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 <emmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16(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_up8x9__sse2_mul16(
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 do {
34 const uint8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const uint8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const uint8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const uint8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const uint8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const uint8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const uint8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const uint8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const uint8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 input = (const uint8_t**) ((uintptr_t) input + input_stride);
80
81 size_t c = channels;
82 const void* w = weights;
83 const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
84 for (; c >= 8; c -= 8) {
85 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
86 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
87
88
89 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
90 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
91 i0 += 8;
92
93 const __m128i vzero = _mm_setzero_si128();
94 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
95 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
96
97 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
98 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
99
100 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
101 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
102
103 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
104 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
105 i1 += 8;
106
107 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
108 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
109
110 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
111 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
112
113 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
114 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
115
116 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
117 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
118 i2 += 8;
119
120 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
121 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
122
123 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
124 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
125
126 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
127 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
128
129 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
130 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
131 i3 += 8;
132
133 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
134 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
135
136 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
137 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
138
139 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
140 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
141
142 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
143 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
144 i4 += 8;
145
146 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
147 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
148
149 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
150 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
151
152 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
153 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
154
155 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
156 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
157 i5 += 8;
158
159 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
160 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
161
162 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
163 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
164
165 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
166 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
167
168 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
169 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
170 i6 += 8;
171
172 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
173 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
174
175 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
176 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
177
178 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
179 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
180
181 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
182 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
183 i7 += 8;
184
185 const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
186 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
187
188 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
189 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
193
194 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
195 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
196 i8 += 8;
197
198 const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
199 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
200
201 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
202 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
203
204 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
205 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
206
207 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t));
208
209 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
210 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
211
212 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
213 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
214 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
215
216 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
217 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
218 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
219
220 vacc0123 = _mm_cvtps_epi32(vscaled0123);
221 vacc4567 = _mm_cvtps_epi32(vscaled4567);
222
223 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
224 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
225
226 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
227
228 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
229 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
230
231 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
232 output += 8;
233 }
234 if XNN_UNLIKELY(c != 0) {
235 {
236 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
237 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
238
239
240 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
241 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
242
243 const __m128i vzero = _mm_setzero_si128();
244 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
245 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
246
247 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
248 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
249
250 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
251 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
252
253 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
254 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
255
256 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
257 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
258
259 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
260 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
261
262 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
263 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
264
265 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
266 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
267
268 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
269 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
270
271 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
272 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
273
274 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
275 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
276
277 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
278 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
279
280 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
281 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
282
283 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
284 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
285
286 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
287 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
288
289 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
290 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
291
292 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
293 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
294
295 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
296 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
297
298 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
299 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
300
301 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
302 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
303
304 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
305 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
306
307 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
308 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
309
310 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
311 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
312
313 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
314 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
315
316 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
317 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
318
319 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
320 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
321
322 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
323 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
324
325 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
326 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
327
328 const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
329 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
330
331 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
332 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
333
334 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
335 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
336
337 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
338 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
339
340 const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
341 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
342
343 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
344 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
345
346 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
347 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
348
349
350 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
351 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
352
353 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
354 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
355 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
356
357 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
358 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
359 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
360
361 vacc0123 = _mm_cvtps_epi32(vscaled0123);
362 vacc4567 = _mm_cvtps_epi32(vscaled4567);
363
364
365 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
366 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
367
368 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
369
370 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
371
372 if (c & 4) {
373 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
374 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
375 output += 4;
376 }
377 if (c & 2) {
378 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
379 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
380 output += 2;
381 }
382 if (c & 1) {
383 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
384 output += 1;
385 }
386 }
387 }
388
389 output = (uint8_t*) ((uintptr_t) output + output_increment);
390 } while (--output_width != 0);
391 }
392