xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up32x9-minmax-fp32-avx2-mul16-vpunpck.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-avx2-mul16-vpunpck.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 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 <immintrin.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16 
17 
xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpunpck(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpunpck(
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_qs8_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     input = (const int8_t**) ((uintptr_t) input + input_stride);
80 
81     size_t c = channels;
82     const void* w = weights;
83     for (; c >= 32; c -= 32) {
84       __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
85       __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
86       __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int32_t)));
87       __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 24 * sizeof(int32_t)));
88 
89       __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
90       __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
91       __m256i vaccGHIJOPQR = _mm256_inserti128_si256(vaccGHIJKLMN, _mm256_castsi256_si128(vaccOPQRSTUV), 1);
92       __m256i vaccKLMNSTUV = _mm256_permute2x128_si256(vaccGHIJKLMN, vaccOPQRSTUV, 0x31);
93 
94 
95       const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
96       const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
97       const __m256i vi0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i0 + 16)));
98       const __m256i vk0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
99       i0 += 32;
100 
101       const __m256i vprod0x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
102       const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
103       const __m256i vprod0xGHIJKLMNOPQRSTUVlo =  _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV);
104       const __m256i vprod0xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod0xGHIJKLMNOPQRSTUVlo, 15);
105 
106       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
107       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
108       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod0xGHIJKLMNOPQRSTUVlo, vprod0xGHIJKLMNOPQRSTUVhi));
109       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod0xGHIJKLMNOPQRSTUVlo, vprod0xGHIJKLMNOPQRSTUVhi));
110 
111       const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
112       const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
113       const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16)));
114       const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
115       i1 += 32;
116 
117       const __m256i vprod1x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
118       const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
119       const __m256i vprod1xGHIJKLMNOPQRSTUVlo =  _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV);
120       const __m256i vprod1xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod1xGHIJKLMNOPQRSTUVlo, 15);
121 
122       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
123       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
124       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod1xGHIJKLMNOPQRSTUVlo, vprod1xGHIJKLMNOPQRSTUVhi));
125       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod1xGHIJKLMNOPQRSTUVlo, vprod1xGHIJKLMNOPQRSTUVhi));
126 
127       const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
128       const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
129       const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16)));
130       const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
131       i2 += 32;
132 
133       const __m256i vprod2x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
134       const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
135       const __m256i vprod2xGHIJKLMNOPQRSTUVlo =  _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV);
136       const __m256i vprod2xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod2xGHIJKLMNOPQRSTUVlo, 15);
137 
138       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
139       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
140       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod2xGHIJKLMNOPQRSTUVlo, vprod2xGHIJKLMNOPQRSTUVhi));
141       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod2xGHIJKLMNOPQRSTUVlo, vprod2xGHIJKLMNOPQRSTUVhi));
142 
143       const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
144       const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
145       const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16)));
146       const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
147       i3 += 32;
148 
149       const __m256i vprod3x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
150       const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
151       const __m256i vprod3xGHIJKLMNOPQRSTUVlo =  _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV);
152       const __m256i vprod3xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod3xGHIJKLMNOPQRSTUVlo, 15);
153 
154       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
155       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
156       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod3xGHIJKLMNOPQRSTUVlo, vprod3xGHIJKLMNOPQRSTUVhi));
157       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod3xGHIJKLMNOPQRSTUVlo, vprod3xGHIJKLMNOPQRSTUVhi));
158 
159       const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
160       const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
161       const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16)));
162       const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
163       i4 += 32;
164 
165       const __m256i vprod4x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
166       const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
167       const __m256i vprod4xGHIJKLMNOPQRSTUVlo =  _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV);
168       const __m256i vprod4xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod4xGHIJKLMNOPQRSTUVlo, 15);
169 
170       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
171       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
172       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod4xGHIJKLMNOPQRSTUVlo, vprod4xGHIJKLMNOPQRSTUVhi));
173       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod4xGHIJKLMNOPQRSTUVlo, vprod4xGHIJKLMNOPQRSTUVhi));
174 
175       const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
176       const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
177       const __m256i vi5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i5 + 16)));
178       const __m256i vk5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
179       i5 += 32;
180 
181       const __m256i vprod5x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
182       const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
183       const __m256i vprod5xGHIJKLMNOPQRSTUVlo =  _mm256_mullo_epi16(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV);
184       const __m256i vprod5xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod5xGHIJKLMNOPQRSTUVlo, 15);
185 
186       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
187       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
188       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod5xGHIJKLMNOPQRSTUVlo, vprod5xGHIJKLMNOPQRSTUVhi));
189       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod5xGHIJKLMNOPQRSTUVlo, vprod5xGHIJKLMNOPQRSTUVhi));
190 
191       const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
192       const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
193       const __m256i vi6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i6 + 16)));
194       const __m256i vk6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
195       i6 += 32;
196 
197       const __m256i vprod6x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
198       const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
199       const __m256i vprod6xGHIJKLMNOPQRSTUVlo =  _mm256_mullo_epi16(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV);
200       const __m256i vprod6xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod6xGHIJKLMNOPQRSTUVlo, 15);
201 
202       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
203       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
204       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod6xGHIJKLMNOPQRSTUVlo, vprod6xGHIJKLMNOPQRSTUVhi));
205       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod6xGHIJKLMNOPQRSTUVlo, vprod6xGHIJKLMNOPQRSTUVhi));
206 
207       const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
208       const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
209       const __m256i vi7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i7 + 16)));
210       const __m256i vk7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
211       i7 += 32;
212 
213       const __m256i vprod7x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
214       const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
215       const __m256i vprod7xGHIJKLMNOPQRSTUVlo =  _mm256_mullo_epi16(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV);
216       const __m256i vprod7xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod7xGHIJKLMNOPQRSTUVlo, 15);
217 
218       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
219       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
220       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod7xGHIJKLMNOPQRSTUVlo, vprod7xGHIJKLMNOPQRSTUVhi));
221       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod7xGHIJKLMNOPQRSTUVlo, vprod7xGHIJKLMNOPQRSTUVhi));
222 
223       const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
224       const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
225       const __m256i vi8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i8 + 16)));
226       const __m256i vk8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
227       i8 += 32;
228 
229       const __m256i vprod8x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
230       const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
231       const __m256i vprod8xGHIJKLMNOPQRSTUVlo =  _mm256_mullo_epi16(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV);
232       const __m256i vprod8xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod8xGHIJKLMNOPQRSTUVlo, 15);
233 
234       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
235       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
236       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod8xGHIJKLMNOPQRSTUVlo, vprod8xGHIJKLMNOPQRSTUVhi));
237       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod8xGHIJKLMNOPQRSTUVlo, vprod8xGHIJKLMNOPQRSTUVhi));
238 
239       w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t));
240 
241       vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
242       vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
243       vaccGHIJKLMN = _mm256_inserti128_si256(vaccGHIJOPQR, _mm256_castsi256_si128(vaccKLMNSTUV), 1);
244       vaccOPQRSTUV = _mm256_permute2x128_si256(vaccGHIJOPQR, vaccKLMNSTUV, 0x31);
245 
246       __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
247       __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
248       __m256 vfpaccGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
249       __m256 vfpaccOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
250 
251       const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
252       vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
253       vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
254       vfpaccGHIJKLMN = _mm256_mul_ps(vfpaccGHIJKLMN, vscale);
255       vfpaccOPQRSTUV = _mm256_mul_ps(vfpaccOPQRSTUV, vscale);
256 
257       const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
258       vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
259       vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
260       vfpaccGHIJKLMN = _mm256_min_ps(vfpaccGHIJKLMN, voutput_max_less_zero_point);
261       vfpaccOPQRSTUV = _mm256_min_ps(vfpaccOPQRSTUV, voutput_max_less_zero_point);
262 
263       vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
264       vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
265       vaccGHIJKLMN = _mm256_cvtps_epi32(vfpaccGHIJKLMN);
266       vaccOPQRSTUV = _mm256_cvtps_epi32(vfpaccOPQRSTUV);
267 
268       const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
269       const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
270       const __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
271 
272       __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
273       __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
274 
275       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
276       vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
277       voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
278 
279       _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
280       _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
281       output += 32;
282     }
283     if XNN_UNLIKELY(c != 0) {
284       const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t));
285       do {
286         __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
287         __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
288 
289         __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
290         __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
291 
292 
293         const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
294         const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k));
295         i0 += 16;
296 
297         const __m256i vprod0x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
298         const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
299 
300         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
301         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
302 
303         const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
304         const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 32)));
305         i1 += 16;
306 
307         const __m256i vprod1x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
308         const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
309 
310         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
311         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
312 
313         const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
314         const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 64)));
315         i2 += 16;
316 
317         const __m256i vprod2x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
318         const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
319 
320         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
321         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
322 
323         const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
324         const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 96)));
325         i3 += 16;
326 
327         const __m256i vprod3x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
328         const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
329 
330         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
331         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
332 
333         const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
334         const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 128)));
335         i4 += 16;
336 
337         const __m256i vprod4x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
338         const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
339 
340         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
341         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
342 
343         const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
344         const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 160)));
345         i5 += 16;
346 
347         const __m256i vprod5x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
348         const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
349 
350         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
351         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
352 
353         const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
354         const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 192)));
355         i6 += 16;
356 
357         const __m256i vprod6x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
358         const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
359 
360         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
361         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
362 
363         const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
364         const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 224)));
365         i7 += 16;
366 
367         const __m256i vprod7x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
368         const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
369 
370         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
371         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
372 
373         const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
374         const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 256)));
375         i8 += 16;
376 
377         const __m256i vprod8x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
378         const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
379 
380         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
381         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
382 
383         vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
384         vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
385 
386         k += 16;
387 
388         __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
389         __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
390 
391         const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
392         vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale);
393         vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale);
394 
395         const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
396         vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
397         vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
398 
399         vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
400         vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
401 
402         w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
403 
404         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
405         __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
406         __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), voutput_zero_point);
407 
408         const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
409 
410         __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
411         vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
412 
413         if XNN_LIKELY(c >= 16) {
414           _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
415           output += 16;
416           c -= 16;
417         } else {
418           if (c & 8) {
419             _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
420             vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
421             output += 8;
422           }
423           if (c & 4) {
424             unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
425             vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
426             output += 4;
427           }
428           if (c & 2) {
429             unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
430             vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
431             output += 2;
432           }
433           if (c & 1) {
434             *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
435             output += 1;
436           }
437           c = 0;
438         }
439       } while (c != 0);
440     }
441 
442     output = (int8_t*) ((uintptr_t) output + output_increment);
443   } while (--output_width != 0);
444 }
445