1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-avx2-mul16-vpmovsx.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_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx(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_up16x9__avx2_mul16_vpmovsx(
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 >= 16; c -= 16) {
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
87
88 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
89 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
90 i0 += 16;
91
92 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
93 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
94 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
95 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
96
97 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
98 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
99 i1 += 16;
100
101 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
102 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
103 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
104 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
105
106 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
107 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
108 i2 += 16;
109
110 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
111 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
112 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
113 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
114
115 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
116 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
117 i3 += 16;
118
119 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
120 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
121 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
122 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
123
124 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
125 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
126 i4 += 16;
127
128 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
129 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
130 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
131 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
132
133 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
134 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
135 i5 += 16;
136
137 const __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
138 const __m128i vprod5x89ABCDEF = _mm256_extracti128_si256(vprod5x0123456789ABCDEF, 1);
139 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod5x0123456789ABCDEF)));
140 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod5x89ABCDEF));
141
142 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
143 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
144 i6 += 16;
145
146 const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
147 const __m128i vprod6x89ABCDEF = _mm256_extracti128_si256(vprod6x0123456789ABCDEF, 1);
148 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod6x0123456789ABCDEF)));
149 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod6x89ABCDEF));
150
151 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
152 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
153 i7 += 16;
154
155 const __m256i vprod7x0123456789ABCDEF = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
156 const __m128i vprod7x89ABCDEF = _mm256_extracti128_si256(vprod7x0123456789ABCDEF, 1);
157 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod7x0123456789ABCDEF)));
158 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod7x89ABCDEF));
159
160 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
161 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
162 i8 += 16;
163
164 const __m256i vprod8x0123456789ABCDEF = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
165 const __m128i vprod8x89ABCDEF = _mm256_extracti128_si256(vprod8x0123456789ABCDEF, 1);
166 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod8x0123456789ABCDEF)));
167 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod8x89ABCDEF));
168
169 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
170
171 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
172 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
173
174 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) w);
175 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(float)));
176 w = (const void*) ((uintptr_t) w + 16 * sizeof(float));
177 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
178 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
179
180 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
181 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
182 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
183
184 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
185 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
186
187 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
188 const __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
189
190 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
191
192 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
193 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
194
195 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
196 output += 16;
197 }
198 if XNN_UNLIKELY(c != 0) {
199 {
200 __m256i vacc01234567 = _mm256_loadu_si256((const __m256i*) w);
201 __m256i vacc89ABCDEF = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
202
203
204 const __m256i vi0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i0));
205 const __m256i vk0x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
206
207 const __m256i vprod0x0123456789ABCDEF = _mm256_mullo_epi16(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF);
208 const __m128i vprod0x89ABCDEF = _mm256_extracti128_si256(vprod0x0123456789ABCDEF, 1);
209 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x0123456789ABCDEF)));
210 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod0x89ABCDEF));
211
212 const __m256i vi1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i1));
213 const __m256i vk1x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
214
215 const __m256i vprod1x0123456789ABCDEF = _mm256_mullo_epi16(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
216 const __m128i vprod1x89ABCDEF = _mm256_extracti128_si256(vprod1x0123456789ABCDEF, 1);
217 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x0123456789ABCDEF)));
218 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod1x89ABCDEF));
219
220 const __m256i vi2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i2));
221 const __m256i vk2x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
222
223 const __m256i vprod2x0123456789ABCDEF = _mm256_mullo_epi16(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF);
224 const __m128i vprod2x89ABCDEF = _mm256_extracti128_si256(vprod2x0123456789ABCDEF, 1);
225 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x0123456789ABCDEF)));
226 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod2x89ABCDEF));
227
228 const __m256i vi3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i3));
229 const __m256i vk3x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
230
231 const __m256i vprod3x0123456789ABCDEF = _mm256_mullo_epi16(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF);
232 const __m128i vprod3x89ABCDEF = _mm256_extracti128_si256(vprod3x0123456789ABCDEF, 1);
233 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x0123456789ABCDEF)));
234 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod3x89ABCDEF));
235
236 const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4));
237 const __m256i vk4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
238
239 const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF);
240 const __m128i vprod4x89ABCDEF = _mm256_extracti128_si256(vprod4x0123456789ABCDEF, 1);
241 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x0123456789ABCDEF)));
242 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod4x89ABCDEF));
243
244 const __m256i vi5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i5));
245 const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
246
247 const __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF);
248 const __m128i vprod5x89ABCDEF = _mm256_extracti128_si256(vprod5x0123456789ABCDEF, 1);
249 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod5x0123456789ABCDEF)));
250 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod5x89ABCDEF));
251
252 const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6));
253 const __m256i vk6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
254
255 const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF);
256 const __m128i vprod6x89ABCDEF = _mm256_extracti128_si256(vprod6x0123456789ABCDEF, 1);
257 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod6x0123456789ABCDEF)));
258 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod6x89ABCDEF));
259
260 const __m256i vi7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i7));
261 const __m256i vk7x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
262
263 const __m256i vprod7x0123456789ABCDEF = _mm256_mullo_epi16(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF);
264 const __m128i vprod7x89ABCDEF = _mm256_extracti128_si256(vprod7x0123456789ABCDEF, 1);
265 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod7x0123456789ABCDEF)));
266 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod7x89ABCDEF));
267
268 const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8));
269 const __m256i vk8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
270
271 const __m256i vprod8x0123456789ABCDEF = _mm256_mullo_epi16(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF);
272 const __m128i vprod8x89ABCDEF = _mm256_extracti128_si256(vprod8x0123456789ABCDEF, 1);
273 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod8x0123456789ABCDEF)));
274 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_cvtepi16_epi32(vprod8x89ABCDEF));
275
276
277 __m256 vfpacc01234567 = _mm256_cvtepi32_ps(vacc01234567);
278 __m256 vfpacc89ABCDEF = _mm256_cvtepi32_ps(vacc89ABCDEF);
279
280 const __m256 vscale01234567 = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
281 const __m256 vscale89ABCDEF = _mm256_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t) + 8 * sizeof(float)));
282 vfpacc01234567 = _mm256_mul_ps(vfpacc01234567, vscale01234567);
283 vfpacc89ABCDEF = _mm256_mul_ps(vfpacc89ABCDEF, vscale89ABCDEF);
284
285 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
286 vfpacc01234567 = _mm256_min_ps(vfpacc01234567, voutput_max_less_zero_point);
287 vfpacc89ABCDEF = _mm256_min_ps(vfpacc89ABCDEF, voutput_max_less_zero_point);
288
289 vacc01234567 = _mm256_cvtps_epi32(vfpacc01234567);
290 vacc89ABCDEF = _mm256_cvtps_epi32(vfpacc89ABCDEF);
291
292
293 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_avx2.output_zero_point);
294 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
295 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc89ABCDEF), _mm256_extracti128_si256(vacc89ABCDEF, 1)), voutput_zero_point);
296
297 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_avx2.output_min);
298
299 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
300 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
301
302 if (c & 8) {
303 _mm_storel_epi64((__m128i*) output, vout0123456789ABCDEF);
304 vout0123456789ABCDEF = _mm_unpackhi_epi64(vout0123456789ABCDEF, vout0123456789ABCDEF);
305 output += 8;
306 }
307 if (c & 4) {
308 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456789ABCDEF));
309 vout0123456789ABCDEF = _mm_srli_epi64(vout0123456789ABCDEF, 32);
310 output += 4;
311 }
312 if (c & 2) {
313 unaligned_store_u16(output, _mm_extract_epi16(vout0123456789ABCDEF, 0));
314 vout0123456789ABCDEF = _mm_srli_epi32(vout0123456789ABCDEF, 16);
315 output += 2;
316 }
317 if (c & 1) {
318 *output = (int8_t) _mm_extract_epi8(vout0123456789ABCDEF, 0);
319 output += 1;
320 }
321 }
322 }
323
324 output = (int8_t*) ((uintptr_t) output + output_increment);
325 } while (--output_width != 0);
326 }
327