1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <emmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_qc8_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__sse2_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_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 vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)));
93 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
94 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)));
95 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16));
96 const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)));
97 i0 += 24;
98
99 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
100 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
101 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
102 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
103 const __m128i vxi0xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi0xGHIJKLMN, vi0xGHIJKLMN), 8);
104 const __m128i vxk0xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk0xGHIJKLMN, vk0xGHIJKLMN), 8);
105
106 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
107 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
108 const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
109 const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
110 const __m128i vprod0xGHIJKLMNlo = _mm_mullo_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
111 const __m128i vprod0xGHIJKLMNhi = _mm_mulhi_epi16(vxi0xGHIJKLMN, vxk0xGHIJKLMN);
112
113 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
114 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
115 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
116 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
117 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod0xGHIJKLMNlo, vprod0xGHIJKLMNhi));
118 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod0xGHIJKLMNlo, vprod0xGHIJKLMNhi));
119
120 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
121 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
122 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
123 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
124 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16));
125 const __m128i vk1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
126 i1 += 24;
127
128 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
129 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
130 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
131 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
132 const __m128i vxi1xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi1xGHIJKLMN, vi1xGHIJKLMN), 8);
133 const __m128i vxk1xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk1xGHIJKLMN, vk1xGHIJKLMN), 8);
134
135 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
136 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
137 const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
138 const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
139 const __m128i vprod1xGHIJKLMNlo = _mm_mullo_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
140 const __m128i vprod1xGHIJKLMNhi = _mm_mulhi_epi16(vxi1xGHIJKLMN, vxk1xGHIJKLMN);
141
142 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
143 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
144 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
145 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
146 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod1xGHIJKLMNlo, vprod1xGHIJKLMNhi));
147 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod1xGHIJKLMNlo, vprod1xGHIJKLMNhi));
148
149 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
150 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
151 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
152 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
153 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16));
154 const __m128i vk2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
155 i2 += 24;
156
157 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
158 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
159 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
160 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
161 const __m128i vxi2xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi2xGHIJKLMN, vi2xGHIJKLMN), 8);
162 const __m128i vxk2xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk2xGHIJKLMN, vk2xGHIJKLMN), 8);
163
164 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
165 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
166 const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
167 const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
168 const __m128i vprod2xGHIJKLMNlo = _mm_mullo_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
169 const __m128i vprod2xGHIJKLMNhi = _mm_mulhi_epi16(vxi2xGHIJKLMN, vxk2xGHIJKLMN);
170
171 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
172 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
173 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
174 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
175 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod2xGHIJKLMNlo, vprod2xGHIJKLMNhi));
176 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod2xGHIJKLMNlo, vprod2xGHIJKLMNhi));
177
178 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
179 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
180 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
181 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
182 const __m128i vi3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i3 + 16));
183 const __m128i vk3xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
184 i3 += 24;
185
186 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
187 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
188 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
189 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
190 const __m128i vxi3xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi3xGHIJKLMN, vi3xGHIJKLMN), 8);
191 const __m128i vxk3xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk3xGHIJKLMN, vk3xGHIJKLMN), 8);
192
193 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
194 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
195 const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
196 const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
197 const __m128i vprod3xGHIJKLMNlo = _mm_mullo_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
198 const __m128i vprod3xGHIJKLMNhi = _mm_mulhi_epi16(vxi3xGHIJKLMN, vxk3xGHIJKLMN);
199
200 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
201 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
202 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
203 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
204 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod3xGHIJKLMNlo, vprod3xGHIJKLMNhi));
205 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod3xGHIJKLMNlo, vprod3xGHIJKLMNhi));
206
207 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
208 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
209 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
210 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
211 const __m128i vi4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i4 + 16));
212 const __m128i vk4xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
213 i4 += 24;
214
215 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
216 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
217 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
218 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
219 const __m128i vxi4xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi4xGHIJKLMN, vi4xGHIJKLMN), 8);
220 const __m128i vxk4xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk4xGHIJKLMN, vk4xGHIJKLMN), 8);
221
222 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
223 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
224 const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
225 const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
226 const __m128i vprod4xGHIJKLMNlo = _mm_mullo_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
227 const __m128i vprod4xGHIJKLMNhi = _mm_mulhi_epi16(vxi4xGHIJKLMN, vxk4xGHIJKLMN);
228
229 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
230 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
231 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
232 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
233 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod4xGHIJKLMNlo, vprod4xGHIJKLMNhi));
234 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod4xGHIJKLMNlo, vprod4xGHIJKLMNhi));
235
236 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
237 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
238 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
239 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
240 const __m128i vi5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i5 + 16));
241 const __m128i vk5xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
242 i5 += 24;
243
244 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
245 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
246 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
247 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
248 const __m128i vxi5xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi5xGHIJKLMN, vi5xGHIJKLMN), 8);
249 const __m128i vxk5xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk5xGHIJKLMN, vk5xGHIJKLMN), 8);
250
251 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
252 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
253 const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
254 const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
255 const __m128i vprod5xGHIJKLMNlo = _mm_mullo_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
256 const __m128i vprod5xGHIJKLMNhi = _mm_mulhi_epi16(vxi5xGHIJKLMN, vxk5xGHIJKLMN);
257
258 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
259 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
260 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
261 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
262 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod5xGHIJKLMNlo, vprod5xGHIJKLMNhi));
263 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod5xGHIJKLMNlo, vprod5xGHIJKLMNhi));
264
265 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
266 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
267 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
268 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
269 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16));
270 const __m128i vk6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
271 i6 += 24;
272
273 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
274 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
275 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
276 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
277 const __m128i vxi6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi6xGHIJKLMN, vi6xGHIJKLMN), 8);
278 const __m128i vxk6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk6xGHIJKLMN, vk6xGHIJKLMN), 8);
279
280 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
281 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
282 const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
283 const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
284 const __m128i vprod6xGHIJKLMNlo = _mm_mullo_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
285 const __m128i vprod6xGHIJKLMNhi = _mm_mulhi_epi16(vxi6xGHIJKLMN, vxk6xGHIJKLMN);
286
287 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
288 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
289 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
290 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
291 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod6xGHIJKLMNlo, vprod6xGHIJKLMNhi));
292 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod6xGHIJKLMNlo, vprod6xGHIJKLMNhi));
293
294 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
295 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
296 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
297 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
298 const __m128i vi7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i7 + 16));
299 const __m128i vk7xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
300 i7 += 24;
301
302 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
303 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
304 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
305 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
306 const __m128i vxi7xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi7xGHIJKLMN, vi7xGHIJKLMN), 8);
307 const __m128i vxk7xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk7xGHIJKLMN, vk7xGHIJKLMN), 8);
308
309 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
310 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
311 const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
312 const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
313 const __m128i vprod7xGHIJKLMNlo = _mm_mullo_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
314 const __m128i vprod7xGHIJKLMNhi = _mm_mulhi_epi16(vxi7xGHIJKLMN, vxk7xGHIJKLMN);
315
316 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
317 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
318 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
319 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
320 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod7xGHIJKLMNlo, vprod7xGHIJKLMNhi));
321 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod7xGHIJKLMNlo, vprod7xGHIJKLMNhi));
322
323 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
324 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
325 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
326 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
327 const __m128i vi8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i8 + 16));
328 const __m128i vk8xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
329 i8 += 24;
330
331 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
332 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
333 const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
334 const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
335 const __m128i vxi8xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi8xGHIJKLMN, vi8xGHIJKLMN), 8);
336 const __m128i vxk8xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vk8xGHIJKLMN, vk8xGHIJKLMN), 8);
337
338 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
339 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
340 const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
341 const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
342 const __m128i vprod8xGHIJKLMNlo = _mm_mullo_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
343 const __m128i vprod8xGHIJKLMNhi = _mm_mulhi_epi16(vxi8xGHIJKLMN, vxk8xGHIJKLMN);
344
345 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
346 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
347 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
348 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
349 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vprod8xGHIJKLMNlo, vprod8xGHIJKLMNhi));
350 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vprod8xGHIJKLMNlo, vprod8xGHIJKLMNhi));
351
352 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t));
353
354 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
355 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
356 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
357 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
358 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
359 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
360
361 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
362 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
363 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
364 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
365 const __m128 vscaleGHIJ = _mm_loadu_ps((const float*) w + 16);
366 const __m128 vscaleKLMN = _mm_loadu_ps((const float*) w + 20);
367 w = (const void*) ((const float*) w + 24);
368 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
369 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
370 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
371 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
372 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscaleGHIJ);
373 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscaleKLMN);
374
375 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
376 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
377 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
378 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
379 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
380 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
381 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
382
383 vacc0123 = _mm_cvtps_epi32(vscaled0123);
384 vacc4567 = _mm_cvtps_epi32(vscaled4567);
385 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
386 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
387 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
388 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
389
390 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
391 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
392 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
393 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
394
395 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
396 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
397 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
398 voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
399
400 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
401 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
402
403
404 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
405 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
406 output += 24;
407 }
408 if XNN_UNLIKELY(c != 0) {
409 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
410 do {
411 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
412 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
413
414
415 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
416 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
417 i0 += 8;
418
419 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
420 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
421
422 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
423 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
424
425 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
426 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
427
428 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
429 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 24));
430 i1 += 8;
431
432 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
433 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
434
435 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
436 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
437
438 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
439 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
440
441 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
442 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
443 i2 += 8;
444
445 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
446 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
447
448 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
449 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
450
451 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
452 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
453
454 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
455 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 72));
456 i3 += 8;
457
458 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
459 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
460
461 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
462 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
463
464 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
465 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
466
467 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
468 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
469 i4 += 8;
470
471 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
472 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
473
474 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
475 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
476
477 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
478 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
479
480 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
481 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 120));
482 i5 += 8;
483
484 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
485 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
486
487 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
488 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
489
490 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
491 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
492
493 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
494 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 144));
495 i6 += 8;
496
497 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
498 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
499
500 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
501 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
502
503 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
504 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
505
506 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
507 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 168));
508 i7 += 8;
509
510 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
511 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
512
513 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
514 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
515
516 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
517 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
518
519 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
520 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 192));
521 i8 += 8;
522
523 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
524 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
525
526 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
527 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
528
529 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
530 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
531
532 k += 8;
533
534 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
535 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
536
537 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)));
538 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t) + 4 * sizeof(float)));
539 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
540 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
541
542 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
543 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
544 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
545
546 vacc0123 = _mm_cvtps_epi32(vscaled0123);
547 vacc4567 = _mm_cvtps_epi32(vscaled4567);
548
549 w = (const void*) ((const int32_t*) w + 8);
550
551 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
552 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
553
554 vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->sse2.output_min));
555
556 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
557
558
559 if XNN_LIKELY(c >= 8) {
560 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
561 output += 8;
562 c -= 8;
563 } else {
564 if (c & 4) {
565 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
566 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
567 output += 4;
568 }
569 if (c & 2) {
570 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
571 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
572 output += 2;
573 }
574 if (c & 1) {
575 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
576 output += 1;
577 }
578 c = 0;
579 }
580 } while (c != 0);
581 }
582
583 output = (int8_t*) ((uintptr_t) output + output_increment);
584 } while (--output_width != 0);
585 }
586