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
16
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)])17 void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16(
18 size_t channels,
19 size_t output_width,
20 const uint8_t** input,
21 const void* weights,
22 uint8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const uint8_t* zero,
27 const union xnn_qu8_conv_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 uint8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const uint8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const uint8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const uint8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const uint8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const uint8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const uint8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const uint8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const uint8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 input = (const uint8_t**) ((uintptr_t) input + input_stride);
79
80 size_t c = channels;
81 const void* w = weights;
82 const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
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 vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
90 i0 += 8;
91
92 const __m128i vzero = _mm_setzero_si128();
93 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
94 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
95
96 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
97 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
98
99 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
100 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
101
102 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
103 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
104 i1 += 8;
105
106 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
107 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
108
109 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
110 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
111
112 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
113 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
114
115 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
116 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
117 i2 += 8;
118
119 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
120 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
121
122 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
123 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
124
125 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
126 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
127
128 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
129 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
130 i3 += 8;
131
132 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
133 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
134
135 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
136 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
137
138 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
139 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
140
141 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
142 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
143 i4 += 8;
144
145 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
146 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
147
148 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
149 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
150
151 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
152 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
153
154 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
155 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
156 i5 += 8;
157
158 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
159 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
160
161 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
162 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
163
164 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
165 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
166
167 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
168 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
169 i6 += 8;
170
171 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
172 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
173
174 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
175 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
176
177 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
178 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
179
180 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
181 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
182 i7 += 8;
183
184 const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
185 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
186
187 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
188 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
189
190 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
191 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
192
193 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
194 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
195 i8 += 8;
196
197 const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
198 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
199
200 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
201 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
202
203 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
204 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
205
206 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t));
207
208 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
209 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
210
211 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
212 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
213 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
214
215 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
216 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
217 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
218
219 vacc0123 = _mm_cvtps_epi32(vscaled0123);
220 vacc4567 = _mm_cvtps_epi32(vscaled4567);
221
222 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
223 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
224
225 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
226
227 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
228 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
229
230 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
231 output += 8;
232 }
233 if XNN_UNLIKELY(c != 0) {
234 {
235 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
236 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
237
238
239 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
240 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
241
242 const __m128i vzero = _mm_setzero_si128();
243 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
244 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
245
246 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
247 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
248
249 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
250 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
251
252 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
253 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
254
255 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
256 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
257
258 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
259 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
260
261 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
262 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
263
264 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
265 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
266
267 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
268 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
269
270 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
271 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
272
273 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
274 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
275
276 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
277 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
278
279 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
280 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
281
282 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
283 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
284
285 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
286 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
287
288 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
289 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
290
291 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
292 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
293
294 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
295 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
296
297 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
298 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
299
300 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
301 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
302
303 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
304 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
305
306 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
307 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
308
309 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
310 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
311
312 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
313 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
314
315 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
316 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
317
318 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
319 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
320
321 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
322 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
323
324 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
325 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
326
327 const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
328 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
329
330 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
331 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
332
333 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
334 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
335
336 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
337 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
338
339 const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
340 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
341
342 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
343 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
344
345 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
346 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
347
348
349 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
350 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
351
352 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
353 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
354 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
355
356 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
357 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
358 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
359
360 vacc0123 = _mm_cvtps_epi32(vscaled0123);
361 vacc4567 = _mm_cvtps_epi32(vscaled4567);
362
363
364 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
365 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
366
367 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
368
369 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
370
371 if (c & 4) {
372 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
373 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
374 output += 4;
375 }
376 if (c & 2) {
377 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
378 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
379 output += 2;
380 }
381 if (c & 1) {
382 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
383 output += 1;
384 }
385 }
386 }
387
388 output = (uint8_t*) ((uintptr_t) output + output_increment);
389 } while (--output_width != 0);
390 }
391