xref: /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/up16x9-minmax-fp32-sse2-mul16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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_up16x9__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_up16x9__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 >= 16; c -= 16) {
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 
89 
90       const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
91       const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
92       const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
93       const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
94       i0 += 16;
95 
96       const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
97       const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
98       const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
99       const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
100 
101       const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
102       const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
103       const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
104       const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
105 
106       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
107       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
108       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
109       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
110 
111       const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
112       const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
113       const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
114       const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
115       i1 += 16;
116 
117       const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
118       const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
119       const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
120       const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
121 
122       const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
123       const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
124       const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
125       const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
126 
127       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
128       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
129       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
130       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
131 
132       const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
133       const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
134       const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
135       const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
136       i2 += 16;
137 
138       const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
139       const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
140       const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
141       const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
142 
143       const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
144       const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
145       const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
146       const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
147 
148       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
149       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
150       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
151       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
152 
153       const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
154       const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
155       const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
156       const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
157       i3 += 16;
158 
159       const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
160       const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
161       const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
162       const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
163 
164       const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
165       const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
166       const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
167       const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
168 
169       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
170       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
171       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
172       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
173 
174       const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
175       const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
176       const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
177       const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
178       i4 += 16;
179 
180       const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
181       const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
182       const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
183       const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
184 
185       const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
186       const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
187       const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
188       const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
189 
190       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
191       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
192       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
193       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
194 
195       const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
196       const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
197       const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
198       const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
199       i5 += 16;
200 
201       const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
202       const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
203       const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
204       const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
205 
206       const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
207       const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
208       const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
209       const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
210 
211       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
212       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
213       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
214       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
215 
216       const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
217       const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
218       const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
219       const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
220       i6 += 16;
221 
222       const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
223       const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
224       const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
225       const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
226 
227       const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
228       const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
229       const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
230       const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
231 
232       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
233       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
234       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
235       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
236 
237       const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
238       const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
239       const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
240       const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
241       i7 += 16;
242 
243       const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
244       const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
245       const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
246       const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
247 
248       const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
249       const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
250       const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
251       const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
252 
253       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
254       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
255       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
256       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
257 
258       const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
259       const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
260       const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
261       const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
262       i8 += 16;
263 
264       const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
265       const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
266       const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
267       const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
268 
269       const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
270       const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
271       const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
272       const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
273 
274       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
275       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
276       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
277       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
278 
279       w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
280 
281       __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
282       __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
283       __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
284       __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
285 
286       const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
287       vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
288       vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
289       vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
290       vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
291 
292       const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
293       vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
294       vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
295       vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
296       vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
297 
298       vacc0123 = _mm_cvtps_epi32(vscaled0123);
299       vacc4567 = _mm_cvtps_epi32(vscaled4567);
300       vacc89AB = _mm_cvtps_epi32(vscaled89AB);
301       vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
302 
303       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
304       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
305       __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
306 
307       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
308       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
309       vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
310 
311       __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
312 
313 
314       _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
315       output += 16;
316     }
317     if XNN_UNLIKELY(c != 0) {
318       const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
319       do {
320         __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
321         __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
322 
323 
324         const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
325         const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
326         i0 += 8;
327 
328         const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
329         const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
330 
331         const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
332         const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
333 
334         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
335         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
336 
337         const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
338         const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
339         i1 += 8;
340 
341         const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
342         const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
343 
344         const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
345         const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
346 
347         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
348         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
349 
350         const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
351         const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
352         i2 += 8;
353 
354         const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
355         const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
356 
357         const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
358         const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
359 
360         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
361         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
362 
363         const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
364         const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
365         i3 += 8;
366 
367         const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
368         const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
369 
370         const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
371         const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
372 
373         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
374         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
375 
376         const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
377         const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
378         i4 += 8;
379 
380         const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
381         const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
382 
383         const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
384         const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
385 
386         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
387         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
388 
389         const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
390         const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
391         i5 += 8;
392 
393         const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
394         const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
395 
396         const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
397         const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
398 
399         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
400         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
401 
402         const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
403         const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
404         i6 += 8;
405 
406         const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
407         const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
408 
409         const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
410         const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
411 
412         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
413         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
414 
415         const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
416         const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
417         i7 += 8;
418 
419         const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
420         const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
421 
422         const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
423         const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
424 
425         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
426         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
427 
428         const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
429         const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
430         i8 += 8;
431 
432         const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
433         const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
434 
435         const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
436         const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
437 
438         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
439         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
440 
441         k += 8;
442 
443         __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
444         __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
445 
446         const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
447         vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
448         vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
449 
450         const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
451         vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
452         vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
453 
454         vacc0123 = _mm_cvtps_epi32(vscaled0123);
455         vacc4567 = _mm_cvtps_epi32(vscaled4567);
456 
457         w = (const void*) ((const int32_t*) w + 8);
458 
459         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
460         __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
461 
462         vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
463 
464         __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
465 
466 
467         if XNN_LIKELY(c >= 8) {
468           _mm_storel_epi64((__m128i*) output, vout0123456701234567);
469           output += 8;
470           c -= 8;
471         } else {
472           if (c & 4) {
473             unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
474             vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
475             output += 4;
476           }
477           if (c & 2) {
478             unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
479             vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
480             output += 2;
481           }
482           if (c & 1) {
483             *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
484             output += 1;
485           }
486           c = 0;
487         }
488       } while (c != 0);
489     }
490 
491     output = (int8_t*) ((uintptr_t) output + output_increment);
492   } while (--output_width != 0);
493 }
494