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_qs8_dwconv_minmax_fp32_ukernel_up16x9__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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__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_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 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 >= 16; c -= 16) {
83 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
84 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
85 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
86 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
87
88
89 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
90 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
91 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
92 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
93 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
94 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
95 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
96 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
97 i0 += 16;
98
99
100 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
101 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
102
103 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
104 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
105 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
106 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
107
108 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
109 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
110 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
111 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
112 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
113 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
114 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
115 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
116 i1 += 16;
117
118
119 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
120 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
121
122 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
123 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
124 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
125 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
126
127 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
128 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
129 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
130 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
131 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
132 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
133 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
134 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
135 i2 += 16;
136
137
138 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
139 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
140
141 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
142 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
143 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
144 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
145
146 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
147 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
148 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
149 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
150 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
151 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
152 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
153 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
154 i3 += 16;
155
156
157 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
158 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
159
160 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
161 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
162 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
163 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
164
165 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
166 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
167 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
168 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
169 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
170 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
171 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
172 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
173 i4 += 16;
174
175
176 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
177 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
178
179 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
180 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
181 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
182 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
183
184 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
185 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
186 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
187 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
188 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
189 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
190 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
191 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
192 i5 += 16;
193
194
195 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
196 vprod89ABCDEF = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
197
198 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
199 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
200 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
201 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
202
203 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
204 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
205 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
206 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
207 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
208 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
209 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
210 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
211 i6 += 16;
212
213
214 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
215 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
216
217 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
218 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
219 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
220 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
221
222 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
223 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
224 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
225 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
226 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
227 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
228 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
229 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
230 i7 += 16;
231
232
233 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
234 vprod89ABCDEF = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
235
236 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
237 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
238 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
239 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
240
241 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
242 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
243 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
244 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
245 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
246 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
247 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
248 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
249 i8 += 16;
250
251
252 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
253 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
254
255 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
256 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
257 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
258 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
259
260 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
261
262 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
263 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
264 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
265 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
266
267 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
268 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
269 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
270 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
271 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
272
273 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
274 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
275 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
276 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
277 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
278
279 vacc0123 = _mm_cvtps_epi32(vscaled0123);
280 vacc4567 = _mm_cvtps_epi32(vscaled4567);
281 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
282 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
283
284 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
285 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
286 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
287
288
289 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
290
291 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
292 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
293
294 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
295 output += 16;
296 }
297 if XNN_UNLIKELY(c != 0) {
298 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
299 do {
300 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
301 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
302
303
304 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
305 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
306 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
307 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
308 i0 += 8;
309
310
311 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
312
313 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
314 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
315
316 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
317 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
318 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
319 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
320 i1 += 8;
321
322
323 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
324
325 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
326 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
327
328 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
329 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
330 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
331 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
332 i2 += 8;
333
334
335 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
336
337 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
338 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
339
340 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
341 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
342 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
343 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
344 i3 += 8;
345
346
347 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
348
349 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
350 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
351
352 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
353 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
354 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
355 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
356 i4 += 8;
357
358
359 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
360
361 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
362 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
363
364 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
365 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
366 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
367 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
368 i5 += 8;
369
370
371 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
372
373 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
374 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
375
376 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
377 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
378 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
379 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
380 i6 += 8;
381
382
383 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
384
385 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
386 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
387
388 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
389 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
390 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
391 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
392 i7 += 8;
393
394
395 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
396
397 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
398 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
399
400 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
401 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
402 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
403 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
404 i8 += 8;
405
406
407 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
408
409 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
410 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
411
412 k += 8;
413
414 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
415 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
416
417 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
418 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
419 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
420
421 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
422 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
423 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
424
425 vacc0123 = _mm_cvtps_epi32(vscaled0123);
426 vacc4567 = _mm_cvtps_epi32(vscaled4567);
427
428 w = (const void*) ((const int32_t*) w + 8);
429
430 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
431 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
432
433
434 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
435
436 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
437
438 if XNN_LIKELY(c >= 8) {
439 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
440 output += 8;
441 c -= 8;
442 } else {
443 if (c & 4) {
444 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
445 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
446 output += 4;
447 }
448 if (c & 2) {
449 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
450 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
451 output += 2;
452 }
453 if (c & 1) {
454 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
455 output += 1;
456 }
457 c = 0;
458 }
459 } while (c != 0);
460 }
461
462 output = (int8_t*) ((uintptr_t) output + output_increment);
463 } while (--output_width != 0);
464 }
465