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_up24x9__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_up24x9__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 >= 24; c -= 24) {
84 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
85 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
86 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
87 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
88 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
89 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
90
91
92 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
93 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
94 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)));
95 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
96 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
97 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
98 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)));
99 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
100 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
101 const __m128i vxi0xGHIJKLMN = _mm_cvtepi8_epi16(vi0xGHIJKLMN);
102 const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)));
103 const __m128i vxk0xGHIJKLMN = _mm_cvtepi8_epi16(vk0xGHIJKLMN);
104 i0 += 24;
105
106
107 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
108 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
109 __m128i vprodGHIJKLMN = _mm_mullo_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
110
111 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
112 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
113 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
114 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
115 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
116 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
117
118 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
119 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
120 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
121 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
122 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
123 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
124 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
125 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
126 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
127 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(vi1xGHIJKLMN);
128 const __m128i vk1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
129 const __m128i vxk1xGHIJKLMN = _mm_cvtepi8_epi16(vk1xGHIJKLMN);
130 i1 += 24;
131
132
133 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
134 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
135 vprodGHIJKLMN = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
136
137 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
138 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
139 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
140 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
141 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
142 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
143
144 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
145 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
146 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
147 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
148 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
149 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
150 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
151 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
152 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
153 const __m128i vxi2xGHIJKLMN = _mm_cvtepi8_epi16(vi2xGHIJKLMN);
154 const __m128i vk2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
155 const __m128i vxk2xGHIJKLMN = _mm_cvtepi8_epi16(vk2xGHIJKLMN);
156 i2 += 24;
157
158
159 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
160 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
161 vprodGHIJKLMN = _mm_mullo_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
162
163 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
164 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
165 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
166 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
167 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
168 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
169
170 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
171 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
172 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
173 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
174 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
175 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
176 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
177 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
178 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
179 const __m128i vxi3xGHIJKLMN = _mm_cvtepi8_epi16(vi3xGHIJKLMN);
180 const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
181 const __m128i vxk3xGHIJKLMN = _mm_cvtepi8_epi16(vk3xGHIJKLMN);
182 i3 += 24;
183
184
185 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
186 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
187 vprodGHIJKLMN = _mm_mullo_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
188
189 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
190 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
191 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
192 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
193 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
194 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
195
196 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
197 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
198 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
199 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
200 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
201 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
202 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
203 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
204 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
205 const __m128i vxi4xGHIJKLMN = _mm_cvtepi8_epi16(vi4xGHIJKLMN);
206 const __m128i vk4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
207 const __m128i vxk4xGHIJKLMN = _mm_cvtepi8_epi16(vk4xGHIJKLMN);
208 i4 += 24;
209
210
211 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
212 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
213 vprodGHIJKLMN = _mm_mullo_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
214
215 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
216 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
217 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
218 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
219 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
220 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
221
222 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
223 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
224 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
225 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
226 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
227 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
228 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
229 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
230 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
231 const __m128i vxi5xGHIJKLMN = _mm_cvtepi8_epi16(vi5xGHIJKLMN);
232 const __m128i vk5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
233 const __m128i vxk5xGHIJKLMN = _mm_cvtepi8_epi16(vk5xGHIJKLMN);
234 i5 += 24;
235
236
237 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
238 vprod89ABCDEF = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
239 vprodGHIJKLMN = _mm_mullo_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
240
241 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
242 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
243 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
244 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
245 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
246 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
247
248 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
249 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
250 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
251 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
252 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
253 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
254 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
255 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
256 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
257 const __m128i vxi6xGHIJKLMN = _mm_cvtepi8_epi16(vi6xGHIJKLMN);
258 const __m128i vk6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
259 const __m128i vxk6xGHIJKLMN = _mm_cvtepi8_epi16(vk6xGHIJKLMN);
260 i6 += 24;
261
262
263 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
264 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
265 vprodGHIJKLMN = _mm_mullo_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
266
267 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
268 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
269 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
270 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
271 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
272 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
273
274 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
275 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
276 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
277 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
278 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
279 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
280 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
281 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
282 const __m128i vi7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i7 + 16));
283 const __m128i vxi7xGHIJKLMN = _mm_cvtepi8_epi16(vi7xGHIJKLMN);
284 const __m128i vk7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
285 const __m128i vxk7xGHIJKLMN = _mm_cvtepi8_epi16(vk7xGHIJKLMN);
286 i7 += 24;
287
288
289 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
290 vprod89ABCDEF = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
291 vprodGHIJKLMN = _mm_mullo_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
292
293 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
294 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
295 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
296 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
297 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
298 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
299
300 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
301 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
302 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
303 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
304 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
305 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
306 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
307 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
308 const __m128i vi8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i8 + 16));
309 const __m128i vxi8xGHIJKLMN = _mm_cvtepi8_epi16(vi8xGHIJKLMN);
310 const __m128i vk8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
311 const __m128i vxk8xGHIJKLMN = _mm_cvtepi8_epi16(vk8xGHIJKLMN);
312 i8 += 24;
313
314
315 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
316 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
317 vprodGHIJKLMN = _mm_mullo_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
318
319 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
320 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
321 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
322 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
323 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
324 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
325
326 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t));
327
328 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
329 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
330 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
331 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
332 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
333 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
334
335 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
336 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
337 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
338 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
339 const __m128 vscaleGHIJ = _mm_loadu_ps((const float*) w + 16);
340 const __m128 vscaleKLMN = _mm_loadu_ps((const float*) w + 20);
341 w = (const void*) ((const float*) w + 24);
342 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
343 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
344 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
345 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
346 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscaleGHIJ);
347 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscaleKLMN);
348
349 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
350 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
351 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
352 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
353 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
354 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
355 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
356
357 vacc0123 = _mm_cvtps_epi32(vscaled0123);
358 vacc4567 = _mm_cvtps_epi32(vscaled4567);
359 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
360 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
361 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
362 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
363
364 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
365 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
366 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
367 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
368
369
370 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
371 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
372
373 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
374 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
375 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
376
377 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
378 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
379 output += 24;
380 }
381 if XNN_UNLIKELY(c != 0) {
382 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
383 do {
384 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
385 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
386
387
388 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
389 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
390 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
391 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
392 i0 += 8;
393
394
395 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
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 vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
401 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
402 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 24));
403 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
404 i1 += 8;
405
406
407 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
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 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
413 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
414 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
415 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
416 i2 += 8;
417
418
419 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
420
421 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
422 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
423
424 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
425 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
426 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 72));
427 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
428 i3 += 8;
429
430
431 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
432
433 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
434 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
435
436 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
437 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
438 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
439 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
440 i4 += 8;
441
442
443 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
444
445 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
446 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
447
448 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
449 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
450 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 120));
451 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
452 i5 += 8;
453
454
455 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
456
457 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
458 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
459
460 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
461 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
462 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
463 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
464 i6 += 8;
465
466
467 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
468
469 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
470 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
471
472 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
473 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
474 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 168));
475 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
476 i7 += 8;
477
478
479 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
480
481 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
482 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
483
484 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
485 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
486 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
487 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
488 i8 += 8;
489
490
491 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
492
493 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
494 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
495
496 k += 8;
497
498 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
499 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
500
501 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)));
502 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t) + 4 * sizeof(float)));
503 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
504 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
505
506 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
507 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
508 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
509
510 vacc0123 = _mm_cvtps_epi32(vscaled0123);
511 vacc4567 = _mm_cvtps_epi32(vscaled4567);
512
513 w = (const void*) ((const int32_t*) w + 8);
514
515 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
516 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
517
518
519 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
520
521 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
522
523 if XNN_LIKELY(c >= 8) {
524 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
525 output += 8;
526 c -= 8;
527 } else {
528 if (c & 4) {
529 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
530 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
531 output += 4;
532 }
533 if (c & 2) {
534 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
535 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
536 output += 2;
537 }
538 if (c & 1) {
539 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
540 output += 1;
541 }
542 c = 0;
543 }
544 } while (c != 0);
545 }
546
547 output = (int8_t*) ((uintptr_t) output + output_increment);
548 } while (--output_width != 0);
549 }
550