268 
269 
270       const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
271       const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
272       const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
273       const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
274 
275       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
276       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
277       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
278       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
279 
280       w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t));
281 
282       __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
283       __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
284       __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
285       __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
286 
287       const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
288       vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
289       vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
290       vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
291       vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
292 
293       const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
294       vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
295       vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
296       vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
297       vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
298 
299       vacc0123 = _mm_cvtps_epi32(vscaled0123);
300       vacc4567 = _mm_cvtps_epi32(vscaled4567);
301       vacc89AB = _mm_cvtps_epi32(vscaled89AB);
302       vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
303 
304       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
305       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
306       __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
307 
308       __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
309 
310       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
311       vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
312 
313       _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
314       output += 16;
315     }
316     if XNN_UNLIKELY(c != 0) {
317       const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
318       do {
319         __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
320         __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
321 
322 
323         const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
324         const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
325         const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
326         const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
327         i0 += 8;
328 
329 
330         const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
331         const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
332 
333         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
334         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
335 
336         const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
337         const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
338         const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
339         const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
340         i1 += 8;
341 
342 
343         const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
344         const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
345 
346         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
347         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
348 
349         const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
350         const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
351         const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
352         const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
353         i2 += 8;
354 
355 
356         const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
357         const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
358 
359         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
360         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
361 
362         const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
363         const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
364         const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
365         const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
366         i3 += 8;
367 
368 
369         const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
370         const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
371 
372         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
373         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
374 
375         const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
376         const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
377         const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
378         const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
379         i4 += 8;
380 
381 
382         const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
383         const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
384 
385         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
386         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
387 
388         const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
389         const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
390         const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
391         const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
392         i5 += 8;
393 
394 
395         const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
396         const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
397 
398         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
399         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
400 
401         const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
402         const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
403         const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
404         const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
405         i6 += 8;
406 
407 
408         const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
409         const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
410 
411         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
412         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
413 
414         const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
415         const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
416         const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
417         const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
418         i7 += 8;
419 
420 
421         const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
422         const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
423 
424         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
425         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
426 
427         const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
428         const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
429         const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
430         const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
431         i8 += 8;
432 
433 
434         const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
435         const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
436 
437         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
438         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
439 
440         k += 8;
441 
442         __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
443         __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
444 
445         const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
446         vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
447         vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
448 
449         const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
450         vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
451         vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
452 
453         vacc0123 = _mm_cvtps_epi32(vscaled0123);
454         vacc4567 = _mm_cvtps_epi32(vscaled4567);
455 
456         w = (const void*) ((const int32_t*) w + 8);
457 
458         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
459         __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
460 
461         __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
462 
463         vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
464 
465         if XNN_LIKELY(c >= 8) {
466           _mm_storel_epi64((__m128i*) output, vout0123456701234567);
467           output += 8;
468           c -= 8;
469         } else {
470           if (c & 4) {
471             unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
472             vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
473             output += 4;
474           }
475           if (c & 2) {
476             unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
477             vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
478             output += 2;
479           }
480           if (c & 1) {
481             *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
482             output += 1;
483           }
484           c = 0;
485         }
486       } while (c != 0);
487     }
488 
489     output = (uint8_t*) ((uintptr_t) output + output_increment);
490   } while (--output_width != 0);
491 }
492