1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/dwconv.h>
20 #include <xnnpack/intrinsics-polyfill.h>
21 #include <xnnpack/unaligned.h>
22
23
xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__xop_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)])24 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__xop_mul32(
25 size_t channels,
26 size_t output_width,
27 const uint8_t** input,
28 const void* weights,
29 uint8_t* output,
30 size_t input_stride,
31 size_t output_increment,
32 size_t input_offset,
33 const uint8_t* zero,
34 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(channels != 0);
37 assert(output_width != 0);
38
39 const __m128i vk_zero_point = _mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) params->fp32_sse2.kernel_zero_point));
40 do {
41 const uint8_t* i0 = input[0];
42 assert(i0 != NULL);
43 if XNN_UNPREDICTABLE(i0 != zero) {
44 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
45 }
46 const uint8_t* i1 = input[1];
47 assert(i1 != NULL);
48 if XNN_UNPREDICTABLE(i1 != zero) {
49 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
50 }
51 const uint8_t* i2 = input[2];
52 assert(i2 != NULL);
53 if XNN_UNPREDICTABLE(i2 != zero) {
54 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
55 }
56 const uint8_t* i3 = input[3];
57 assert(i3 != NULL);
58 if XNN_UNPREDICTABLE(i3 != zero) {
59 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
60 }
61 const uint8_t* i4 = input[4];
62 assert(i4 != NULL);
63 if XNN_UNPREDICTABLE(i4 != zero) {
64 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
65 }
66 const uint8_t* i5 = input[5];
67 assert(i5 != NULL);
68 if XNN_UNPREDICTABLE(i5 != zero) {
69 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
70 }
71 const uint8_t* i6 = input[6];
72 assert(i6 != NULL);
73 if XNN_UNPREDICTABLE(i6 != zero) {
74 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
75 }
76 const uint8_t* i7 = input[7];
77 assert(i7 != NULL);
78 if XNN_UNPREDICTABLE(i7 != zero) {
79 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
80 }
81 const uint8_t* i8 = input[8];
82 assert(i8 != NULL);
83 if XNN_UNPREDICTABLE(i8 != zero) {
84 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
85 }
86 const uint8_t* i9 = input[9];
87 assert(i9 != NULL);
88 if XNN_UNPREDICTABLE(i9 != zero) {
89 i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
90 }
91 const uint8_t* i10 = input[10];
92 assert(i10 != NULL);
93 if XNN_UNPREDICTABLE(i10 != zero) {
94 i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
95 }
96 const uint8_t* i11 = input[11];
97 assert(i11 != NULL);
98 if XNN_UNPREDICTABLE(i11 != zero) {
99 i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
100 }
101 const uint8_t* i12 = input[12];
102 assert(i12 != NULL);
103 if XNN_UNPREDICTABLE(i12 != zero) {
104 i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
105 }
106 const uint8_t* i13 = input[13];
107 assert(i13 != NULL);
108 if XNN_UNPREDICTABLE(i13 != zero) {
109 i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
110 }
111 const uint8_t* i14 = input[14];
112 assert(i14 != NULL);
113 if XNN_UNPREDICTABLE(i14 != zero) {
114 i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
115 }
116 const uint8_t* i15 = input[15];
117 assert(i15 != NULL);
118 if XNN_UNPREDICTABLE(i15 != zero) {
119 i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
120 }
121 const uint8_t* i16 = input[16];
122 assert(i16 != NULL);
123 if XNN_UNPREDICTABLE(i16 != zero) {
124 i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
125 }
126 const uint8_t* i17 = input[17];
127 assert(i17 != NULL);
128 if XNN_UNPREDICTABLE(i17 != zero) {
129 i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
130 }
131 const uint8_t* i18 = input[18];
132 assert(i18 != NULL);
133 if XNN_UNPREDICTABLE(i18 != zero) {
134 i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
135 }
136 const uint8_t* i19 = input[19];
137 assert(i19 != NULL);
138 if XNN_UNPREDICTABLE(i19 != zero) {
139 i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
140 }
141 const uint8_t* i20 = input[20];
142 assert(i20 != NULL);
143 if XNN_UNPREDICTABLE(i20 != zero) {
144 i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
145 }
146 const uint8_t* i21 = input[21];
147 assert(i21 != NULL);
148 if XNN_UNPREDICTABLE(i21 != zero) {
149 i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
150 }
151 const uint8_t* i22 = input[22];
152 assert(i22 != NULL);
153 if XNN_UNPREDICTABLE(i22 != zero) {
154 i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
155 }
156 const uint8_t* i23 = input[23];
157 assert(i23 != NULL);
158 if XNN_UNPREDICTABLE(i23 != zero) {
159 i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
160 }
161 const uint8_t* i24 = input[24];
162 assert(i24 != NULL);
163 if XNN_UNPREDICTABLE(i24 != zero) {
164 i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
165 }
166 input = (const uint8_t**) ((uintptr_t) input + input_stride);
167
168 size_t c = channels;
169 const void* w = weights;
170 for (; c >= 8; c -= 8) {
171 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
172 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
173
174
175 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
176 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t))))), vk_zero_point);
177 const __m128i vi0x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
178 const __m128i vk0x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(uint8_t))))), vk_zero_point);
179 i0 += 8;
180
181 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
182 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
183
184 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
185 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t))))), vk_zero_point);
186 const __m128i vi1x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
187 const __m128i vk1x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(uint8_t))))), vk_zero_point);
188 i1 += 8;
189
190 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
191 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
192
193 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
194 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t))))), vk_zero_point);
195 const __m128i vi2x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
196 const __m128i vk2x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(uint8_t))))), vk_zero_point);
197 i2 += 8;
198
199 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
200 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
201
202 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
203 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t))))), vk_zero_point);
204 const __m128i vi3x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
205 const __m128i vk3x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(uint8_t))))), vk_zero_point);
206 i3 += 8;
207
208 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
209 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
210
211 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
212 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t))))), vk_zero_point);
213 const __m128i vi4x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
214 const __m128i vk4x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(uint8_t))))), vk_zero_point);
215 i4 += 8;
216
217 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
218 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
219
220 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
221 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t))))), vk_zero_point);
222 const __m128i vi5x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
223 const __m128i vk5x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(uint8_t))))), vk_zero_point);
224 i5 += 8;
225
226 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
227 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
228
229 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
230 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t))))), vk_zero_point);
231 const __m128i vi6x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
232 const __m128i vk6x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 52 * sizeof(uint8_t))))), vk_zero_point);
233 i6 += 8;
234
235 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
236 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
237
238 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
239 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t))))), vk_zero_point);
240 const __m128i vi7x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
241 const __m128i vk7x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 60 * sizeof(uint8_t))))), vk_zero_point);
242 i7 += 8;
243
244 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
245 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
246
247 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
248 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t))))), vk_zero_point);
249 const __m128i vi8x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
250 const __m128i vk8x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 68 * sizeof(uint8_t))))), vk_zero_point);
251 i8 += 8;
252
253 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
254 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
255
256 const __m128i vi9x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
257 const __m128i vk9x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t))))), vk_zero_point);
258 const __m128i vi9x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 4)));
259 const __m128i vk9x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 76 * sizeof(uint8_t))))), vk_zero_point);
260 i9 += 8;
261
262 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
263 vacc4567 = _mm_macc_epi32(vi9x4567, vk9x4567, vacc4567);
264
265 const __m128i vi10x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
266 const __m128i vk10x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t))))), vk_zero_point);
267 const __m128i vi10x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 4)));
268 const __m128i vk10x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 84 * sizeof(uint8_t))))), vk_zero_point);
269 i10 += 8;
270
271 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
272 vacc4567 = _mm_macc_epi32(vi10x4567, vk10x4567, vacc4567);
273
274 const __m128i vi11x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
275 const __m128i vk11x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t))))), vk_zero_point);
276 const __m128i vi11x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 4)));
277 const __m128i vk11x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 92 * sizeof(uint8_t))))), vk_zero_point);
278 i11 += 8;
279
280 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
281 vacc4567 = _mm_macc_epi32(vi11x4567, vk11x4567, vacc4567);
282
283 const __m128i vi12x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
284 const __m128i vk12x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t))))), vk_zero_point);
285 const __m128i vi12x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 4)));
286 const __m128i vk12x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 100 * sizeof(uint8_t))))), vk_zero_point);
287 i12 += 8;
288
289 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
290 vacc4567 = _mm_macc_epi32(vi12x4567, vk12x4567, vacc4567);
291
292 const __m128i vi13x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
293 const __m128i vk13x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t))))), vk_zero_point);
294 const __m128i vi13x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 4)));
295 const __m128i vk13x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 108 * sizeof(uint8_t))))), vk_zero_point);
296 i13 += 8;
297
298 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
299 vacc4567 = _mm_macc_epi32(vi13x4567, vk13x4567, vacc4567);
300
301 const __m128i vi14x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
302 const __m128i vk14x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t))))), vk_zero_point);
303 const __m128i vi14x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 4)));
304 const __m128i vk14x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 116 * sizeof(uint8_t))))), vk_zero_point);
305 i14 += 8;
306
307 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
308 vacc4567 = _mm_macc_epi32(vi14x4567, vk14x4567, vacc4567);
309
310 const __m128i vi15x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
311 const __m128i vk15x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t))))), vk_zero_point);
312 const __m128i vi15x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 4)));
313 const __m128i vk15x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 124 * sizeof(uint8_t))))), vk_zero_point);
314 i15 += 8;
315
316 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
317 vacc4567 = _mm_macc_epi32(vi15x4567, vk15x4567, vacc4567);
318
319 const __m128i vi16x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
320 const __m128i vk16x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t))))), vk_zero_point);
321 const __m128i vi16x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 4)));
322 const __m128i vk16x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 132 * sizeof(uint8_t))))), vk_zero_point);
323 i16 += 8;
324
325 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
326 vacc4567 = _mm_macc_epi32(vi16x4567, vk16x4567, vacc4567);
327
328 const __m128i vi17x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
329 const __m128i vk17x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t))))), vk_zero_point);
330 const __m128i vi17x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 4)));
331 const __m128i vk17x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 140 * sizeof(uint8_t))))), vk_zero_point);
332 i17 += 8;
333
334 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
335 vacc4567 = _mm_macc_epi32(vi17x4567, vk17x4567, vacc4567);
336
337 const __m128i vi18x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
338 const __m128i vk18x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t))))), vk_zero_point);
339 const __m128i vi18x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 4)));
340 const __m128i vk18x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 148 * sizeof(uint8_t))))), vk_zero_point);
341 i18 += 8;
342
343 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
344 vacc4567 = _mm_macc_epi32(vi18x4567, vk18x4567, vacc4567);
345
346 const __m128i vi19x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
347 const __m128i vk19x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t))))), vk_zero_point);
348 const __m128i vi19x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 4)));
349 const __m128i vk19x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 156 * sizeof(uint8_t))))), vk_zero_point);
350 i19 += 8;
351
352 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
353 vacc4567 = _mm_macc_epi32(vi19x4567, vk19x4567, vacc4567);
354
355 const __m128i vi20x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
356 const __m128i vk20x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t))))), vk_zero_point);
357 const __m128i vi20x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 4)));
358 const __m128i vk20x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 164 * sizeof(uint8_t))))), vk_zero_point);
359 i20 += 8;
360
361 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
362 vacc4567 = _mm_macc_epi32(vi20x4567, vk20x4567, vacc4567);
363
364 const __m128i vi21x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
365 const __m128i vk21x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t))))), vk_zero_point);
366 const __m128i vi21x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 4)));
367 const __m128i vk21x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 172 * sizeof(uint8_t))))), vk_zero_point);
368 i21 += 8;
369
370 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
371 vacc4567 = _mm_macc_epi32(vi21x4567, vk21x4567, vacc4567);
372
373 const __m128i vi22x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
374 const __m128i vk22x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t))))), vk_zero_point);
375 const __m128i vi22x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 4)));
376 const __m128i vk22x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 180 * sizeof(uint8_t))))), vk_zero_point);
377 i22 += 8;
378
379 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
380 vacc4567 = _mm_macc_epi32(vi22x4567, vk22x4567, vacc4567);
381
382 const __m128i vi23x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
383 const __m128i vk23x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t))))), vk_zero_point);
384 const __m128i vi23x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 4)));
385 const __m128i vk23x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 188 * sizeof(uint8_t))))), vk_zero_point);
386 i23 += 8;
387
388 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
389 vacc4567 = _mm_macc_epi32(vi23x4567, vk23x4567, vacc4567);
390
391 const __m128i vi24x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
392 const __m128i vk24x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t))))), vk_zero_point);
393 const __m128i vi24x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 4)));
394 const __m128i vk24x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 196 * sizeof(uint8_t))))), vk_zero_point);
395 i24 += 8;
396
397 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
398 vacc4567 = _mm_macc_epi32(vi24x4567, vk24x4567, vacc4567);
399
400 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t));
401
402 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
403 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
404
405 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
406 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
407 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
408
409 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
410 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
411 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
412
413 vacc0123 = _mm_cvtps_epi32(vscaled0123);
414 vacc4567 = _mm_cvtps_epi32(vscaled4567);
415
416 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
417 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
418
419 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
420 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
421 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
422
423 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
424 output += 8;
425 }
426 if XNN_UNLIKELY(c != 0) {
427 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 8);
428 do {
429 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
430
431 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
432 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) k))), vk_zero_point);
433 i0 += 4;
434
435 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
436 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
437 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 8)))), vk_zero_point);
438 i1 += 4;
439
440 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
441 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
442 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16)))), vk_zero_point);
443 i2 += 4;
444
445 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
446 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
447 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 24)))), vk_zero_point);
448 i3 += 4;
449
450 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
451 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
452 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32)))), vk_zero_point);
453 i4 += 4;
454
455 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
456 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
457 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 40)))), vk_zero_point);
458 i5 += 4;
459
460 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
461 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
462 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48)))), vk_zero_point);
463 i6 += 4;
464
465 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
466 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
467 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 56)))), vk_zero_point);
468 i7 += 4;
469
470 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
471 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
472 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64)))), vk_zero_point);
473 i8 += 4;
474
475 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
476 const __m128i vi9x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
477 const __m128i vk9x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 72)))), vk_zero_point);
478 i9 += 4;
479
480 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
481 const __m128i vi10x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
482 const __m128i vk10x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 80)))), vk_zero_point);
483 i10 += 4;
484
485 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
486 const __m128i vi11x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
487 const __m128i vk11x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 88)))), vk_zero_point);
488 i11 += 4;
489
490 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
491 const __m128i vi12x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
492 const __m128i vk12x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96)))), vk_zero_point);
493 i12 += 4;
494
495 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
496 const __m128i vi13x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
497 const __m128i vk13x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 104)))), vk_zero_point);
498 i13 += 4;
499
500 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
501 const __m128i vi14x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
502 const __m128i vk14x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 112)))), vk_zero_point);
503 i14 += 4;
504
505 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
506 const __m128i vi15x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
507 const __m128i vk15x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 120)))), vk_zero_point);
508 i15 += 4;
509
510 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
511 const __m128i vi16x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
512 const __m128i vk16x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 128)))), vk_zero_point);
513 i16 += 4;
514
515 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
516 const __m128i vi17x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
517 const __m128i vk17x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 136)))), vk_zero_point);
518 i17 += 4;
519
520 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
521 const __m128i vi18x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
522 const __m128i vk18x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144)))), vk_zero_point);
523 i18 += 4;
524
525 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
526 const __m128i vi19x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
527 const __m128i vk19x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 152)))), vk_zero_point);
528 i19 += 4;
529
530 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
531 const __m128i vi20x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
532 const __m128i vk20x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 160)))), vk_zero_point);
533 i20 += 4;
534
535 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
536 const __m128i vi21x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
537 const __m128i vk21x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 168)))), vk_zero_point);
538 i21 += 4;
539
540 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
541 const __m128i vi22x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
542 const __m128i vk22x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 176)))), vk_zero_point);
543 i22 += 4;
544
545 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
546 const __m128i vi23x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
547 const __m128i vk23x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 184)))), vk_zero_point);
548 i23 += 4;
549
550 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
551 const __m128i vi24x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
552 const __m128i vk24x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192)))), vk_zero_point);
553 i24 += 4;
554
555 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
556
557 k += 4;
558
559 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
560 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.scale));
561 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.output_max_less_zero_point));
562 vacc0123 = _mm_cvtps_epi32(vscaled0123);
563
564 w = (const void*) ((const int32_t*) w + 4);
565
566 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
567 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
568
569 vout0123 = _mm_packus_epi16(vout0123, vout0123);
570 vout0123 = _mm_max_epu8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
571
572 if XNN_LIKELY(c >= 4) {
573 _mm_storeu_si32(output, vout0123);
574 output += 4;
575 c -= 4;
576 } else {
577 if (c & 2) {
578 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
579 vout0123 = _mm_srli_epi32(vout0123, 16);
580 output += 2;
581 }
582 if (c & 1) {
583 *output = (uint8_t) _mm_extract_epi8(vout0123, 0);
584 output += 1;
585 }
586 c = 0;
587 }
588 } while (c != 0);
589 }
590
591 output = (uint8_t*) ((uintptr_t) output + output_increment);
592 } while (--output_width != 0);
593 }
594