1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx2-mul32.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 <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32(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_up32x9__avx2_mul32(
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 const __m256i vk_zero_point = _mm256_cvtepu16_epi32(_mm_load_si128((const __m128i*) params->fp32_avx2.kernel_zero_point));
34 do {
35 const uint8_t* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
39 }
40 const uint8_t* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
44 }
45 const uint8_t* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
49 }
50 const uint8_t* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
54 }
55 const uint8_t* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
59 }
60 const uint8_t* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
64 }
65 const uint8_t* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
69 }
70 const uint8_t* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
74 }
75 const uint8_t* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
79 }
80 input = (const uint8_t**) ((uintptr_t) input + input_stride);
81
82 size_t c = channels;
83 const void* w = weights;
84 for (; c >= 32; c -= 32) {
85 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
86 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 8));
87 __m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 16));
88 __m256i vaccOPQRSTUV = _mm256_loadu_si256((const __m256i*) ((const int32_t*) w + 24));
89
90
91 const __m256i vi0x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i0));
92 const __m256i vk0x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
93 const __m256i vi0x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
94 const __m256i vk0x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 8 * sizeof(uint8_t)))), vk_zero_point);
95 const __m256i vi0xGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 16)));
96 const __m256i vk0xGHIJKLMN = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
97 const __m256i vi0xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 24)));
98 const __m256i vk0xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 24 * sizeof(uint8_t)))), vk_zero_point);
99 i0 += 32;
100
101 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
102 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi0x89ABCDEF, vk0x89ABCDEF));
103 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi0xGHIJKLMN, vk0xGHIJKLMN));
104 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi0xOPQRSTUV, vk0xOPQRSTUV));
105
106 const __m256i vi1x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i1));
107 const __m256i vk1x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
108 const __m256i vi1x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
109 const __m256i vk1x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 40 * sizeof(uint8_t)))), vk_zero_point);
110 const __m256i vi1xGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 16)));
111 const __m256i vk1xGHIJKLMN = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
112 const __m256i vi1xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 24)));
113 const __m256i vk1xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 56 * sizeof(uint8_t)))), vk_zero_point);
114 i1 += 32;
115
116 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
117 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi1x89ABCDEF, vk1x89ABCDEF));
118 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi1xGHIJKLMN, vk1xGHIJKLMN));
119 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi1xOPQRSTUV, vk1xOPQRSTUV));
120
121 const __m256i vi2x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i2));
122 const __m256i vk2x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
123 const __m256i vi2x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
124 const __m256i vk2x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 72 * sizeof(uint8_t)))), vk_zero_point);
125 const __m256i vi2xGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 16)));
126 const __m256i vk2xGHIJKLMN = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
127 const __m256i vi2xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i2 + 24)));
128 const __m256i vk2xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 88 * sizeof(uint8_t)))), vk_zero_point);
129 i2 += 32;
130
131 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
132 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi2x89ABCDEF, vk2x89ABCDEF));
133 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi2xGHIJKLMN, vk2xGHIJKLMN));
134 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi2xOPQRSTUV, vk2xOPQRSTUV));
135
136 const __m256i vi3x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i3));
137 const __m256i vk3x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
138 const __m256i vi3x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
139 const __m256i vk3x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 104 * sizeof(uint8_t)))), vk_zero_point);
140 const __m256i vi3xGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 16)));
141 const __m256i vk3xGHIJKLMN = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
142 const __m256i vi3xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i3 + 24)));
143 const __m256i vk3xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 120 * sizeof(uint8_t)))), vk_zero_point);
144 i3 += 32;
145
146 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
147 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi3x89ABCDEF, vk3x89ABCDEF));
148 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN));
149 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi3xOPQRSTUV, vk3xOPQRSTUV));
150
151 const __m256i vi4x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i4));
152 const __m256i vk4x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
153 const __m256i vi4x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
154 const __m256i vk4x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 136 * sizeof(uint8_t)))), vk_zero_point);
155 const __m256i vi4xGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 16)));
156 const __m256i vk4xGHIJKLMN = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(uint8_t)))), vk_zero_point);
157 const __m256i vi4xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i4 + 24)));
158 const __m256i vk4xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 152 * sizeof(uint8_t)))), vk_zero_point);
159 i4 += 32;
160
161 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
162 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi4x89ABCDEF, vk4x89ABCDEF));
163 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi4xGHIJKLMN, vk4xGHIJKLMN));
164 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi4xOPQRSTUV, vk4xOPQRSTUV));
165
166 const __m256i vi5x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i5));
167 const __m256i vk5x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(uint8_t)))), vk_zero_point);
168 const __m256i vi5x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
169 const __m256i vk5x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 168 * sizeof(uint8_t)))), vk_zero_point);
170 const __m256i vi5xGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 16)));
171 const __m256i vk5xGHIJKLMN = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(uint8_t)))), vk_zero_point);
172 const __m256i vi5xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i5 + 24)));
173 const __m256i vk5xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 184 * sizeof(uint8_t)))), vk_zero_point);
174 i5 += 32;
175
176 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
177 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi5x89ABCDEF, vk5x89ABCDEF));
178 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi5xGHIJKLMN, vk5xGHIJKLMN));
179 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi5xOPQRSTUV, vk5xOPQRSTUV));
180
181 const __m256i vi6x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i6));
182 const __m256i vk6x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(uint8_t)))), vk_zero_point);
183 const __m256i vi6x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
184 const __m256i vk6x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 200 * sizeof(uint8_t)))), vk_zero_point);
185 const __m256i vi6xGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 16)));
186 const __m256i vk6xGHIJKLMN = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(uint8_t)))), vk_zero_point);
187 const __m256i vi6xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i6 + 24)));
188 const __m256i vk6xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 216 * sizeof(uint8_t)))), vk_zero_point);
189 i6 += 32;
190
191 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
192 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi6x89ABCDEF, vk6x89ABCDEF));
193 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi6xGHIJKLMN, vk6xGHIJKLMN));
194 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi6xOPQRSTUV, vk6xOPQRSTUV));
195
196 const __m256i vi7x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i7));
197 const __m256i vk7x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(uint8_t)))), vk_zero_point);
198 const __m256i vi7x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 8)));
199 const __m256i vk7x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 232 * sizeof(uint8_t)))), vk_zero_point);
200 const __m256i vi7xGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 16)));
201 const __m256i vk7xGHIJKLMN = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(uint8_t)))), vk_zero_point);
202 const __m256i vi7xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i7 + 24)));
203 const __m256i vk7xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 248 * sizeof(uint8_t)))), vk_zero_point);
204 i7 += 32;
205
206 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
207 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi7x89ABCDEF, vk7x89ABCDEF));
208 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi7xGHIJKLMN, vk7xGHIJKLMN));
209 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi7xOPQRSTUV, vk7xOPQRSTUV));
210
211 const __m256i vi8x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i8));
212 const __m256i vk8x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(uint8_t)))), vk_zero_point);
213 const __m256i vi8x89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 8)));
214 const __m256i vk8x89ABCDEF = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 264 * sizeof(uint8_t)))), vk_zero_point);
215 const __m256i vi8xGHIJKLMN = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 16)));
216 const __m256i vk8xGHIJKLMN = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(uint8_t)))), vk_zero_point);
217 const __m256i vi8xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i8 + 24)));
218 const __m256i vk8xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 280 * sizeof(uint8_t)))), vk_zero_point);
219 i8 += 32;
220
221 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
222 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vi8x89ABCDEF, vk8x89ABCDEF));
223 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi8xGHIJKLMN, vk8xGHIJKLMN));
224 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi8xOPQRSTUV, vk8xOPQRSTUV));
225
226 w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(uint8_t));
227
228 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
229 __m256 vscaled89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
230 __m256 vscaledGHIJKLMN = _mm256_cvtepi32_ps(vaccGHIJKLMN);
231 __m256 vscaledOPQRSTUV = _mm256_cvtepi32_ps(vaccOPQRSTUV);
232
233 const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
234 vscaled01234567 = _mm256_mul_ps(vscaled01234567, vscale);
235 vscaled89ABCDEF = _mm256_mul_ps(vscaled89ABCDEF, vscale);
236 vscaledGHIJKLMN = _mm256_mul_ps(vscaledGHIJKLMN, vscale);
237 vscaledOPQRSTUV = _mm256_mul_ps(vscaledOPQRSTUV, vscale);
238
239 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
240 vscaled01234567 = _mm256_min_ps(vscaled01234567, voutput_max_less_zero_point);
241 vscaled89ABCDEF = _mm256_min_ps(vscaled89ABCDEF, voutput_max_less_zero_point);
242 vscaledGHIJKLMN = _mm256_min_ps(vscaledGHIJKLMN, voutput_max_less_zero_point);
243 vscaledOPQRSTUV = _mm256_min_ps(vscaledOPQRSTUV, voutput_max_less_zero_point);
244
245 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
246 vacc89ABCDEF = _mm256_cvtps_epi32(vscaled89ABCDEF);
247 vaccGHIJKLMN = _mm256_cvtps_epi32(vscaledGHIJKLMN);
248 vaccOPQRSTUV = _mm256_cvtps_epi32(vscaledOPQRSTUV);
249
250 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
251 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
252 __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
253
254 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
255 __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
256
257 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
258 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
259 voutGHIJKLMNOPQRSTUV = _mm_max_epu8(voutGHIJKLMNOPQRSTUV, voutput_min);
260
261 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
262 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
263 output += 32;
264 }
265 if XNN_UNLIKELY(c != 0) {
266 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 32);
267 do {
268 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
269
270
271 const __m256i vi0x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i0));
272 const __m256i vk0x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) k)), vk_zero_point);
273 i0 += 8;
274
275 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi0x01234567, vk0x01234567));
276
277 const __m256i vi1x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i1));
278 const __m256i vk1x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 32))), vk_zero_point);
279 i1 += 8;
280
281 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi1x01234567, vk1x01234567));
282
283 const __m256i vi2x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i2));
284 const __m256i vk2x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 64))), vk_zero_point);
285 i2 += 8;
286
287 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi2x01234567, vk2x01234567));
288
289 const __m256i vi3x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i3));
290 const __m256i vk3x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 96))), vk_zero_point);
291 i3 += 8;
292
293 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi3x01234567, vk3x01234567));
294
295 const __m256i vi4x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i4));
296 const __m256i vk4x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 128))), vk_zero_point);
297 i4 += 8;
298
299 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi4x01234567, vk4x01234567));
300
301 const __m256i vi5x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i5));
302 const __m256i vk5x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 160))), vk_zero_point);
303 i5 += 8;
304
305 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi5x01234567, vk5x01234567));
306
307 const __m256i vi6x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i6));
308 const __m256i vk6x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 192))), vk_zero_point);
309 i6 += 8;
310
311 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi6x01234567, vk6x01234567));
312
313 const __m256i vi7x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i7));
314 const __m256i vk7x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 224))), vk_zero_point);
315 i7 += 8;
316
317 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi7x01234567, vk7x01234567));
318
319 const __m256i vi8x01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) i8));
320 const __m256i vk8x01234567 = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (k + 256))), vk_zero_point);
321 i8 += 8;
322
323 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vi8x01234567, vk8x01234567));
324
325 k += 8;
326
327 __m256 vscaled01234567 = _mm256_cvtepi32_ps(vacc01234567);
328 vscaled01234567 = _mm256_mul_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.scale));
329 vscaled01234567 = _mm256_min_ps(vscaled01234567, _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point));
330 vacc01234567 = _mm256_cvtps_epi32(vscaled01234567);
331
332 w = (const void*) ((const int32_t*) w + 8);
333
334 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
335 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
336
337 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
338
339 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
340 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
341
342 if XNN_LIKELY(c >= 8) {
343 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
344 output += 8;
345 c -= 8;
346 } else {
347 if (c & 4) {
348 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
349 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
350 output += 4;
351 }
352 if (c & 2) {
353 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
354 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
355 output += 2;
356 }
357 if (c & 1) {
358 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
359 output += 1;
360 }
361 c = 0;
362 }
363 } while (c != 0);
364 }
365
366 output = (uint8_t*) ((uintptr_t) output + output_increment);
367 } while (--output_width != 0);
368 }
369