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
16
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_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 input = (const int8_t**) ((uintptr_t) input + input_stride);
79
80 size_t c = channels;
81 const void* w = weights;
82 for (; c >= 8; c -= 8) {
83 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
84 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
85
86
87 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
88 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
89 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
90 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
91 i0 += 8;
92
93
94 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
95
96 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
97 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
98
99 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
100 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
101 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
102 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
103 i1 += 8;
104
105
106 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
107
108 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
109 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
110
111 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
112 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
113 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
114 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
115 i2 += 8;
116
117
118 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
119
120 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
121 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
122
123 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
124 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
125 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
126 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
127 i3 += 8;
128
129
130 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
131
132 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
133 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
134
135 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
136 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
137 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
138 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
139 i4 += 8;
140
141
142 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
143
144 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
145 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
146
147 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
148 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
149 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
150 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
151 i5 += 8;
152
153
154 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
155
156 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
157 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
158
159 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
160 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
161 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
162 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
163 i6 += 8;
164
165
166 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
167
168 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
169 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
170
171 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
172 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
173 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
174 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
175 i7 += 8;
176
177
178 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
179
180 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
181 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
182
183 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
184 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
185 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
186 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
187 i8 += 8;
188
189
190 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
191
192 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
193 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
194
195 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
196
197 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
198 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
199
200 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
201 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
202 w = (const void*) ((const float*) w + 8);
203 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
204 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
205
206 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
207 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
208 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
209
210 vacc0123 = _mm_cvtps_epi32(vscaled0123);
211 vacc4567 = _mm_cvtps_epi32(vscaled4567);
212
213 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
214 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
215
216
217 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
218
219 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
220 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
221
222 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
223 output += 8;
224 }
225 if XNN_UNLIKELY(c != 0) {
226 {
227 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
228 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
229
230
231 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
232 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
233 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
234 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
235
236
237 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
238
239 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
240 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
241
242 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
243 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
244 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
245 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
246
247
248 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
249
250 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
251 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
252
253 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
254 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
255 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
256 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
257
258
259 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
260
261 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
262 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
263
264 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
265 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
266 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
267 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
268
269
270 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
271
272 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
273 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
274
275 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
276 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
277 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
278 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
279
280
281 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
282
283 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
284 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
285
286 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
287 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
288 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
289 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
290
291
292 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
293
294 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
295 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
296
297 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
298 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
299 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
300 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
301
302
303 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
304
305 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
306 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
307
308 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
309 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
310 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
311 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
312
313
314 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
315
316 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
317 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
318
319 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
320 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
321 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
322 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
323
324
325 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
326
327 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
328 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
329
330
331 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
332 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
333
334 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
335 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t) + 4 * sizeof(float)));
336 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
337 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
338
339 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
340 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
341 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
342
343 vacc0123 = _mm_cvtps_epi32(vscaled0123);
344 vacc4567 = _mm_cvtps_epi32(vscaled4567);
345
346
347 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
348 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
349
350
351 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
352
353 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->sse4.output_min));
354
355 if (c & 4) {
356 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
357 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
358 output += 4;
359 }
360 if (c & 2) {
361 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
362 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
363 output += 2;
364 }
365 if (c & 1) {
366 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
367 output += 1;
368 }
369 }
370 }
371
372 output = (int8_t*) ((uintptr_t) output + output_increment);
373 } while (--output_width != 0);
374 }
375