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 #include <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_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)])19 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul32(
20 size_t channels,
21 size_t output_width,
22 const uint8_t** input,
23 const void* weights,
24 uint8_t* output,
25 size_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const uint8_t* zero,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(channels != 0);
32 assert(output_width != 0);
33
34 const __m128i vk_zero_point = _mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) params->fp32_sse2.kernel_zero_point));
35 do {
36 const uint8_t* i0 = input[0];
37 assert(i0 != NULL);
38 if XNN_UNPREDICTABLE(i0 != zero) {
39 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
40 }
41 const uint8_t* i1 = input[1];
42 assert(i1 != NULL);
43 if XNN_UNPREDICTABLE(i1 != zero) {
44 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
45 }
46 const uint8_t* i2 = input[2];
47 assert(i2 != NULL);
48 if XNN_UNPREDICTABLE(i2 != zero) {
49 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
50 }
51 const uint8_t* i3 = input[3];
52 assert(i3 != NULL);
53 if XNN_UNPREDICTABLE(i3 != zero) {
54 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
55 }
56 const uint8_t* i4 = input[4];
57 assert(i4 != NULL);
58 if XNN_UNPREDICTABLE(i4 != zero) {
59 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
60 }
61 const uint8_t* i5 = input[5];
62 assert(i5 != NULL);
63 if XNN_UNPREDICTABLE(i5 != zero) {
64 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
65 }
66 const uint8_t* i6 = input[6];
67 assert(i6 != NULL);
68 if XNN_UNPREDICTABLE(i6 != zero) {
69 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
70 }
71 const uint8_t* i7 = input[7];
72 assert(i7 != NULL);
73 if XNN_UNPREDICTABLE(i7 != zero) {
74 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
75 }
76 const uint8_t* i8 = input[8];
77 assert(i8 != NULL);
78 if XNN_UNPREDICTABLE(i8 != zero) {
79 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
80 }
81 const uint8_t* i9 = input[9];
82 assert(i9 != NULL);
83 if XNN_UNPREDICTABLE(i9 != zero) {
84 i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
85 }
86 const uint8_t* i10 = input[10];
87 assert(i10 != NULL);
88 if XNN_UNPREDICTABLE(i10 != zero) {
89 i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
90 }
91 const uint8_t* i11 = input[11];
92 assert(i11 != NULL);
93 if XNN_UNPREDICTABLE(i11 != zero) {
94 i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
95 }
96 const uint8_t* i12 = input[12];
97 assert(i12 != NULL);
98 if XNN_UNPREDICTABLE(i12 != zero) {
99 i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
100 }
101 const uint8_t* i13 = input[13];
102 assert(i13 != NULL);
103 if XNN_UNPREDICTABLE(i13 != zero) {
104 i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
105 }
106 const uint8_t* i14 = input[14];
107 assert(i14 != NULL);
108 if XNN_UNPREDICTABLE(i14 != zero) {
109 i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
110 }
111 const uint8_t* i15 = input[15];
112 assert(i15 != NULL);
113 if XNN_UNPREDICTABLE(i15 != zero) {
114 i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
115 }
116 const uint8_t* i16 = input[16];
117 assert(i16 != NULL);
118 if XNN_UNPREDICTABLE(i16 != zero) {
119 i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
120 }
121 const uint8_t* i17 = input[17];
122 assert(i17 != NULL);
123 if XNN_UNPREDICTABLE(i17 != zero) {
124 i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
125 }
126 const uint8_t* i18 = input[18];
127 assert(i18 != NULL);
128 if XNN_UNPREDICTABLE(i18 != zero) {
129 i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
130 }
131 const uint8_t* i19 = input[19];
132 assert(i19 != NULL);
133 if XNN_UNPREDICTABLE(i19 != zero) {
134 i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
135 }
136 const uint8_t* i20 = input[20];
137 assert(i20 != NULL);
138 if XNN_UNPREDICTABLE(i20 != zero) {
139 i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
140 }
141 const uint8_t* i21 = input[21];
142 assert(i21 != NULL);
143 if XNN_UNPREDICTABLE(i21 != zero) {
144 i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
145 }
146 const uint8_t* i22 = input[22];
147 assert(i22 != NULL);
148 if XNN_UNPREDICTABLE(i22 != zero) {
149 i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
150 }
151 const uint8_t* i23 = input[23];
152 assert(i23 != NULL);
153 if XNN_UNPREDICTABLE(i23 != zero) {
154 i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
155 }
156 const uint8_t* i24 = input[24];
157 assert(i24 != NULL);
158 if XNN_UNPREDICTABLE(i24 != zero) {
159 i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
160 }
161 input = (const uint8_t**) ((uintptr_t) input + input_stride);
162
163 size_t c = channels;
164 const void* w = weights;
165 for (; c >= 8; c -= 8) {
166 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
167 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
168
169
170 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
171 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);
172 const __m128i vi0x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
173 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);
174 i0 += 8;
175
176 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
177 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
178
179 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
180 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);
181 const __m128i vi1x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
182 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);
183 i1 += 8;
184
185 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
186 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
187
188 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
189 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);
190 const __m128i vi2x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
191 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);
192 i2 += 8;
193
194 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
195 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
196
197 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
198 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);
199 const __m128i vi3x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
200 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);
201 i3 += 8;
202
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
204 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
205
206 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
207 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);
208 const __m128i vi4x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
209 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);
210 i4 += 8;
211
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
214
215 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
216 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);
217 const __m128i vi5x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
218 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);
219 i5 += 8;
220
221 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
222 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
223
224 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
225 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);
226 const __m128i vi6x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
227 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);
228 i6 += 8;
229
230 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
231 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
232
233 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
234 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);
235 const __m128i vi7x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
236 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);
237 i7 += 8;
238
239 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
240 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
241
242 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
243 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);
244 const __m128i vi8x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
245 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);
246 i8 += 8;
247
248 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
249 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
250
251 const __m128i vi9x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
252 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);
253 const __m128i vi9x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 4)));
254 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);
255 i9 += 8;
256
257 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
258 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi9x4567, vk9x4567));
259
260 const __m128i vi10x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
261 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);
262 const __m128i vi10x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 4)));
263 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);
264 i10 += 8;
265
266 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
267 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi10x4567, vk10x4567));
268
269 const __m128i vi11x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
270 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);
271 const __m128i vi11x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 4)));
272 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);
273 i11 += 8;
274
275 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
276 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi11x4567, vk11x4567));
277
278 const __m128i vi12x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
279 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);
280 const __m128i vi12x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 4)));
281 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);
282 i12 += 8;
283
284 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
285 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi12x4567, vk12x4567));
286
287 const __m128i vi13x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
288 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);
289 const __m128i vi13x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 4)));
290 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);
291 i13 += 8;
292
293 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
294 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi13x4567, vk13x4567));
295
296 const __m128i vi14x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
297 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);
298 const __m128i vi14x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 4)));
299 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);
300 i14 += 8;
301
302 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
303 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi14x4567, vk14x4567));
304
305 const __m128i vi15x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
306 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);
307 const __m128i vi15x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 4)));
308 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);
309 i15 += 8;
310
311 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
312 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi15x4567, vk15x4567));
313
314 const __m128i vi16x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
315 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);
316 const __m128i vi16x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 4)));
317 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);
318 i16 += 8;
319
320 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
321 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi16x4567, vk16x4567));
322
323 const __m128i vi17x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
324 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);
325 const __m128i vi17x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 4)));
326 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);
327 i17 += 8;
328
329 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
330 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi17x4567, vk17x4567));
331
332 const __m128i vi18x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
333 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);
334 const __m128i vi18x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 4)));
335 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);
336 i18 += 8;
337
338 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
339 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi18x4567, vk18x4567));
340
341 const __m128i vi19x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
342 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);
343 const __m128i vi19x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 4)));
344 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);
345 i19 += 8;
346
347 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
348 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi19x4567, vk19x4567));
349
350 const __m128i vi20x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
351 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);
352 const __m128i vi20x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 4)));
353 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);
354 i20 += 8;
355
356 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
357 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi20x4567, vk20x4567));
358
359 const __m128i vi21x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
360 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);
361 const __m128i vi21x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 4)));
362 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);
363 i21 += 8;
364
365 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
366 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi21x4567, vk21x4567));
367
368 const __m128i vi22x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
369 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);
370 const __m128i vi22x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 4)));
371 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);
372 i22 += 8;
373
374 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
375 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi22x4567, vk22x4567));
376
377 const __m128i vi23x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
378 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);
379 const __m128i vi23x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 4)));
380 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);
381 i23 += 8;
382
383 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
384 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi23x4567, vk23x4567));
385
386 const __m128i vi24x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
387 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);
388 const __m128i vi24x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 4)));
389 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);
390 i24 += 8;
391
392 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
393 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi24x4567, vk24x4567));
394
395 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t));
396
397 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
398 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
399
400 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
401 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
402 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
403
404 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
405 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
406 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
407
408 vacc0123 = _mm_cvtps_epi32(vscaled0123);
409 vacc4567 = _mm_cvtps_epi32(vscaled4567);
410
411 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
412 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
413
414 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
415 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
416 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
417
418 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
419 output += 8;
420 }
421 if XNN_UNLIKELY(c != 0) {
422 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 8);
423 do {
424 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
425
426 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
427 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) k))), vk_zero_point);
428 i0 += 4;
429
430 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
431 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
432 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 8)))), vk_zero_point);
433 i1 += 4;
434
435 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
436 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
437 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16)))), vk_zero_point);
438 i2 += 4;
439
440 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
441 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
442 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 24)))), vk_zero_point);
443 i3 += 4;
444
445 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
446 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
447 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32)))), vk_zero_point);
448 i4 += 4;
449
450 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
451 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
452 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 40)))), vk_zero_point);
453 i5 += 4;
454
455 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
456 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
457 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48)))), vk_zero_point);
458 i6 += 4;
459
460 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
461 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
462 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 56)))), vk_zero_point);
463 i7 += 4;
464
465 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
466 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
467 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64)))), vk_zero_point);
468 i8 += 4;
469
470 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
471 const __m128i vi9x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
472 const __m128i vk9x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 72)))), vk_zero_point);
473 i9 += 4;
474
475 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
476 const __m128i vi10x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
477 const __m128i vk10x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 80)))), vk_zero_point);
478 i10 += 4;
479
480 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
481 const __m128i vi11x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
482 const __m128i vk11x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 88)))), vk_zero_point);
483 i11 += 4;
484
485 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
486 const __m128i vi12x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
487 const __m128i vk12x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96)))), vk_zero_point);
488 i12 += 4;
489
490 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
491 const __m128i vi13x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
492 const __m128i vk13x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 104)))), vk_zero_point);
493 i13 += 4;
494
495 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
496 const __m128i vi14x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
497 const __m128i vk14x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 112)))), vk_zero_point);
498 i14 += 4;
499
500 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
501 const __m128i vi15x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
502 const __m128i vk15x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 120)))), vk_zero_point);
503 i15 += 4;
504
505 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
506 const __m128i vi16x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
507 const __m128i vk16x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 128)))), vk_zero_point);
508 i16 += 4;
509
510 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
511 const __m128i vi17x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
512 const __m128i vk17x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 136)))), vk_zero_point);
513 i17 += 4;
514
515 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
516 const __m128i vi18x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
517 const __m128i vk18x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144)))), vk_zero_point);
518 i18 += 4;
519
520 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
521 const __m128i vi19x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
522 const __m128i vk19x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 152)))), vk_zero_point);
523 i19 += 4;
524
525 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
526 const __m128i vi20x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
527 const __m128i vk20x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 160)))), vk_zero_point);
528 i20 += 4;
529
530 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
531 const __m128i vi21x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
532 const __m128i vk21x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 168)))), vk_zero_point);
533 i21 += 4;
534
535 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
536 const __m128i vi22x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
537 const __m128i vk22x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 176)))), vk_zero_point);
538 i22 += 4;
539
540 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
541 const __m128i vi23x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
542 const __m128i vk23x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 184)))), vk_zero_point);
543 i23 += 4;
544
545 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
546 const __m128i vi24x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
547 const __m128i vk24x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192)))), vk_zero_point);
548 i24 += 4;
549
550 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
551
552 k += 4;
553
554 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
555 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.scale));
556 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.output_max_less_zero_point));
557 vacc0123 = _mm_cvtps_epi32(vscaled0123);
558
559 w = (const void*) ((const int32_t*) w + 4);
560
561 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
562 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
563
564 vout0123 = _mm_packus_epi16(vout0123, vout0123);
565 vout0123 = _mm_max_epu8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
566
567 if XNN_LIKELY(c >= 4) {
568 _mm_storeu_si32(output, vout0123);
569 output += 4;
570 c -= 4;
571 } else {
572 if (c & 2) {
573 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
574 vout0123 = _mm_srli_epi32(vout0123, 16);
575 output += 2;
576 }
577 if (c & 1) {
578 *output = (uint8_t) _mm_extract_epi8(vout0123, 0);
579 output += 1;
580 }
581 c = 0;
582 }
583 } while (c != 0);
584 }
585
586 output = (uint8_t*) ((uintptr_t) output + output_increment);
587 } while (--output_width != 0);
588 }
589