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_qs8_dwconv_minmax_fp32_ukernel_up8x25__xop_mul32(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__xop_mul32(
25 size_t channels,
26 size_t output_width,
27 const int8_t** input,
28 const void* weights,
29 int8_t* output,
30 size_t input_stride,
31 size_t output_increment,
32 size_t input_offset,
33 const int8_t* zero,
34 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(channels != 0);
37 assert(output_width != 0);
38
39 do {
40 const int8_t* i0 = input[0];
41 assert(i0 != NULL);
42 if XNN_UNPREDICTABLE(i0 != zero) {
43 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
44 }
45 const int8_t* i1 = input[1];
46 assert(i1 != NULL);
47 if XNN_UNPREDICTABLE(i1 != zero) {
48 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
49 }
50 const int8_t* i2 = input[2];
51 assert(i2 != NULL);
52 if XNN_UNPREDICTABLE(i2 != zero) {
53 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
54 }
55 const int8_t* i3 = input[3];
56 assert(i3 != NULL);
57 if XNN_UNPREDICTABLE(i3 != zero) {
58 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
59 }
60 const int8_t* i4 = input[4];
61 assert(i4 != NULL);
62 if XNN_UNPREDICTABLE(i4 != zero) {
63 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
64 }
65 const int8_t* i5 = input[5];
66 assert(i5 != NULL);
67 if XNN_UNPREDICTABLE(i5 != zero) {
68 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
69 }
70 const int8_t* i6 = input[6];
71 assert(i6 != NULL);
72 if XNN_UNPREDICTABLE(i6 != zero) {
73 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
74 }
75 const int8_t* i7 = input[7];
76 assert(i7 != NULL);
77 if XNN_UNPREDICTABLE(i7 != zero) {
78 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
79 }
80 const int8_t* i8 = input[8];
81 assert(i8 != NULL);
82 if XNN_UNPREDICTABLE(i8 != zero) {
83 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
84 }
85 const int8_t* i9 = input[9];
86 assert(i9 != NULL);
87 if XNN_UNPREDICTABLE(i9 != zero) {
88 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
89 }
90 const int8_t* i10 = input[10];
91 assert(i10 != NULL);
92 if XNN_UNPREDICTABLE(i10 != zero) {
93 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
94 }
95 const int8_t* i11 = input[11];
96 assert(i11 != NULL);
97 if XNN_UNPREDICTABLE(i11 != zero) {
98 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
99 }
100 const int8_t* i12 = input[12];
101 assert(i12 != NULL);
102 if XNN_UNPREDICTABLE(i12 != zero) {
103 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
104 }
105 const int8_t* i13 = input[13];
106 assert(i13 != NULL);
107 if XNN_UNPREDICTABLE(i13 != zero) {
108 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
109 }
110 const int8_t* i14 = input[14];
111 assert(i14 != NULL);
112 if XNN_UNPREDICTABLE(i14 != zero) {
113 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
114 }
115 const int8_t* i15 = input[15];
116 assert(i15 != NULL);
117 if XNN_UNPREDICTABLE(i15 != zero) {
118 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
119 }
120 const int8_t* i16 = input[16];
121 assert(i16 != NULL);
122 if XNN_UNPREDICTABLE(i16 != zero) {
123 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
124 }
125 const int8_t* i17 = input[17];
126 assert(i17 != NULL);
127 if XNN_UNPREDICTABLE(i17 != zero) {
128 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
129 }
130 const int8_t* i18 = input[18];
131 assert(i18 != NULL);
132 if XNN_UNPREDICTABLE(i18 != zero) {
133 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
134 }
135 const int8_t* i19 = input[19];
136 assert(i19 != NULL);
137 if XNN_UNPREDICTABLE(i19 != zero) {
138 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
139 }
140 const int8_t* i20 = input[20];
141 assert(i20 != NULL);
142 if XNN_UNPREDICTABLE(i20 != zero) {
143 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
144 }
145 const int8_t* i21 = input[21];
146 assert(i21 != NULL);
147 if XNN_UNPREDICTABLE(i21 != zero) {
148 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
149 }
150 const int8_t* i22 = input[22];
151 assert(i22 != NULL);
152 if XNN_UNPREDICTABLE(i22 != zero) {
153 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
154 }
155 const int8_t* i23 = input[23];
156 assert(i23 != NULL);
157 if XNN_UNPREDICTABLE(i23 != zero) {
158 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
159 }
160 const int8_t* i24 = input[24];
161 assert(i24 != NULL);
162 if XNN_UNPREDICTABLE(i24 != zero) {
163 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
164 }
165 input = (const int8_t**) ((uintptr_t) input + input_stride);
166
167 size_t c = channels;
168 const void* w = weights;
169 for (; c >= 8; c -= 8) {
170 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
171 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
172
173
174 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
175 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
176 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
177 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
178 i0 += 8;
179
180 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
181 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
182
183 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
184 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
185 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
186 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
187 i1 += 8;
188
189 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
190 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
191
192 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
193 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
194 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
195 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
196 i2 += 8;
197
198 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
199 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
200
201 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
202 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
203 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
204 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
205 i3 += 8;
206
207 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
208 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
209
210 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
211 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
212 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
213 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
214 i4 += 8;
215
216 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
217 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
218
219 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
220 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
221 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
222 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
223 i5 += 8;
224
225 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
226 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
227
228 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
229 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
230 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
231 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
232 i6 += 8;
233
234 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
235 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
236
237 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
238 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
239 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
240 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
241 i7 += 8;
242
243 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
244 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
245
246 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
247 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)))));
248 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
249 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 68 * sizeof(int8_t)))));
250 i8 += 8;
251
252 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
253 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
254
255 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
256 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)))));
257 const __m128i vi9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9 + 4)));
258 const __m128i vk9x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 76 * sizeof(int8_t)))));
259 i9 += 8;
260
261 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
262 vacc4567 = _mm_macc_epi32(vi9x4567, vk9x4567, vacc4567);
263
264 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
265 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t)))));
266 const __m128i vi10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10 + 4)));
267 const __m128i vk10x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 84 * sizeof(int8_t)))));
268 i10 += 8;
269
270 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
271 vacc4567 = _mm_macc_epi32(vi10x4567, vk10x4567, vacc4567);
272
273 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
274 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t)))));
275 const __m128i vi11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11 + 4)));
276 const __m128i vk11x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 92 * sizeof(int8_t)))));
277 i11 += 8;
278
279 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
280 vacc4567 = _mm_macc_epi32(vi11x4567, vk11x4567, vacc4567);
281
282 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
283 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t)))));
284 const __m128i vi12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12 + 4)));
285 const __m128i vk12x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 100 * sizeof(int8_t)))));
286 i12 += 8;
287
288 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
289 vacc4567 = _mm_macc_epi32(vi12x4567, vk12x4567, vacc4567);
290
291 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
292 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t)))));
293 const __m128i vi13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13 + 4)));
294 const __m128i vk13x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 108 * sizeof(int8_t)))));
295 i13 += 8;
296
297 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
298 vacc4567 = _mm_macc_epi32(vi13x4567, vk13x4567, vacc4567);
299
300 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
301 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t)))));
302 const __m128i vi14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14 + 4)));
303 const __m128i vk14x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 116 * sizeof(int8_t)))));
304 i14 += 8;
305
306 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
307 vacc4567 = _mm_macc_epi32(vi14x4567, vk14x4567, vacc4567);
308
309 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
310 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t)))));
311 const __m128i vi15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15 + 4)));
312 const __m128i vk15x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 124 * sizeof(int8_t)))));
313 i15 += 8;
314
315 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
316 vacc4567 = _mm_macc_epi32(vi15x4567, vk15x4567, vacc4567);
317
318 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
319 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t)))));
320 const __m128i vi16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16 + 4)));
321 const __m128i vk16x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 132 * sizeof(int8_t)))));
322 i16 += 8;
323
324 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
325 vacc4567 = _mm_macc_epi32(vi16x4567, vk16x4567, vacc4567);
326
327 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
328 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t)))));
329 const __m128i vi17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17 + 4)));
330 const __m128i vk17x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 140 * sizeof(int8_t)))));
331 i17 += 8;
332
333 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
334 vacc4567 = _mm_macc_epi32(vi17x4567, vk17x4567, vacc4567);
335
336 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
337 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t)))));
338 const __m128i vi18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18 + 4)));
339 const __m128i vk18x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 148 * sizeof(int8_t)))));
340 i18 += 8;
341
342 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
343 vacc4567 = _mm_macc_epi32(vi18x4567, vk18x4567, vacc4567);
344
345 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
346 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t)))));
347 const __m128i vi19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19 + 4)));
348 const __m128i vk19x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 156 * sizeof(int8_t)))));
349 i19 += 8;
350
351 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
352 vacc4567 = _mm_macc_epi32(vi19x4567, vk19x4567, vacc4567);
353
354 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
355 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t)))));
356 const __m128i vi20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20 + 4)));
357 const __m128i vk20x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 164 * sizeof(int8_t)))));
358 i20 += 8;
359
360 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
361 vacc4567 = _mm_macc_epi32(vi20x4567, vk20x4567, vacc4567);
362
363 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
364 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t)))));
365 const __m128i vi21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21 + 4)));
366 const __m128i vk21x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 172 * sizeof(int8_t)))));
367 i21 += 8;
368
369 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
370 vacc4567 = _mm_macc_epi32(vi21x4567, vk21x4567, vacc4567);
371
372 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
373 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t)))));
374 const __m128i vi22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22 + 4)));
375 const __m128i vk22x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 180 * sizeof(int8_t)))));
376 i22 += 8;
377
378 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
379 vacc4567 = _mm_macc_epi32(vi22x4567, vk22x4567, vacc4567);
380
381 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
382 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t)))));
383 const __m128i vi23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23 + 4)));
384 const __m128i vk23x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 188 * sizeof(int8_t)))));
385 i23 += 8;
386
387 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
388 vacc4567 = _mm_macc_epi32(vi23x4567, vk23x4567, vacc4567);
389
390 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
391 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t)))));
392 const __m128i vi24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24 + 4)));
393 const __m128i vk24x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 8 * sizeof(int32_t) + 196 * sizeof(int8_t)))));
394 i24 += 8;
395
396 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
397 vacc4567 = _mm_macc_epi32(vi24x4567, vk24x4567, vacc4567);
398
399 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t));
400
401 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
402 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
403
404 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
405 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
406 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
407
408 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
409 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
410 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
411
412 vacc0123 = _mm_cvtps_epi32(vscaled0123);
413 vacc4567 = _mm_cvtps_epi32(vscaled4567);
414
415 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
416 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
417
418 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
419 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
420 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
421
422 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
423 output += 8;
424 }
425 if XNN_UNLIKELY(c != 0) {
426 const int8_t* k = (const int8_t*) ((const int32_t*) w + 8);
427 do {
428 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
429
430 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
431 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) k)));
432 i0 += 4;
433
434 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
435 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
436 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 8))));
437 i1 += 4;
438
439 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
440 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
441 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16))));
442 i2 += 4;
443
444 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
445 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
446 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 24))));
447 i3 += 4;
448
449 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
450 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
451 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32))));
452 i4 += 4;
453
454 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
455 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
456 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 40))));
457 i5 += 4;
458
459 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
460 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
461 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48))));
462 i6 += 4;
463
464 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
465 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
466 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 56))));
467 i7 += 4;
468
469 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
470 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
471 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64))));
472 i8 += 4;
473
474 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
475 const __m128i vi9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i9)));
476 const __m128i vk9x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 72))));
477 i9 += 4;
478
479 vacc0123 = _mm_macc_epi32(vi9x0123, vk9x0123, vacc0123);
480 const __m128i vi10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i10)));
481 const __m128i vk10x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 80))));
482 i10 += 4;
483
484 vacc0123 = _mm_macc_epi32(vi10x0123, vk10x0123, vacc0123);
485 const __m128i vi11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i11)));
486 const __m128i vk11x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 88))));
487 i11 += 4;
488
489 vacc0123 = _mm_macc_epi32(vi11x0123, vk11x0123, vacc0123);
490 const __m128i vi12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i12)));
491 const __m128i vk12x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96))));
492 i12 += 4;
493
494 vacc0123 = _mm_macc_epi32(vi12x0123, vk12x0123, vacc0123);
495 const __m128i vi13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i13)));
496 const __m128i vk13x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 104))));
497 i13 += 4;
498
499 vacc0123 = _mm_macc_epi32(vi13x0123, vk13x0123, vacc0123);
500 const __m128i vi14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i14)));
501 const __m128i vk14x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 112))));
502 i14 += 4;
503
504 vacc0123 = _mm_macc_epi32(vi14x0123, vk14x0123, vacc0123);
505 const __m128i vi15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i15)));
506 const __m128i vk15x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 120))));
507 i15 += 4;
508
509 vacc0123 = _mm_macc_epi32(vi15x0123, vk15x0123, vacc0123);
510 const __m128i vi16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i16)));
511 const __m128i vk16x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 128))));
512 i16 += 4;
513
514 vacc0123 = _mm_macc_epi32(vi16x0123, vk16x0123, vacc0123);
515 const __m128i vi17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i17)));
516 const __m128i vk17x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 136))));
517 i17 += 4;
518
519 vacc0123 = _mm_macc_epi32(vi17x0123, vk17x0123, vacc0123);
520 const __m128i vi18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i18)));
521 const __m128i vk18x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144))));
522 i18 += 4;
523
524 vacc0123 = _mm_macc_epi32(vi18x0123, vk18x0123, vacc0123);
525 const __m128i vi19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i19)));
526 const __m128i vk19x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 152))));
527 i19 += 4;
528
529 vacc0123 = _mm_macc_epi32(vi19x0123, vk19x0123, vacc0123);
530 const __m128i vi20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i20)));
531 const __m128i vk20x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 160))));
532 i20 += 4;
533
534 vacc0123 = _mm_macc_epi32(vi20x0123, vk20x0123, vacc0123);
535 const __m128i vi21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i21)));
536 const __m128i vk21x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 168))));
537 i21 += 4;
538
539 vacc0123 = _mm_macc_epi32(vi21x0123, vk21x0123, vacc0123);
540 const __m128i vi22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i22)));
541 const __m128i vk22x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 176))));
542 i22 += 4;
543
544 vacc0123 = _mm_macc_epi32(vi22x0123, vk22x0123, vacc0123);
545 const __m128i vi23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i23)));
546 const __m128i vk23x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 184))));
547 i23 += 4;
548
549 vacc0123 = _mm_macc_epi32(vi23x0123, vk23x0123, vacc0123);
550 const __m128i vi24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i24)));
551 const __m128i vk24x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192))));
552 i24 += 4;
553
554 vacc0123 = _mm_macc_epi32(vi24x0123, vk24x0123, vacc0123);
555
556 k += 4;
557
558 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
559 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
560 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
561 vacc0123 = _mm_cvtps_epi32(vscaled0123);
562
563 w = (const void*) ((const int32_t*) w + 4);
564
565 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
566 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
567
568 vout0123 = _mm_packs_epi16(vout0123, vout0123);
569 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
570
571 if XNN_LIKELY(c >= 4) {
572 _mm_storeu_si32(output, vout0123);
573 output += 4;
574 c -= 4;
575 } else {
576 if (c & 2) {
577 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
578 vout0123 = _mm_srli_epi32(vout0123, 16);
579 output += 2;
580 }
581 if (c & 1) {
582 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
583 output += 1;
584 }
585 c = 0;
586 }
587 } while (c != 0);
588 }
589
590 output = (int8_t*) ((uintptr_t) output + output_increment);
591 } while (--output_width != 0);
592 }
593