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
17
xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__avx_mul32(
19 size_t channels,
20 size_t output_width,
21 const uint8_t** input,
22 const void* weights,
23 uint8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m128i vk_zero_point = _mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) params->fp32_sse2.kernel_zero_point));
34 do {
35 const uint8_t* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
39 }
40 const uint8_t* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
44 }
45 const uint8_t* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
49 }
50 const uint8_t* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
54 }
55 const uint8_t* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
59 }
60 const uint8_t* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
64 }
65 const uint8_t* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
69 }
70 const uint8_t* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
74 }
75 const uint8_t* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
79 }
80 const uint8_t* i9 = input[9];
81 assert(i9 != NULL);
82 if XNN_UNPREDICTABLE(i9 != zero) {
83 i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
84 }
85 const uint8_t* i10 = input[10];
86 assert(i10 != NULL);
87 if XNN_UNPREDICTABLE(i10 != zero) {
88 i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
89 }
90 const uint8_t* i11 = input[11];
91 assert(i11 != NULL);
92 if XNN_UNPREDICTABLE(i11 != zero) {
93 i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
94 }
95 const uint8_t* i12 = input[12];
96 assert(i12 != NULL);
97 if XNN_UNPREDICTABLE(i12 != zero) {
98 i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
99 }
100 const uint8_t* i13 = input[13];
101 assert(i13 != NULL);
102 if XNN_UNPREDICTABLE(i13 != zero) {
103 i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
104 }
105 const uint8_t* i14 = input[14];
106 assert(i14 != NULL);
107 if XNN_UNPREDICTABLE(i14 != zero) {
108 i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
109 }
110 const uint8_t* i15 = input[15];
111 assert(i15 != NULL);
112 if XNN_UNPREDICTABLE(i15 != zero) {
113 i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
114 }
115 const uint8_t* i16 = input[16];
116 assert(i16 != NULL);
117 if XNN_UNPREDICTABLE(i16 != zero) {
118 i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
119 }
120 const uint8_t* i17 = input[17];
121 assert(i17 != NULL);
122 if XNN_UNPREDICTABLE(i17 != zero) {
123 i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
124 }
125 const uint8_t* i18 = input[18];
126 assert(i18 != NULL);
127 if XNN_UNPREDICTABLE(i18 != zero) {
128 i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
129 }
130 const uint8_t* i19 = input[19];
131 assert(i19 != NULL);
132 if XNN_UNPREDICTABLE(i19 != zero) {
133 i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
134 }
135 const uint8_t* i20 = input[20];
136 assert(i20 != NULL);
137 if XNN_UNPREDICTABLE(i20 != zero) {
138 i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
139 }
140 const uint8_t* i21 = input[21];
141 assert(i21 != NULL);
142 if XNN_UNPREDICTABLE(i21 != zero) {
143 i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
144 }
145 const uint8_t* i22 = input[22];
146 assert(i22 != NULL);
147 if XNN_UNPREDICTABLE(i22 != zero) {
148 i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
149 }
150 const uint8_t* i23 = input[23];
151 assert(i23 != NULL);
152 if XNN_UNPREDICTABLE(i23 != zero) {
153 i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
154 }
155 const uint8_t* i24 = input[24];
156 assert(i24 != NULL);
157 if XNN_UNPREDICTABLE(i24 != zero) {
158 i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
159 }
160 input = (const uint8_t**) ((uintptr_t) input + input_stride);
161
162 size_t c = channels;
163 const void* w = weights;
164 for (; c >= 8; c -= 8) {
165 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
166 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
167
168
169 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i0));
170 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
171 const __m128i vi0x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i0 + 4));
172 const __m128i vk0x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(uint8_t)))), vk_zero_point);
173 i0 += 8;
174
175 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
176 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
177
178 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i1));
179 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)))), vk_zero_point);
180 const __m128i vi1x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i1 + 4));
181 const __m128i vk1x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(uint8_t)))), vk_zero_point);
182 i1 += 8;
183
184 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
185 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
186
187 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i2));
188 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
189 const __m128i vi2x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i2 + 4));
190 const __m128i vk2x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(uint8_t)))), vk_zero_point);
191 i2 += 8;
192
193 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
194 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
195
196 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i3));
197 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)))), vk_zero_point);
198 const __m128i vi3x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i3 + 4));
199 const __m128i vk3x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(uint8_t)))), vk_zero_point);
200 i3 += 8;
201
202 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
203 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
204
205 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i4));
206 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
207 const __m128i vi4x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i4 + 4));
208 const __m128i vk4x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(uint8_t)))), vk_zero_point);
209 i4 += 8;
210
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
213
214 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i5));
215 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)))), vk_zero_point);
216 const __m128i vi5x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i5 + 4));
217 const __m128i vk5x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(uint8_t)))), vk_zero_point);
218 i5 += 8;
219
220 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
221 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
222
223 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i6));
224 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
225 const __m128i vi6x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i6 + 4));
226 const __m128i vk6x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 52 * sizeof(uint8_t)))), vk_zero_point);
227 i6 += 8;
228
229 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
230 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
231
232 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i7));
233 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)))), vk_zero_point);
234 const __m128i vi7x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i7 + 4));
235 const __m128i vk7x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 60 * sizeof(uint8_t)))), vk_zero_point);
236 i7 += 8;
237
238 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
239 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
240
241 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i8));
242 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
243 const __m128i vi8x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i8 + 4));
244 const __m128i vk8x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 68 * sizeof(uint8_t)))), vk_zero_point);
245 i8 += 8;
246
247 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
248 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
249
250 const __m128i vi9x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i9));
251 const __m128i vk9x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)))), vk_zero_point);
252 const __m128i vi9x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i9 + 4));
253 const __m128i vk9x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 76 * sizeof(uint8_t)))), vk_zero_point);
254 i9 += 8;
255
256 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
257 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi9x4567, vk9x4567));
258
259 const __m128i vi10x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i10));
260 const __m128i vk10x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
261 const __m128i vi10x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i10 + 4));
262 const __m128i vk10x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 84 * sizeof(uint8_t)))), vk_zero_point);
263 i10 += 8;
264
265 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
266 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi10x4567, vk10x4567));
267
268 const __m128i vi11x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i11));
269 const __m128i vk11x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t)))), vk_zero_point);
270 const __m128i vi11x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i11 + 4));
271 const __m128i vk11x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 92 * sizeof(uint8_t)))), vk_zero_point);
272 i11 += 8;
273
274 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
275 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi11x4567, vk11x4567));
276
277 const __m128i vi12x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i12));
278 const __m128i vk12x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
279 const __m128i vi12x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i12 + 4));
280 const __m128i vk12x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 100 * sizeof(uint8_t)))), vk_zero_point);
281 i12 += 8;
282
283 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
284 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi12x4567, vk12x4567));
285
286 const __m128i vi13x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i13));
287 const __m128i vk13x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t)))), vk_zero_point);
288 const __m128i vi13x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i13 + 4));
289 const __m128i vk13x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 108 * sizeof(uint8_t)))), vk_zero_point);
290 i13 += 8;
291
292 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
293 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi13x4567, vk13x4567));
294
295 const __m128i vi14x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i14));
296 const __m128i vk14x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
297 const __m128i vi14x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i14 + 4));
298 const __m128i vk14x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 116 * sizeof(uint8_t)))), vk_zero_point);
299 i14 += 8;
300
301 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
302 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi14x4567, vk14x4567));
303
304 const __m128i vi15x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i15));
305 const __m128i vk15x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t)))), vk_zero_point);
306 const __m128i vi15x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i15 + 4));
307 const __m128i vk15x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 124 * sizeof(uint8_t)))), vk_zero_point);
308 i15 += 8;
309
310 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
311 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi15x4567, vk15x4567));
312
313 const __m128i vi16x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i16));
314 const __m128i vk16x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
315 const __m128i vi16x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i16 + 4));
316 const __m128i vk16x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 132 * sizeof(uint8_t)))), vk_zero_point);
317 i16 += 8;
318
319 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
320 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi16x4567, vk16x4567));
321
322 const __m128i vi17x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i17));
323 const __m128i vk17x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t)))), vk_zero_point);
324 const __m128i vi17x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i17 + 4));
325 const __m128i vk17x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 140 * sizeof(uint8_t)))), vk_zero_point);
326 i17 += 8;
327
328 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
329 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi17x4567, vk17x4567));
330
331 const __m128i vi18x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i18));
332 const __m128i vk18x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t)))), vk_zero_point);
333 const __m128i vi18x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i18 + 4));
334 const __m128i vk18x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 148 * sizeof(uint8_t)))), vk_zero_point);
335 i18 += 8;
336
337 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
338 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi18x4567, vk18x4567));
339
340 const __m128i vi19x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i19));
341 const __m128i vk19x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t)))), vk_zero_point);
342 const __m128i vi19x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i19 + 4));
343 const __m128i vk19x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 156 * sizeof(uint8_t)))), vk_zero_point);
344 i19 += 8;
345
346 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
347 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi19x4567, vk19x4567));
348
349 const __m128i vi20x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i20));
350 const __m128i vk20x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t)))), vk_zero_point);
351 const __m128i vi20x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i20 + 4));
352 const __m128i vk20x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 164 * sizeof(uint8_t)))), vk_zero_point);
353 i20 += 8;
354
355 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
356 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi20x4567, vk20x4567));
357
358 const __m128i vi21x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i21));
359 const __m128i vk21x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t)))), vk_zero_point);
360 const __m128i vi21x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i21 + 4));
361 const __m128i vk21x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 172 * sizeof(uint8_t)))), vk_zero_point);
362 i21 += 8;
363
364 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
365 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi21x4567, vk21x4567));
366
367 const __m128i vi22x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i22));
368 const __m128i vk22x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t)))), vk_zero_point);
369 const __m128i vi22x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i22 + 4));
370 const __m128i vk22x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 180 * sizeof(uint8_t)))), vk_zero_point);
371 i22 += 8;
372
373 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
374 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi22x4567, vk22x4567));
375
376 const __m128i vi23x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i23));
377 const __m128i vk23x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t)))), vk_zero_point);
378 const __m128i vi23x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i23 + 4));
379 const __m128i vk23x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 188 * sizeof(uint8_t)))), vk_zero_point);
380 i23 += 8;
381
382 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
383 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi23x4567, vk23x4567));
384
385 const __m128i vi24x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i24));
386 const __m128i vk24x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t)))), vk_zero_point);
387 const __m128i vi24x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i24 + 4));
388 const __m128i vk24x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 196 * sizeof(uint8_t)))), vk_zero_point);
389 i24 += 8;
390
391 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
392 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi24x4567, vk24x4567));
393
394 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t));
395
396 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
397 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
398
399 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
400 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
401 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
402
403 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
404 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
405 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
406
407 vacc0123 = _mm_cvtps_epi32(vscaled0123);
408 vacc4567 = _mm_cvtps_epi32(vscaled4567);
409
410 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
411 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
412
413 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
414 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
415 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
416
417 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
418 output += 8;
419 }
420 if XNN_UNLIKELY(c != 0) {
421 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 8);
422 do {
423 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
424
425 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i0));
426 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32(k)), vk_zero_point);
427 i0 += 4;
428
429 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
430 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i1));
431 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 8))), vk_zero_point);
432 i1 += 4;
433
434 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
435 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i2));
436 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 16))), vk_zero_point);
437 i2 += 4;
438
439 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
440 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i3));
441 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 24))), vk_zero_point);
442 i3 += 4;
443
444 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
445 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i4));
446 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 32))), vk_zero_point);
447 i4 += 4;
448
449 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
450 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i5));
451 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 40))), vk_zero_point);
452 i5 += 4;
453
454 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
455 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i6));
456 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 48))), vk_zero_point);
457 i6 += 4;
458
459 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
460 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i7));
461 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 56))), vk_zero_point);
462 i7 += 4;
463
464 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
465 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i8));
466 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 64))), vk_zero_point);
467 i8 += 4;
468
469 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
470 const __m128i vi9x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i9));
471 const __m128i vk9x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 72))), vk_zero_point);
472 i9 += 4;
473
474 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi9x0123, vk9x0123));
475 const __m128i vi10x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i10));
476 const __m128i vk10x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 80))), vk_zero_point);
477 i10 += 4;
478
479 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi10x0123, vk10x0123));
480 const __m128i vi11x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i11));
481 const __m128i vk11x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 88))), vk_zero_point);
482 i11 += 4;
483
484 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi11x0123, vk11x0123));
485 const __m128i vi12x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i12));
486 const __m128i vk12x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 96))), vk_zero_point);
487 i12 += 4;
488
489 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi12x0123, vk12x0123));
490 const __m128i vi13x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i13));
491 const __m128i vk13x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 104))), vk_zero_point);
492 i13 += 4;
493
494 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi13x0123, vk13x0123));
495 const __m128i vi14x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i14));
496 const __m128i vk14x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 112))), vk_zero_point);
497 i14 += 4;
498
499 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi14x0123, vk14x0123));
500 const __m128i vi15x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i15));
501 const __m128i vk15x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 120))), vk_zero_point);
502 i15 += 4;
503
504 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi15x0123, vk15x0123));
505 const __m128i vi16x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i16));
506 const __m128i vk16x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 128))), vk_zero_point);
507 i16 += 4;
508
509 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi16x0123, vk16x0123));
510 const __m128i vi17x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i17));
511 const __m128i vk17x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 136))), vk_zero_point);
512 i17 += 4;
513
514 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi17x0123, vk17x0123));
515 const __m128i vi18x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i18));
516 const __m128i vk18x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 144))), vk_zero_point);
517 i18 += 4;
518
519 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi18x0123, vk18x0123));
520 const __m128i vi19x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i19));
521 const __m128i vk19x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 152))), vk_zero_point);
522 i19 += 4;
523
524 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi19x0123, vk19x0123));
525 const __m128i vi20x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i20));
526 const __m128i vk20x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 160))), vk_zero_point);
527 i20 += 4;
528
529 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi20x0123, vk20x0123));
530 const __m128i vi21x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i21));
531 const __m128i vk21x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 168))), vk_zero_point);
532 i21 += 4;
533
534 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi21x0123, vk21x0123));
535 const __m128i vi22x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i22));
536 const __m128i vk22x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 176))), vk_zero_point);
537 i22 += 4;
538
539 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi22x0123, vk22x0123));
540 const __m128i vi23x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i23));
541 const __m128i vk23x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 184))), vk_zero_point);
542 i23 += 4;
543
544 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi23x0123, vk23x0123));
545 const __m128i vi24x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i24));
546 const __m128i vk24x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 192))), vk_zero_point);
547 i24 += 4;
548
549 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi24x0123, vk24x0123));
550
551 k += 4;
552
553 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
554 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.scale));
555 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.output_max_less_zero_point));
556 vacc0123 = _mm_cvtps_epi32(vscaled0123);
557
558 w = (const void*) ((const int32_t*) w + 4);
559
560 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
561 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
562
563 vout0123 = _mm_packus_epi16(vout0123, vout0123);
564 vout0123 = _mm_max_epu8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
565
566 if XNN_LIKELY(c >= 4) {
567 _mm_storeu_si32(output, vout0123);
568 output += 4;
569 c -= 4;
570 } else {
571 if (c & 2) {
572 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
573 vout0123 = _mm_srli_epi32(vout0123, 16);
574 output += 2;
575 }
576 if (c & 1) {
577 *output = (uint8_t) _mm_extract_epi8(vout0123, 0);
578 output += 1;
579 }
580 c = 0;
581 }
582 } while (c != 0);
583 }
584
585 output = (uint8_t*) ((uintptr_t) output + output_increment);
586 } while (--output_width != 0);
587 }
588