1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <emmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse2_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)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__sse2_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_qs8_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 vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)));
94 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
95 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)));
96 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
97 const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)));
98 i0 += 24;
99
100 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
101 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
102 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
103 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
104 const __m128i vxi0xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi0xGHIJKLMN, vi0xGHIJKLMN), 8);
105 const __m128i vxk0xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk0xGHIJKLMN, vk0xGHIJKLMN), 8);
106
107 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
108 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
109 const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
110 const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
111 const __m128i vprod0xGHIJKLMNlo = _mm_mullo_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
112 const __m128i vprod0xGHIJKLMNhi = _mm_mulhi_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
113
114 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
115 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
116 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
117 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
118 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod0xGHIJKLMNlo, vprod0xGHIJKLMNhi));
119 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod0xGHIJKLMNlo, vprod0xGHIJKLMNhi));
120
121 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
122 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
123 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
124 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
125 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
126 const __m128i vk1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
127 i1 += 24;
128
129 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
130 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
131 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
132 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
133 const __m128i vxi1xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi1xGHIJKLMN, vi1xGHIJKLMN), 8);
134 const __m128i vxk1xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk1xGHIJKLMN, vk1xGHIJKLMN), 8);
135
136 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
137 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
138 const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
139 const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
140 const __m128i vprod1xGHIJKLMNlo = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
141 const __m128i vprod1xGHIJKLMNhi = _mm_mulhi_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
142
143 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
144 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
145 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
146 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
147 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod1xGHIJKLMNlo, vprod1xGHIJKLMNhi));
148 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod1xGHIJKLMNlo, vprod1xGHIJKLMNhi));
149
150 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
151 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
152 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
153 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
154 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
155 const __m128i vk2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
156 i2 += 24;
157
158 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
159 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
160 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
161 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
162 const __m128i vxi2xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi2xGHIJKLMN, vi2xGHIJKLMN), 8);
163 const __m128i vxk2xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk2xGHIJKLMN, vk2xGHIJKLMN), 8);
164
165 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
166 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
167 const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
168 const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
169 const __m128i vprod2xGHIJKLMNlo = _mm_mullo_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
170 const __m128i vprod2xGHIJKLMNhi = _mm_mulhi_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
171
172 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
173 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
174 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
175 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
176 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod2xGHIJKLMNlo, vprod2xGHIJKLMNhi));
177 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod2xGHIJKLMNlo, vprod2xGHIJKLMNhi));
178
179 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
180 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
181 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
182 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
183 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
184 const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
185 i3 += 24;
186
187 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
188 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
189 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
190 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
191 const __m128i vxi3xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi3xGHIJKLMN, vi3xGHIJKLMN), 8);
192 const __m128i vxk3xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk3xGHIJKLMN, vk3xGHIJKLMN), 8);
193
194 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
195 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
196 const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
197 const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
198 const __m128i vprod3xGHIJKLMNlo = _mm_mullo_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
199 const __m128i vprod3xGHIJKLMNhi = _mm_mulhi_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
200
201 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
202 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
203 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
204 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
205 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod3xGHIJKLMNlo, vprod3xGHIJKLMNhi));
206 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod3xGHIJKLMNlo, vprod3xGHIJKLMNhi));
207
208 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
209 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
210 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
211 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
212 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
213 const __m128i vk4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
214 i4 += 24;
215
216 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
217 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
218 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
219 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
220 const __m128i vxi4xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi4xGHIJKLMN, vi4xGHIJKLMN), 8);
221 const __m128i vxk4xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk4xGHIJKLMN, vk4xGHIJKLMN), 8);
222
223 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
224 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
225 const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
226 const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
227 const __m128i vprod4xGHIJKLMNlo = _mm_mullo_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
228 const __m128i vprod4xGHIJKLMNhi = _mm_mulhi_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
229
230 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
231 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
232 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
233 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
234 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod4xGHIJKLMNlo, vprod4xGHIJKLMNhi));
235 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod4xGHIJKLMNlo, vprod4xGHIJKLMNhi));
236
237 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
238 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
239 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
240 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
241 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
242 const __m128i vk5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
243 i5 += 24;
244
245 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
246 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
247 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
248 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
249 const __m128i vxi5xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi5xGHIJKLMN, vi5xGHIJKLMN), 8);
250 const __m128i vxk5xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk5xGHIJKLMN, vk5xGHIJKLMN), 8);
251
252 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
253 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
254 const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
255 const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
256 const __m128i vprod5xGHIJKLMNlo = _mm_mullo_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
257 const __m128i vprod5xGHIJKLMNhi = _mm_mulhi_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
258
259 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
260 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
261 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
262 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
263 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod5xGHIJKLMNlo, vprod5xGHIJKLMNhi));
264 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod5xGHIJKLMNlo, vprod5xGHIJKLMNhi));
265
266 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
267 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
268 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
269 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
270 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
271 const __m128i vk6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
272 i6 += 24;
273
274 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
275 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
276 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
277 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
278 const __m128i vxi6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi6xGHIJKLMN, vi6xGHIJKLMN), 8);
279 const __m128i vxk6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk6xGHIJKLMN, vk6xGHIJKLMN), 8);
280
281 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
282 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
283 const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
284 const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
285 const __m128i vprod6xGHIJKLMNlo = _mm_mullo_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
286 const __m128i vprod6xGHIJKLMNhi = _mm_mulhi_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
287
288 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
289 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
290 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
291 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
292 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod6xGHIJKLMNlo, vprod6xGHIJKLMNhi));
293 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod6xGHIJKLMNlo, vprod6xGHIJKLMNhi));
294
295 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
296 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
297 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
298 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
299 const __m128i vi7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i7 + 16));
300 const __m128i vk7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
301 i7 += 24;
302
303 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
304 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
305 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
306 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
307 const __m128i vxi7xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi7xGHIJKLMN, vi7xGHIJKLMN), 8);
308 const __m128i vxk7xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk7xGHIJKLMN, vk7xGHIJKLMN), 8);
309
310 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
311 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
312 const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
313 const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
314 const __m128i vprod7xGHIJKLMNlo = _mm_mullo_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
315 const __m128i vprod7xGHIJKLMNhi = _mm_mulhi_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
316
317 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
318 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
319 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
320 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
321 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod7xGHIJKLMNlo, vprod7xGHIJKLMNhi));
322 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod7xGHIJKLMNlo, vprod7xGHIJKLMNhi));
323
324 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
325 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
326 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
327 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
328 const __m128i vi8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i8 + 16));
329 const __m128i vk8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
330 i8 += 24;
331
332 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
333 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
334 const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
335 const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
336 const __m128i vxi8xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi8xGHIJKLMN, vi8xGHIJKLMN), 8);
337 const __m128i vxk8xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk8xGHIJKLMN, vk8xGHIJKLMN), 8);
338
339 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
340 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
341 const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
342 const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
343 const __m128i vprod8xGHIJKLMNlo = _mm_mullo_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
344 const __m128i vprod8xGHIJKLMNhi = _mm_mulhi_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
345
346 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
347 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
348 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
349 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
350 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod8xGHIJKLMNlo, vprod8xGHIJKLMNhi));
351 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod8xGHIJKLMNlo, vprod8xGHIJKLMNhi));
352
353 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t));
354
355 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
356 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
357 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
358 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
359 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
360 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
361
362 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
363 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
364 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
365 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
366 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
367 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale);
368 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale);
369
370 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
371 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
372 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
373 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
374 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
375 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
376 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
377
378 vacc0123 = _mm_cvtps_epi32(vscaled0123);
379 vacc4567 = _mm_cvtps_epi32(vscaled4567);
380 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
381 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
382 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
383 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
384
385 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
386 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
387 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
388 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
389
390 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
391 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
392 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
393 voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
394
395 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
396 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
397
398
399 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
400 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
401 output += 24;
402 }
403 if XNN_UNLIKELY(c != 0) {
404 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
405 do {
406 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
407 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
408
409
410 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
411 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
412 i0 += 8;
413
414 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
415 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
416
417 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
418 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
419
420 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
421 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
422
423 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
424 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 24));
425 i1 += 8;
426
427 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
428 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
429
430 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
431 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
432
433 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
434 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
435
436 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
437 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
438 i2 += 8;
439
440 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
441 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
442
443 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
444 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
445
446 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
447 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
448
449 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
450 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 72));
451 i3 += 8;
452
453 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
454 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
455
456 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
457 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
458
459 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
460 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
461
462 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
463 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
464 i4 += 8;
465
466 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
467 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
468
469 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
470 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
471
472 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
473 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
474
475 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
476 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 120));
477 i5 += 8;
478
479 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
480 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
481
482 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
483 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
484
485 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
486 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
487
488 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
489 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
490 i6 += 8;
491
492 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
493 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
494
495 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
496 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
497
498 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
499 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
500
501 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
502 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 168));
503 i7 += 8;
504
505 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
506 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
507
508 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
509 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
510
511 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
512 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
513
514 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
515 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
516 i8 += 8;
517
518 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
519 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
520
521 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
522 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
523
524 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
525 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
526
527 k += 8;
528
529 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
530 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
531
532 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
533 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
534 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
535
536 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
537 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
538 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
539
540 vacc0123 = _mm_cvtps_epi32(vscaled0123);
541 vacc4567 = _mm_cvtps_epi32(vscaled4567);
542
543 w = (const void*) ((const int32_t*) w + 8);
544
545 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
546 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
547
548 vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
549
550 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
551
552
553 if XNN_LIKELY(c >= 8) {
554 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
555 output += 8;
556 c -= 8;
557 } else {
558 if (c & 4) {
559 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
560 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
561 output += 4;
562 }
563 if (c & 2) {
564 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
565 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
566 output += 2;
567 }
568 if (c & 1) {
569 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
570 output += 1;
571 }
572 c = 0;
573 }
574 } while (c != 0);
575 }
576
577 output = (int8_t*) ((uintptr_t) output + output_increment);
578 } while (--output_width != 0);
579 }
580