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_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__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 >= 24; c -= 24) {
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 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
88 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
89
90
91 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
92 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
93 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)));
94 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
95 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
96 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
97 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)));
98 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
99 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
100 const __m128i vxi0xGHIJKLMN = _mm_cvtepi8_epi16(vi0xGHIJKLMN);
101 const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)));
102 const __m128i vxk0xGHIJKLMN = _mm_cvtepi8_epi16(vk0xGHIJKLMN);
103 i0 += 24;
104
105
106 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
107 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
108 __m128i vprodGHIJKLMN = _mm_mullo_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
109
110 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
111 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
112 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
113 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
114 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
115 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
116
117 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
118 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
119 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
120 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
121 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
122 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
123 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
124 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
125 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
126 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(vi1xGHIJKLMN);
127 const __m128i vk1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
128 const __m128i vxk1xGHIJKLMN = _mm_cvtepi8_epi16(vk1xGHIJKLMN);
129 i1 += 24;
130
131
132 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
133 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
134 vprodGHIJKLMN = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
135
136 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
137 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
138 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
139 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
140 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
141 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
142
143 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
144 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
145 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
146 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
147 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
148 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
149 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
150 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
151 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
152 const __m128i vxi2xGHIJKLMN = _mm_cvtepi8_epi16(vi2xGHIJKLMN);
153 const __m128i vk2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
154 const __m128i vxk2xGHIJKLMN = _mm_cvtepi8_epi16(vk2xGHIJKLMN);
155 i2 += 24;
156
157
158 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
159 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
160 vprodGHIJKLMN = _mm_mullo_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
161
162 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
163 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
164 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
165 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
166 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
167 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
168
169 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
170 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
171 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
172 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
173 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
174 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
175 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
176 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
177 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
178 const __m128i vxi3xGHIJKLMN = _mm_cvtepi8_epi16(vi3xGHIJKLMN);
179 const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
180 const __m128i vxk3xGHIJKLMN = _mm_cvtepi8_epi16(vk3xGHIJKLMN);
181 i3 += 24;
182
183
184 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
185 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
186 vprodGHIJKLMN = _mm_mullo_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
187
188 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
189 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
190 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
191 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
192 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
193 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
194
195 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
196 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
197 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
198 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
199 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
200 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
201 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
202 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
203 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
204 const __m128i vxi4xGHIJKLMN = _mm_cvtepi8_epi16(vi4xGHIJKLMN);
205 const __m128i vk4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
206 const __m128i vxk4xGHIJKLMN = _mm_cvtepi8_epi16(vk4xGHIJKLMN);
207 i4 += 24;
208
209
210 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
211 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
212 vprodGHIJKLMN = _mm_mullo_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
213
214 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
215 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
216 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
217 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
218 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
219 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
220
221 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
222 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
223 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
224 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
225 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
226 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
227 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
228 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
229 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
230 const __m128i vxi5xGHIJKLMN = _mm_cvtepi8_epi16(vi5xGHIJKLMN);
231 const __m128i vk5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
232 const __m128i vxk5xGHIJKLMN = _mm_cvtepi8_epi16(vk5xGHIJKLMN);
233 i5 += 24;
234
235
236 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
237 vprod89ABCDEF = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
238 vprodGHIJKLMN = _mm_mullo_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
239
240 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
241 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
242 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
243 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
244 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
245 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
246
247 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
248 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
249 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
250 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
251 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
252 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
253 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
254 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
255 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
256 const __m128i vxi6xGHIJKLMN = _mm_cvtepi8_epi16(vi6xGHIJKLMN);
257 const __m128i vk6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
258 const __m128i vxk6xGHIJKLMN = _mm_cvtepi8_epi16(vk6xGHIJKLMN);
259 i6 += 24;
260
261
262 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
263 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
264 vprodGHIJKLMN = _mm_mullo_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
265
266 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
267 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
268 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
269 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
270 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
271 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
272
273 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
274 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
275 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
276 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
277 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
278 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
279 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
280 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
281 const __m128i vi7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i7 + 16));
282 const __m128i vxi7xGHIJKLMN = _mm_cvtepi8_epi16(vi7xGHIJKLMN);
283 const __m128i vk7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
284 const __m128i vxk7xGHIJKLMN = _mm_cvtepi8_epi16(vk7xGHIJKLMN);
285 i7 += 24;
286
287
288 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
289 vprod89ABCDEF = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
290 vprodGHIJKLMN = _mm_mullo_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
291
292 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
293 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
294 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
295 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
296 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
297 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
298
299 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
300 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
301 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
302 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
303 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
304 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
305 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
306 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
307 const __m128i vi8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i8 + 16));
308 const __m128i vxi8xGHIJKLMN = _mm_cvtepi8_epi16(vi8xGHIJKLMN);
309 const __m128i vk8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
310 const __m128i vxk8xGHIJKLMN = _mm_cvtepi8_epi16(vk8xGHIJKLMN);
311 i8 += 24;
312
313
314 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
315 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
316 vprodGHIJKLMN = _mm_mullo_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
317
318 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
319 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
320 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
321 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
322 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_cvtepi16_epi32(vprodGHIJKLMN));
323 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_srai_epi32(_mm_unpackhi_epi16(vprodGHIJKLMN, vprodGHIJKLMN), 16));
324
325 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t));
326
327 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
328 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
329 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
330 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
331 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
332 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
333
334 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
335 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
336 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
337 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
338 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
339 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale);
340 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale);
341
342 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
343 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
344 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
345 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
346 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
347 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
348 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
349
350 vacc0123 = _mm_cvtps_epi32(vscaled0123);
351 vacc4567 = _mm_cvtps_epi32(vscaled4567);
352 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
353 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
354 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
355 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
356
357 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
358 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
359 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
360 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
361
362
363 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
364 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
365
366 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
367 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
368 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
369
370 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
371 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
372 output += 24;
373 }
374 if XNN_UNLIKELY(c != 0) {
375 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
376 do {
377 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
378 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
379
380
381 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
382 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
383 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
384 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
385 i0 += 8;
386
387
388 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
389
390 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
391 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
392
393 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
394 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
395 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 24));
396 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
397 i1 += 8;
398
399
400 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
401
402 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
403 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
404
405 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
406 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
407 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
408 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
409 i2 += 8;
410
411
412 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
413
414 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
415 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
416
417 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
418 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
419 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 72));
420 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
421 i3 += 8;
422
423
424 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
425
426 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
427 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
428
429 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
430 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
431 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
432 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
433 i4 += 8;
434
435
436 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
437
438 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
439 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
440
441 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
442 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
443 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 120));
444 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
445 i5 += 8;
446
447
448 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
449
450 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
451 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
452
453 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
454 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
455 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
456 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
457 i6 += 8;
458
459
460 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
461
462 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
463 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
464
465 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
466 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
467 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 168));
468 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
469 i7 += 8;
470
471
472 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
473
474 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
475 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
476
477 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
478 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
479 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
480 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
481 i8 += 8;
482
483
484 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
485
486 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
487 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
488
489 k += 8;
490
491 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
492 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
493
494 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
495 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
496 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
497
498 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
499 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
500 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
501
502 vacc0123 = _mm_cvtps_epi32(vscaled0123);
503 vacc4567 = _mm_cvtps_epi32(vscaled4567);
504
505 w = (const void*) ((const int32_t*) w + 8);
506
507 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
508 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
509
510
511 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
512
513 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
514
515 if XNN_LIKELY(c >= 8) {
516 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
517 output += 8;
518 c -= 8;
519 } else {
520 if (c & 4) {
521 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
522 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
523 output += 4;
524 }
525 if (c & 2) {
526 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
527 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
528 output += 2;
529 }
530 if (c & 1) {
531 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
532 output += 1;
533 }
534 c = 0;
535 }
536 } while (c != 0);
537 }
538
539 output = (int8_t*) ((uintptr_t) output + output_increment);
540 } while (--output_width != 0);
541 }
542