xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up32x9-minmax-fp32-avx2-mul16-add16-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_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_add16_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_add16_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_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     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       __m256i vacc0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
102       __m256i vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV);
103 
104 
105       const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
106       const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
107       const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16)));
108       const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
109       i1 += 32;
110 
111       vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
112       vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV));
113 
114       __m256i vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
115       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
116       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
117       __m256i vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
118       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
119       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
120 
121       const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
122       const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
123       const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16)));
124       const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
125       i2 += 32;
126 
127       vacc0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
128       vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV);
129 
130 
131       const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
132       const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
133       const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16)));
134       const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
135       i3 += 32;
136 
137       vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
138       vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV));
139 
140       vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
141       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
142       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
143       vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
144       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
145       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
146 
147       const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
148       const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
149       const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16)));
150       const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
151       i4 += 32;
152 
153       vacc0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
154       vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV);
155 
156 
157       const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
158       const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
159       const __m256i vi5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i5 + 16)));
160       const __m256i vk5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
161       i5 += 32;
162 
163       vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
164       vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV));
165 
166       vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
167       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
168       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
169       vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
170       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
171       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
172 
173       const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
174       const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
175       const __m256i vi6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i6 + 16)));
176       const __m256i vk6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
177       i6 += 32;
178 
179       vacc0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
180       vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV);
181 
182 
183       const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
184       const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
185       const __m256i vi7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i7 + 16)));
186       const __m256i vk7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
187       i7 += 32;
188 
189       vacc0123456789ABCDEF = _mm256_add_epi16(vacc0123456789ABCDEF, _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
190       vaccGHIJKLMNOPQRSTUV = _mm256_add_epi16(vaccGHIJKLMNOPQRSTUV, _mm256_mullo_epi16(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV));
191 
192       vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
193       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
194       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
195       vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
196       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
197       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
198 
199       const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
200       const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
201       const __m256i vi8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i8 + 16)));
202       const __m256i vk8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
203       i8 += 32;
204 
205       vacc0123456789ABCDEF = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
206       vaccGHIJKLMNOPQRSTUV = _mm256_mullo_epi16(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV);
207 
208       vsignacc0123456789ABCDEF = _mm256_srai_epi16(vacc0123456789ABCDEF, 15);
209       vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
210       vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vacc0123456789ABCDEF, vsignacc0123456789ABCDEF));
211       vsignaccGHIJKLMNOPQRSTUV = _mm256_srai_epi16(vaccGHIJKLMNOPQRSTUV, 15);
212       vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
213       vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vaccGHIJKLMNOPQRSTUV, vsignaccGHIJKLMNOPQRSTUV));
214 
215       w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t));
216 
217       vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
218       vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
219       vaccGHIJKLMN = _mm256_inserti128_si256(vaccGHIJOPQR, _mm256_castsi256_si128(vaccKLMNSTUV), 1);
220       vaccOPQRSTUV = _mm256_permute2x128_si256(vaccGHIJOPQR, vaccKLMNSTUV, 0x31);
221 
222       __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
223       __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
224       __m256 vfpaccGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
225       __m256 vfpaccOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
226 
227       const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w);
228       const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float)));
229       const __m256 vscaleGHIJKLMN = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(float)));
230       const __m256 vscaleOPQRSTUV = _mm256_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(float)));
231       w = (const void*) ((uintptr_t) w + 32 * sizeof(float));
232       vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
233       vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
234       vfpaccGHIJKLMN = _mm256_mul_ps(vfpaccGHIJKLMN, vscaleGHIJKLMN);
235       vfpaccOPQRSTUV = _mm256_mul_ps(vfpaccOPQRSTUV, vscaleOPQRSTUV);
236 
237       const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
238       vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
239       vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
240       vfpaccGHIJKLMN = _mm256_min_ps(vfpaccGHIJKLMN, voutput_max_less_zero_point);
241       vfpaccOPQRSTUV = _mm256_min_ps(vfpaccOPQRSTUV, voutput_max_less_zero_point);
242 
243       vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
244       vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
245       vaccGHIJKLMN = _mm256_cvtps_epi32(vfpaccGHIJKLMN);
246       vaccOPQRSTUV = _mm256_cvtps_epi32(vfpaccOPQRSTUV);
247 
248       const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
249       const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
250       const __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
251 
252       __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
253       __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
254 
255       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
256       vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
257       voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
258 
259       _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
260       _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
261       output += 32;
262     }
263     if XNN_UNLIKELY(c != 0) {
264       const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t));
265       do {
266         __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
267         __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
268 
269         __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
270         __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
271 
272 
273         const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
274         const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k));
275         i0 += 16;
276 
277         const __m256i vprod0x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
278         const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
279 
280         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
281         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
282 
283         const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
284         const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 32)));
285         i1 += 16;
286 
287         const __m256i vprod1x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
288         const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
289 
290         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
291         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
292 
293         const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
294         const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 64)));
295         i2 += 16;
296 
297         const __m256i vprod2x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
298         const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
299 
300         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
301         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
302 
303         const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
304         const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 96)));
305         i3 += 16;
306 
307         const __m256i vprod3x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
308         const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
309 
310         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
311         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
312 
313         const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
314         const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 128)));
315         i4 += 16;
316 
317         const __m256i vprod4x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
318         const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
319 
320         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
321         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
322 
323         const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
324         const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 160)));
325         i5 += 16;
326 
327         const __m256i vprod5x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
328         const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
329 
330         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
331         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
332 
333         const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
334         const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 192)));
335         i6 += 16;
336 
337         const __m256i vprod6x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
338         const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
339 
340         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
341         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
342 
343         const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
344         const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 224)));
345         i7 += 16;
346 
347         const __m256i vprod7x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
348         const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
349 
350         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
351         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
352 
353         const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
354         const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 256)));
355         i8 += 16;
356 
357         const __m256i vprod8x0123456789ABCDEFlo =  _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
358         const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
359 
360         vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
361         vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
362 
363         vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
364         vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
365 
366         k += 16;
367 
368         __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
369         __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
370 
371         const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t)));
372         const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t) + 8 * sizeof(float)));
373         vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
374         vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
375 
376         const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
377         vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
378         vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
379 
380         vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
381         vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
382 
383         w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
384 
385         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
386         __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
387         __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), voutput_zero_point);
388 
389         const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
390 
391         __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
392         vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
393 
394         if XNN_LIKELY(c >= 16) {
395           _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
396           output += 16;
397           c -= 16;
398         } else {
399           if (c & 8) {
400             _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
401             vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
402             output += 8;
403           }
404           if (c & 4) {
405             unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
406             vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
407             output += 4;
408           }
409           if (c & 2) {
410             unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0));
411             vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
412             output += 2;
413           }
414           if (c & 1) {
415             *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
416             output += 1;
417           }
418           c = 0;
419         }
420       } while (c != 0);
421     }
422 
423     output = (int8_t*) ((uintptr_t) output + output_increment);
424   } while (--output_width != 0);
425 }
426