1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.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 <smmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const int8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const int8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const int8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const int8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const int8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const int8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 input = (const int8_t**) ((uintptr_t) input + input_stride);
80
81 size_t c = channels;
82 const void* w = weights;
83 for (; c >= 8; c -= 8) {
84 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
85 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
86
87
88 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
89 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
90 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
91 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
92 i0 += 8;
93
94
95 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
96
97 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
98 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
99
100 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
101 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
102 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
103 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
104 i1 += 8;
105
106
107 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
108
109 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
110 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
111
112 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
113 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
114 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
115 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
116 i2 += 8;
117
118
119 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
120
121 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
122 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
123
124 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
125 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
126 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
127 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
128 i3 += 8;
129
130
131 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
132
133 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
134 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
135
136 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
137 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
138 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
139 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
140 i4 += 8;
141
142
143 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
144
145 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
146 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
147
148 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
149 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
150 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
151 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
152 i5 += 8;
153
154
155 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
156
157 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
158 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
159
160 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
161 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
162 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
163 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
164 i6 += 8;
165
166
167 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
168
169 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
170 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
171
172 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
173 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
174 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
175 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
176 i7 += 8;
177
178
179 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
180
181 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
182 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
183
184 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
185 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
186 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
187 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
188 i8 += 8;
189
190
191 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
192
193 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
194 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
195
196 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
197
198 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
199 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
200
201 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
202 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
203 w = (const void*) ((const float*) w + 8);
204 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
205 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
206
207 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
208 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
209 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
210
211 vacc0123 = _mm_cvtps_epi32(vscaled0123);
212 vacc4567 = _mm_cvtps_epi32(vscaled4567);
213
214 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
215 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
216
217
218 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
219
220 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
221 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
222
223 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
224 output += 8;
225 }
226 if XNN_UNLIKELY(c != 0) {
227 {
228 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
229 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
230
231
232 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
233 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
234 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
235 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
236
237
238 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
239
240 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
241 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
242
243 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
244 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
245 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
246 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
247
248
249 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
250
251 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
252 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
253
254 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
255 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
256 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
257 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
258
259
260 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
261
262 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
263 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
264
265 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
266 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
267 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
268 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
269
270
271 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
272
273 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
274 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
275
276 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
277 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
278 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
279 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
280
281
282 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
283
284 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
285 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
286
287 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
288 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
289 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
290 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
291
292
293 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
294
295 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
296 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
297
298 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
299 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
300 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
301 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
302
303
304 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
305
306 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
307 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
308
309 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
310 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
311 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
312 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
313
314
315 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
316
317 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
318 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
319
320 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
321 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
322 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
323 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
324
325
326 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
327
328 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
329 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
330
331
332 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
333 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
334
335 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
336 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t) + 4 * sizeof(float)));
337 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
338 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
339
340 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
341 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
342 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
343
344 vacc0123 = _mm_cvtps_epi32(vscaled0123);
345 vacc4567 = _mm_cvtps_epi32(vscaled4567);
346
347
348 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
349 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
350
351
352 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
353
354 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
355
356 if (c & 4) {
357 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
358 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
359 output += 4;
360 }
361 if (c & 2) {
362 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
363 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
364 output += 2;
365 }
366 if (c & 1) {
367 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
368 output += 1;
369 }
370 }
371 }
372
373 output = (int8_t*) ((uintptr_t) output + output_increment);
374 } while (--output_width != 0);
375 }
376