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_up16x25__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_up16x25__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 >= 16; c -= 16) {
171 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
172 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
173 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
174 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
175
176
177 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
178 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t))))), vk_zero_point);
179 const __m128i vi0x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
180 const __m128i vk0x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(uint8_t))))), vk_zero_point);
181 const __m128i vi0x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
182 const __m128i vk0x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t))))), vk_zero_point);
183 const __m128i vi0xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
184 const __m128i vk0xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(uint8_t))))), vk_zero_point);
185 i0 += 16;
186
187 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
188 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
189 vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
190 vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
191
192 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
193 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t))))), vk_zero_point);
194 const __m128i vi1x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
195 const __m128i vk1x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(uint8_t))))), vk_zero_point);
196 const __m128i vi1x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
197 const __m128i vk1x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t))))), vk_zero_point);
198 const __m128i vi1xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
199 const __m128i vk1xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(uint8_t))))), vk_zero_point);
200 i1 += 16;
201
202 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
203 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
204 vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
205 vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
206
207 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
208 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t))))), vk_zero_point);
209 const __m128i vi2x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
210 const __m128i vk2x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(uint8_t))))), vk_zero_point);
211 const __m128i vi2x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
212 const __m128i vk2x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t))))), vk_zero_point);
213 const __m128i vi2xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
214 const __m128i vk2xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(uint8_t))))), vk_zero_point);
215 i2 += 16;
216
217 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
218 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
219 vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
220 vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
221
222 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
223 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t))))), vk_zero_point);
224 const __m128i vi3x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
225 const __m128i vk3x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(uint8_t))))), vk_zero_point);
226 const __m128i vi3x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
227 const __m128i vk3x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t))))), vk_zero_point);
228 const __m128i vi3xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
229 const __m128i vk3xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(uint8_t))))), vk_zero_point);
230 i3 += 16;
231
232 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
233 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
234 vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
235 vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
236
237 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
238 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t))))), vk_zero_point);
239 const __m128i vi4x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
240 const __m128i vk4x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(uint8_t))))), vk_zero_point);
241 const __m128i vi4x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
242 const __m128i vk4x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t))))), vk_zero_point);
243 const __m128i vi4xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
244 const __m128i vk4xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(uint8_t))))), vk_zero_point);
245 i4 += 16;
246
247 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
248 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
249 vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
250 vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
251
252 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
253 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t))))), vk_zero_point);
254 const __m128i vi5x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
255 const __m128i vk5x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(uint8_t))))), vk_zero_point);
256 const __m128i vi5x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
257 const __m128i vk5x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t))))), vk_zero_point);
258 const __m128i vi5xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
259 const __m128i vk5xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(uint8_t))))), vk_zero_point);
260 i5 += 16;
261
262 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
263 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
264 vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
265 vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
266
267 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
268 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t))))), vk_zero_point);
269 const __m128i vi6x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
270 const __m128i vk6x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(uint8_t))))), vk_zero_point);
271 const __m128i vi6x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
272 const __m128i vk6x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t))))), vk_zero_point);
273 const __m128i vi6xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
274 const __m128i vk6xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(uint8_t))))), vk_zero_point);
275 i6 += 16;
276
277 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
278 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
279 vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
280 vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
281
282 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
283 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t))))), vk_zero_point);
284 const __m128i vi7x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
285 const __m128i vk7x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(uint8_t))))), vk_zero_point);
286 const __m128i vi7x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
287 const __m128i vk7x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t))))), vk_zero_point);
288 const __m128i vi7xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
289 const __m128i vk7xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(uint8_t))))), vk_zero_point);
290 i7 += 16;
291
292 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
293 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
294 vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
295 vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
296
297 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
298 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t))))), vk_zero_point);
299 const __m128i vi8x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
300 const __m128i vk8x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(uint8_t))))), vk_zero_point);
301 const __m128i vi8x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
302 const __m128i vk8x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t))))), vk_zero_point);
303 const __m128i vi8xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
304 const __m128i vk8xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(uint8_t))))), vk_zero_point);
305 i8 += 16;
306
307 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
308 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
309 vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
310 vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
311
312 const __m128i vi9x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
313 const __m128i vk9x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t))))), vk_zero_point);
314 const __m128i vi9x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 4)));
315 const __m128i vk9x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 148 * sizeof(uint8_t))))), vk_zero_point);
316 const __m128i vi9x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 8)));
317 const __m128i vk9x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 152 * sizeof(uint8_t))))), vk_zero_point);
318 const __m128i vi9xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 12)));
319 const __m128i vk9xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 156 * sizeof(uint8_t))))), vk_zero_point);
320 i9 += 16;
321
322 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
323 vacc4567 = _mm_macc_epi32(vi9x4567, vk9x4567, vacc4567);
324 vacc89AB = _mm_macc_epi32(vi9x89AB, vk9x89AB, vacc89AB);
325 vaccCDEF = _mm_macc_epi32(vi9xCDEF, vk9xCDEF, vaccCDEF);
326
327 const __m128i vi10x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
328 const __m128i vk10x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 160 * sizeof(uint8_t))))), vk_zero_point);
329 const __m128i vi10x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 4)));
330 const __m128i vk10x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 164 * sizeof(uint8_t))))), vk_zero_point);
331 const __m128i vi10x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 8)));
332 const __m128i vk10x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 168 * sizeof(uint8_t))))), vk_zero_point);
333 const __m128i vi10xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 12)));
334 const __m128i vk10xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 172 * sizeof(uint8_t))))), vk_zero_point);
335 i10 += 16;
336
337 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
338 vacc4567 = _mm_macc_epi32(vi10x4567, vk10x4567, vacc4567);
339 vacc89AB = _mm_macc_epi32(vi10x89AB, vk10x89AB, vacc89AB);
340 vaccCDEF = _mm_macc_epi32(vi10xCDEF, vk10xCDEF, vaccCDEF);
341
342 const __m128i vi11x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
343 const __m128i vk11x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 176 * sizeof(uint8_t))))), vk_zero_point);
344 const __m128i vi11x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 4)));
345 const __m128i vk11x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 180 * sizeof(uint8_t))))), vk_zero_point);
346 const __m128i vi11x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 8)));
347 const __m128i vk11x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 184 * sizeof(uint8_t))))), vk_zero_point);
348 const __m128i vi11xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 12)));
349 const __m128i vk11xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 188 * sizeof(uint8_t))))), vk_zero_point);
350 i11 += 16;
351
352 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
353 vacc4567 = _mm_macc_epi32(vi11x4567, vk11x4567, vacc4567);
354 vacc89AB = _mm_macc_epi32(vi11x89AB, vk11x89AB, vacc89AB);
355 vaccCDEF = _mm_macc_epi32(vi11xCDEF, vk11xCDEF, vaccCDEF);
356
357 const __m128i vi12x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
358 const __m128i vk12x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 192 * sizeof(uint8_t))))), vk_zero_point);
359 const __m128i vi12x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 4)));
360 const __m128i vk12x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 196 * sizeof(uint8_t))))), vk_zero_point);
361 const __m128i vi12x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 8)));
362 const __m128i vk12x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 200 * sizeof(uint8_t))))), vk_zero_point);
363 const __m128i vi12xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 12)));
364 const __m128i vk12xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 204 * sizeof(uint8_t))))), vk_zero_point);
365 i12 += 16;
366
367 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
368 vacc4567 = _mm_macc_epi32(vi12x4567, vk12x4567, vacc4567);
369 vacc89AB = _mm_macc_epi32(vi12x89AB, vk12x89AB, vacc89AB);
370 vaccCDEF = _mm_macc_epi32(vi12xCDEF, vk12xCDEF, vaccCDEF);
371
372 const __m128i vi13x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
373 const __m128i vk13x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 208 * sizeof(uint8_t))))), vk_zero_point);
374 const __m128i vi13x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 4)));
375 const __m128i vk13x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 212 * sizeof(uint8_t))))), vk_zero_point);
376 const __m128i vi13x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 8)));
377 const __m128i vk13x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 216 * sizeof(uint8_t))))), vk_zero_point);
378 const __m128i vi13xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 12)));
379 const __m128i vk13xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 220 * sizeof(uint8_t))))), vk_zero_point);
380 i13 += 16;
381
382 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
383 vacc4567 = _mm_macc_epi32(vi13x4567, vk13x4567, vacc4567);
384 vacc89AB = _mm_macc_epi32(vi13x89AB, vk13x89AB, vacc89AB);
385 vaccCDEF = _mm_macc_epi32(vi13xCDEF, vk13xCDEF, vaccCDEF);
386
387 const __m128i vi14x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
388 const __m128i vk14x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 224 * sizeof(uint8_t))))), vk_zero_point);
389 const __m128i vi14x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 4)));
390 const __m128i vk14x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 228 * sizeof(uint8_t))))), vk_zero_point);
391 const __m128i vi14x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 8)));
392 const __m128i vk14x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 232 * sizeof(uint8_t))))), vk_zero_point);
393 const __m128i vi14xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 12)));
394 const __m128i vk14xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 236 * sizeof(uint8_t))))), vk_zero_point);
395 i14 += 16;
396
397 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
398 vacc4567 = _mm_macc_epi32(vi14x4567, vk14x4567, vacc4567);
399 vacc89AB = _mm_macc_epi32(vi14x89AB, vk14x89AB, vacc89AB);
400 vaccCDEF = _mm_macc_epi32(vi14xCDEF, vk14xCDEF, vaccCDEF);
401
402 const __m128i vi15x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
403 const __m128i vk15x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 240 * sizeof(uint8_t))))), vk_zero_point);
404 const __m128i vi15x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 4)));
405 const __m128i vk15x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 244 * sizeof(uint8_t))))), vk_zero_point);
406 const __m128i vi15x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 8)));
407 const __m128i vk15x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 248 * sizeof(uint8_t))))), vk_zero_point);
408 const __m128i vi15xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 12)));
409 const __m128i vk15xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 252 * sizeof(uint8_t))))), vk_zero_point);
410 i15 += 16;
411
412 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
413 vacc4567 = _mm_macc_epi32(vi15x4567, vk15x4567, vacc4567);
414 vacc89AB = _mm_macc_epi32(vi15x89AB, vk15x89AB, vacc89AB);
415 vaccCDEF = _mm_macc_epi32(vi15xCDEF, vk15xCDEF, vaccCDEF);
416
417 const __m128i vi16x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
418 const __m128i vk16x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 256 * sizeof(uint8_t))))), vk_zero_point);
419 const __m128i vi16x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 4)));
420 const __m128i vk16x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 260 * sizeof(uint8_t))))), vk_zero_point);
421 const __m128i vi16x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 8)));
422 const __m128i vk16x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 264 * sizeof(uint8_t))))), vk_zero_point);
423 const __m128i vi16xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 12)));
424 const __m128i vk16xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 268 * sizeof(uint8_t))))), vk_zero_point);
425 i16 += 16;
426
427 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
428 vacc4567 = _mm_macc_epi32(vi16x4567, vk16x4567, vacc4567);
429 vacc89AB = _mm_macc_epi32(vi16x89AB, vk16x89AB, vacc89AB);
430 vaccCDEF = _mm_macc_epi32(vi16xCDEF, vk16xCDEF, vaccCDEF);
431
432 const __m128i vi17x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
433 const __m128i vk17x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 272 * sizeof(uint8_t))))), vk_zero_point);
434 const __m128i vi17x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 4)));
435 const __m128i vk17x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 276 * sizeof(uint8_t))))), vk_zero_point);
436 const __m128i vi17x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 8)));
437 const __m128i vk17x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 280 * sizeof(uint8_t))))), vk_zero_point);
438 const __m128i vi17xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 12)));
439 const __m128i vk17xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 284 * sizeof(uint8_t))))), vk_zero_point);
440 i17 += 16;
441
442 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
443 vacc4567 = _mm_macc_epi32(vi17x4567, vk17x4567, vacc4567);
444 vacc89AB = _mm_macc_epi32(vi17x89AB, vk17x89AB, vacc89AB);
445 vaccCDEF = _mm_macc_epi32(vi17xCDEF, vk17xCDEF, vaccCDEF);
446
447 const __m128i vi18x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
448 const __m128i vk18x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 288 * sizeof(uint8_t))))), vk_zero_point);
449 const __m128i vi18x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 4)));
450 const __m128i vk18x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 292 * sizeof(uint8_t))))), vk_zero_point);
451 const __m128i vi18x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 8)));
452 const __m128i vk18x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 296 * sizeof(uint8_t))))), vk_zero_point);
453 const __m128i vi18xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 12)));
454 const __m128i vk18xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 300 * sizeof(uint8_t))))), vk_zero_point);
455 i18 += 16;
456
457 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
458 vacc4567 = _mm_macc_epi32(vi18x4567, vk18x4567, vacc4567);
459 vacc89AB = _mm_macc_epi32(vi18x89AB, vk18x89AB, vacc89AB);
460 vaccCDEF = _mm_macc_epi32(vi18xCDEF, vk18xCDEF, vaccCDEF);
461
462 const __m128i vi19x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
463 const __m128i vk19x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 304 * sizeof(uint8_t))))), vk_zero_point);
464 const __m128i vi19x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 4)));
465 const __m128i vk19x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 308 * sizeof(uint8_t))))), vk_zero_point);
466 const __m128i vi19x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 8)));
467 const __m128i vk19x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 312 * sizeof(uint8_t))))), vk_zero_point);
468 const __m128i vi19xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 12)));
469 const __m128i vk19xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 316 * sizeof(uint8_t))))), vk_zero_point);
470 i19 += 16;
471
472 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
473 vacc4567 = _mm_macc_epi32(vi19x4567, vk19x4567, vacc4567);
474 vacc89AB = _mm_macc_epi32(vi19x89AB, vk19x89AB, vacc89AB);
475 vaccCDEF = _mm_macc_epi32(vi19xCDEF, vk19xCDEF, vaccCDEF);
476
477 const __m128i vi20x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
478 const __m128i vk20x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 320 * sizeof(uint8_t))))), vk_zero_point);
479 const __m128i vi20x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 4)));
480 const __m128i vk20x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 324 * sizeof(uint8_t))))), vk_zero_point);
481 const __m128i vi20x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 8)));
482 const __m128i vk20x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 328 * sizeof(uint8_t))))), vk_zero_point);
483 const __m128i vi20xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 12)));
484 const __m128i vk20xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 332 * sizeof(uint8_t))))), vk_zero_point);
485 i20 += 16;
486
487 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
488 vacc4567 = _mm_macc_epi32(vi20x4567, vk20x4567, vacc4567);
489 vacc89AB = _mm_macc_epi32(vi20x89AB, vk20x89AB, vacc89AB);
490 vaccCDEF = _mm_macc_epi32(vi20xCDEF, vk20xCDEF, vaccCDEF);
491
492 const __m128i vi21x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
493 const __m128i vk21x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 336 * sizeof(uint8_t))))), vk_zero_point);
494 const __m128i vi21x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 4)));
495 const __m128i vk21x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 340 * sizeof(uint8_t))))), vk_zero_point);
496 const __m128i vi21x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 8)));
497 const __m128i vk21x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 344 * sizeof(uint8_t))))), vk_zero_point);
498 const __m128i vi21xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 12)));
499 const __m128i vk21xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 348 * sizeof(uint8_t))))), vk_zero_point);
500 i21 += 16;
501
502 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
503 vacc4567 = _mm_macc_epi32(vi21x4567, vk21x4567, vacc4567);
504 vacc89AB = _mm_macc_epi32(vi21x89AB, vk21x89AB, vacc89AB);
505 vaccCDEF = _mm_macc_epi32(vi21xCDEF, vk21xCDEF, vaccCDEF);
506
507 const __m128i vi22x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
508 const __m128i vk22x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 352 * sizeof(uint8_t))))), vk_zero_point);
509 const __m128i vi22x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 4)));
510 const __m128i vk22x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 356 * sizeof(uint8_t))))), vk_zero_point);
511 const __m128i vi22x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 8)));
512 const __m128i vk22x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 360 * sizeof(uint8_t))))), vk_zero_point);
513 const __m128i vi22xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 12)));
514 const __m128i vk22xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 364 * sizeof(uint8_t))))), vk_zero_point);
515 i22 += 16;
516
517 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
518 vacc4567 = _mm_macc_epi32(vi22x4567, vk22x4567, vacc4567);
519 vacc89AB = _mm_macc_epi32(vi22x89AB, vk22x89AB, vacc89AB);
520 vaccCDEF = _mm_macc_epi32(vi22xCDEF, vk22xCDEF, vaccCDEF);
521
522 const __m128i vi23x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
523 const __m128i vk23x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 368 * sizeof(uint8_t))))), vk_zero_point);
524 const __m128i vi23x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 4)));
525 const __m128i vk23x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 372 * sizeof(uint8_t))))), vk_zero_point);
526 const __m128i vi23x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 8)));
527 const __m128i vk23x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 376 * sizeof(uint8_t))))), vk_zero_point);
528 const __m128i vi23xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 12)));
529 const __m128i vk23xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 380 * sizeof(uint8_t))))), vk_zero_point);
530 i23 += 16;
531
532 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
533 vacc4567 = _mm_macc_epi32(vi23x4567, vk23x4567, vacc4567);
534 vacc89AB = _mm_macc_epi32(vi23x89AB, vk23x89AB, vacc89AB);
535 vaccCDEF = _mm_macc_epi32(vi23xCDEF, vk23xCDEF, vaccCDEF);
536
537 const __m128i vi24x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
538 const __m128i vk24x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 384 * sizeof(uint8_t))))), vk_zero_point);
539 const __m128i vi24x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 4)));
540 const __m128i vk24x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 388 * sizeof(uint8_t))))), vk_zero_point);
541 const __m128i vi24x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 8)));
542 const __m128i vk24x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 392 * sizeof(uint8_t))))), vk_zero_point);
543 const __m128i vi24xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 12)));
544 const __m128i vk24xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 396 * sizeof(uint8_t))))), vk_zero_point);
545 i24 += 16;
546
547 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
548 vacc4567 = _mm_macc_epi32(vi24x4567, vk24x4567, vacc4567);
549 vacc89AB = _mm_macc_epi32(vi24x89AB, vk24x89AB, vacc89AB);
550 vaccCDEF = _mm_macc_epi32(vi24xCDEF, vk24xCDEF, vaccCDEF);
551
552 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 400 * sizeof(uint8_t));
553
554 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
555 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
556 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
557 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
558
559 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
560 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
561 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
562 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
563 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
564
565 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
566 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
567 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
568 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
569 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
570
571 vacc0123 = _mm_cvtps_epi32(vscaled0123);
572 vacc4567 = _mm_cvtps_epi32(vscaled4567);
573 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
574 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
575
576 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
577 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
578 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
579
580 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
581 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
582 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
583
584 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
585 output += 16;
586 }
587 if XNN_UNLIKELY(c != 0) {
588 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
589 do {
590 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
591
592 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
593 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) k))), vk_zero_point);
594 i0 += 4;
595
596 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
597 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
598 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16)))), vk_zero_point);
599 i1 += 4;
600
601 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
602 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
603 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32)))), vk_zero_point);
604 i2 += 4;
605
606 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
607 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
608 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48)))), vk_zero_point);
609 i3 += 4;
610
611 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
612 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
613 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64)))), vk_zero_point);
614 i4 += 4;
615
616 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
617 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
618 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 80)))), vk_zero_point);
619 i5 += 4;
620
621 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
622 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
623 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96)))), vk_zero_point);
624 i6 += 4;
625
626 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
627 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
628 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 112)))), vk_zero_point);
629 i7 += 4;
630
631 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
632 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
633 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 128)))), vk_zero_point);
634 i8 += 4;
635
636 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
637 const __m128i vi9x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
638 const __m128i vk9x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144)))), vk_zero_point);
639 i9 += 4;
640
641 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
642 const __m128i vi10x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
643 const __m128i vk10x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 160)))), vk_zero_point);
644 i10 += 4;
645
646 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
647 const __m128i vi11x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
648 const __m128i vk11x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 176)))), vk_zero_point);
649 i11 += 4;
650
651 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
652 const __m128i vi12x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
653 const __m128i vk12x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192)))), vk_zero_point);
654 i12 += 4;
655
656 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
657 const __m128i vi13x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
658 const __m128i vk13x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 208)))), vk_zero_point);
659 i13 += 4;
660
661 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
662 const __m128i vi14x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
663 const __m128i vk14x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 224)))), vk_zero_point);
664 i14 += 4;
665
666 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
667 const __m128i vi15x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
668 const __m128i vk15x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 240)))), vk_zero_point);
669 i15 += 4;
670
671 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
672 const __m128i vi16x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
673 const __m128i vk16x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 256)))), vk_zero_point);
674 i16 += 4;
675
676 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
677 const __m128i vi17x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
678 const __m128i vk17x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 272)))), vk_zero_point);
679 i17 += 4;
680
681 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
682 const __m128i vi18x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
683 const __m128i vk18x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 288)))), vk_zero_point);
684 i18 += 4;
685
686 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
687 const __m128i vi19x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
688 const __m128i vk19x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 304)))), vk_zero_point);
689 i19 += 4;
690
691 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
692 const __m128i vi20x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
693 const __m128i vk20x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 320)))), vk_zero_point);
694 i20 += 4;
695
696 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
697 const __m128i vi21x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
698 const __m128i vk21x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 336)))), vk_zero_point);
699 i21 += 4;
700
701 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
702 const __m128i vi22x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
703 const __m128i vk22x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 352)))), vk_zero_point);
704 i22 += 4;
705
706 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
707 const __m128i vi23x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
708 const __m128i vk23x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 368)))), vk_zero_point);
709 i23 += 4;
710
711 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
712 const __m128i vi24x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
713 const __m128i vk24x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 384)))), vk_zero_point);
714 i24 += 4;
715
716 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
717
718 k += 4;
719
720 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
721 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.scale));
722 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.output_max_less_zero_point));
723 vacc0123 = _mm_cvtps_epi32(vscaled0123);
724
725 w = (const void*) ((const int32_t*) w + 4);
726
727 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
728 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
729
730 vout0123 = _mm_packus_epi16(vout0123, vout0123);
731 vout0123 = _mm_max_epu8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
732
733 if XNN_LIKELY(c >= 4) {
734 _mm_storeu_si32(output, vout0123);
735 output += 4;
736 c -= 4;
737 } else {
738 if (c & 2) {
739 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
740 vout0123 = _mm_srli_epi32(vout0123, 16);
741 output += 2;
742 }
743 if (c & 1) {
744 *output = (uint8_t) _mm_extract_epi8(vout0123, 0);
745 output += 1;
746 }
747 c = 0;
748 }
749 } while (c != 0);
750 }
751
752 output = (uint8_t*) ((uintptr_t) output + output_increment);
753 } while (--output_width != 0);
754 }
755