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
16
xnn_qc8_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpunpck(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 input = (const int8_t**) ((uintptr_t) input + input_stride);
79
80 size_t c = channels;
81 const void* w = weights;
82 for (; c >= 32; c -= 32) {
83 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
84 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
85 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int32_t)));
86 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 24 * sizeof(int32_t)));
87
88 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
89 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
90 __m256i vaccGHIJOPQR = _mm256_inserti128_si256(vaccGHIJKLMN, _mm256_castsi256_si128(vaccOPQRSTUV), 1);
91 __m256i vaccKLMNSTUV = _mm256_permute2x128_si256(vaccGHIJKLMN, vaccOPQRSTUV, 0x31);
92
93
94 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
95 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
96 const __m256i vi0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i0 + 16)));
97 const __m256i vk0xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
98 i0 += 32;
99
100 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
101 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
102 const __m256i vprod0xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV);
103 const __m256i vprod0xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod0xGHIJKLMNOPQRSTUVlo, 15);
104
105 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
106 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
107 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod0xGHIJKLMNOPQRSTUVlo, vprod0xGHIJKLMNOPQRSTUVhi));
108 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod0xGHIJKLMNOPQRSTUVlo, vprod0xGHIJKLMNOPQRSTUVhi));
109
110 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
111 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
112 const __m256i vi1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i1 + 16)));
113 const __m256i vk1xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
114 i1 += 32;
115
116 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
117 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
118 const __m256i vprod1xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV);
119 const __m256i vprod1xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod1xGHIJKLMNOPQRSTUVlo, 15);
120
121 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
122 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
123 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod1xGHIJKLMNOPQRSTUVlo, vprod1xGHIJKLMNOPQRSTUVhi));
124 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod1xGHIJKLMNOPQRSTUVlo, vprod1xGHIJKLMNOPQRSTUVhi));
125
126 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
127 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
128 const __m256i vi2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i2 + 16)));
129 const __m256i vk2xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
130 i2 += 32;
131
132 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
133 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
134 const __m256i vprod2xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV);
135 const __m256i vprod2xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod2xGHIJKLMNOPQRSTUVlo, 15);
136
137 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
138 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
139 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod2xGHIJKLMNOPQRSTUVlo, vprod2xGHIJKLMNOPQRSTUVhi));
140 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod2xGHIJKLMNOPQRSTUVlo, vprod2xGHIJKLMNOPQRSTUVhi));
141
142 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
143 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
144 const __m256i vi3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i3 + 16)));
145 const __m256i vk3xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
146 i3 += 32;
147
148 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
149 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
150 const __m256i vprod3xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV);
151 const __m256i vprod3xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod3xGHIJKLMNOPQRSTUVlo, 15);
152
153 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
154 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
155 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod3xGHIJKLMNOPQRSTUVlo, vprod3xGHIJKLMNOPQRSTUVhi));
156 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod3xGHIJKLMNOPQRSTUVlo, vprod3xGHIJKLMNOPQRSTUVhi));
157
158 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
159 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
160 const __m256i vi4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i4 + 16)));
161 const __m256i vk4xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
162 i4 += 32;
163
164 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
165 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
166 const __m256i vprod4xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV);
167 const __m256i vprod4xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod4xGHIJKLMNOPQRSTUVlo, 15);
168
169 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
170 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
171 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod4xGHIJKLMNOPQRSTUVlo, vprod4xGHIJKLMNOPQRSTUVhi));
172 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod4xGHIJKLMNOPQRSTUVlo, vprod4xGHIJKLMNOPQRSTUVhi));
173
174 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
175 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
176 const __m256i vi5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i5 + 16)));
177 const __m256i vk5xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
178 i5 += 32;
179
180 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
181 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
182 const __m256i vprod5xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV);
183 const __m256i vprod5xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod5xGHIJKLMNOPQRSTUVlo, 15);
184
185 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
186 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
187 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod5xGHIJKLMNOPQRSTUVlo, vprod5xGHIJKLMNOPQRSTUVhi));
188 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod5xGHIJKLMNOPQRSTUVlo, vprod5xGHIJKLMNOPQRSTUVhi));
189
190 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
191 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
192 const __m256i vi6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i6 + 16)));
193 const __m256i vk6xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
194 i6 += 32;
195
196 const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
197 const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
198 const __m256i vprod6xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV);
199 const __m256i vprod6xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod6xGHIJKLMNOPQRSTUVlo, 15);
200
201 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
202 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
203 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod6xGHIJKLMNOPQRSTUVlo, vprod6xGHIJKLMNOPQRSTUVhi));
204 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod6xGHIJKLMNOPQRSTUVlo, vprod6xGHIJKLMNOPQRSTUVhi));
205
206 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
207 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
208 const __m256i vi7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i7 + 16)));
209 const __m256i vk7xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
210 i7 += 32;
211
212 const __m256i vprod7x0123456789ABCDEFlo = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
213 const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
214 const __m256i vprod7xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV);
215 const __m256i vprod7xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod7xGHIJKLMNOPQRSTUVlo, 15);
216
217 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
218 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
219 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod7xGHIJKLMNOPQRSTUVlo, vprod7xGHIJKLMNOPQRSTUVhi));
220 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod7xGHIJKLMNOPQRSTUVlo, vprod7xGHIJKLMNOPQRSTUVhi));
221
222 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
223 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
224 const __m256i vi8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (i8 + 16)));
225 const __m256i vk8xGHIJKLMNOPQRSTUV = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
226 i8 += 32;
227
228 const __m256i vprod8x0123456789ABCDEFlo = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
229 const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
230 const __m256i vprod8xGHIJKLMNOPQRSTUVlo = _mm256_mullo_epi16(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV);
231 const __m256i vprod8xGHIJKLMNOPQRSTUVhi = _mm256_srai_epi16(vprod8xGHIJKLMNOPQRSTUVlo, 15);
232
233 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
234 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
235 vaccGHIJOPQR = _mm256_add_epi32(vaccGHIJOPQR, _mm256_unpacklo_epi16(vprod8xGHIJKLMNOPQRSTUVlo, vprod8xGHIJKLMNOPQRSTUVhi));
236 vaccKLMNSTUV = _mm256_add_epi32(vaccKLMNSTUV, _mm256_unpackhi_epi16(vprod8xGHIJKLMNOPQRSTUVlo, vprod8xGHIJKLMNOPQRSTUVhi));
237
238 w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t));
239
240 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
241 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
242 vaccGHIJKLMN = _mm256_inserti128_si256(vaccGHIJOPQR, _mm256_castsi256_si128(vaccKLMNSTUV), 1);
243 vaccOPQRSTUV = _mm256_permute2x128_si256(vaccGHIJOPQR, vaccKLMNSTUV, 0x31);
244
245 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
246 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
247 __m256 vfpaccGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
248 __m256 vfpaccOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
249
250 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w);
251 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float)));
252 const __m256 vscaleGHIJKLMN = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(float)));
253 const __m256 vscaleOPQRSTUV = _mm256_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(float)));
254 w = (const void*) ((uintptr_t) w + 32 * sizeof(float));
255 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
256 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
257 vfpaccGHIJKLMN = _mm256_mul_ps(vfpaccGHIJKLMN, vscaleGHIJKLMN);
258 vfpaccOPQRSTUV = _mm256_mul_ps(vfpaccOPQRSTUV, vscaleOPQRSTUV);
259
260 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->avx2.output_max_less_zero_point);
261 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
262 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
263 vfpaccGHIJKLMN = _mm256_min_ps(vfpaccGHIJKLMN, voutput_max_less_zero_point);
264 vfpaccOPQRSTUV = _mm256_min_ps(vfpaccOPQRSTUV, voutput_max_less_zero_point);
265
266 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
267 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
268 vaccGHIJKLMN = _mm256_cvtps_epi32(vfpaccGHIJKLMN);
269 vaccOPQRSTUV = _mm256_cvtps_epi32(vfpaccOPQRSTUV);
270
271 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
272 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
273 const __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
274
275 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
276 __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
277
278 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
279 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
280 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
281
282 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
283 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
284 output += 32;
285 }
286 if XNN_UNLIKELY(c != 0) {
287 const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t));
288 do {
289 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
290 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
291
292 __m256i vacc012389AB = _mm256_inserti128_si256(vacc01234567, _mm256_castsi256_si128(vacc89ABCDEF), 1);
293 __m256i vacc4567CDEF = _mm256_permute2x128_si256(vacc01234567, vacc89ABCDEF, 0x31);
294
295
296 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
297 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) k));
298 i0 += 16;
299
300 const __m256i vprod0x0123456789ABCDEFlo = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
301 const __m256i vprod0x0123456789ABCDEFhi = _mm256_srai_epi16(vprod0x0123456789ABCDEFlo, 15);
302
303 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
304 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod0x0123456789ABCDEFlo, vprod0x0123456789ABCDEFhi));
305
306 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
307 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 32)));
308 i1 += 16;
309
310 const __m256i vprod1x0123456789ABCDEFlo = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
311 const __m256i vprod1x0123456789ABCDEFhi = _mm256_srai_epi16(vprod1x0123456789ABCDEFlo, 15);
312
313 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
314 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod1x0123456789ABCDEFlo, vprod1x0123456789ABCDEFhi));
315
316 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
317 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 64)));
318 i2 += 16;
319
320 const __m256i vprod2x0123456789ABCDEFlo = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
321 const __m256i vprod2x0123456789ABCDEFhi = _mm256_srai_epi16(vprod2x0123456789ABCDEFlo, 15);
322
323 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
324 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod2x0123456789ABCDEFlo, vprod2x0123456789ABCDEFhi));
325
326 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
327 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 96)));
328 i3 += 16;
329
330 const __m256i vprod3x0123456789ABCDEFlo = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
331 const __m256i vprod3x0123456789ABCDEFhi = _mm256_srai_epi16(vprod3x0123456789ABCDEFlo, 15);
332
333 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
334 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod3x0123456789ABCDEFlo, vprod3x0123456789ABCDEFhi));
335
336 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
337 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 128)));
338 i4 += 16;
339
340 const __m256i vprod4x0123456789ABCDEFlo = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
341 const __m256i vprod4x0123456789ABCDEFhi = _mm256_srai_epi16(vprod4x0123456789ABCDEFlo, 15);
342
343 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
344 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod4x0123456789ABCDEFlo, vprod4x0123456789ABCDEFhi));
345
346 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
347 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 160)));
348 i5 += 16;
349
350 const __m256i vprod5x0123456789ABCDEFlo = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
351 const __m256i vprod5x0123456789ABCDEFhi = _mm256_srai_epi16(vprod5x0123456789ABCDEFlo, 15);
352
353 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
354 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod5x0123456789ABCDEFlo, vprod5x0123456789ABCDEFhi));
355
356 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
357 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 192)));
358 i6 += 16;
359
360 const __m256i vprod6x0123456789ABCDEFlo = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
361 const __m256i vprod6x0123456789ABCDEFhi = _mm256_srai_epi16(vprod6x0123456789ABCDEFlo, 15);
362
363 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
364 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod6x0123456789ABCDEFlo, vprod6x0123456789ABCDEFhi));
365
366 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
367 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 224)));
368 i7 += 16;
369
370 const __m256i vprod7x0123456789ABCDEFlo = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
371 const __m256i vprod7x0123456789ABCDEFhi = _mm256_srai_epi16(vprod7x0123456789ABCDEFlo, 15);
372
373 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
374 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod7x0123456789ABCDEFlo, vprod7x0123456789ABCDEFhi));
375
376 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
377 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 256)));
378 i8 += 16;
379
380 const __m256i vprod8x0123456789ABCDEFlo = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
381 const __m256i vprod8x0123456789ABCDEFhi = _mm256_srai_epi16(vprod8x0123456789ABCDEFlo, 15);
382
383 vacc012389AB = _mm256_add_epi32(vacc012389AB, _mm256_unpacklo_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
384 vacc4567CDEF = _mm256_add_epi32(vacc4567CDEF, _mm256_unpackhi_epi16(vprod8x0123456789ABCDEFlo, vprod8x0123456789ABCDEFhi));
385
386 vacc01234567 = _mm256_inserti128_si256(vacc012389AB, _mm256_castsi256_si128(vacc4567CDEF), 1);
387 vacc89ABCDEF = _mm256_permute2x128_si256(vacc012389AB, vacc4567CDEF, 0x31);
388
389 k += 16;
390
391 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
392 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
393
394 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t)));
395 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t) + 8 * sizeof(float)));
396 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
397 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
398
399 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->avx2.output_max_less_zero_point);
400 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
401 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
402
403 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
404 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
405
406 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
407
408 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx2.output_zero_point);
409 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
410 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), voutput_zero_point);
411
412 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
413
414 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
415 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
416
417 if XNN_LIKELY(c >= 16) {
418 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
419 output += 16;
420 c -= 16;
421 } else {
422 if (c & 8) {
423 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
424 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
425 output += 8;
426 }
427 if (c & 4) {
428 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF);
429 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
430 output += 4;
431 }
432 if (c & 2) {
433 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456789ABCDEF, 0);
434 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
435 output += 2;
436 }
437 if (c & 1) {
438 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
439 output += 1;
440 }
441 c = 0;
442 }
443 } while (c != 0);
444 }
445
446 output = (int8_t*) ((uintptr_t) output + output_increment);
447 } while (--output_width != 0);
448 }
449