1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-sse-mul16.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <smmintrin.h>
13 
14 #include <xnnpack/dwconv.h>
15 
16 
xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16(
18     size_t channels,
19     size_t output_width,
20     const int8_t** input,
21     const void* weights,
22     int8_t* output,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const int8_t* zero,
27     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
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 >= 16; c -= 16) {
83       __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
84       __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 4 * sizeof(int32_t)));
85       __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t)));
86       __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 12 * sizeof(int32_t)));
87 
88 
89       const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
90       const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
91       const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
92       const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
93       const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
94       const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
95       const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
96       const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
97       i0 += 16;
98 
99 
100       const __m128i vp0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
101       const __m128i vp0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
102       const __m128i vp0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
103       const __m128i vp0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
104 
105       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp0x01234567lo, vp0x01234567hi));
106       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp0x01234567lo, vp0x01234567hi));
107       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp0x89ABCDEFlo, vp0x89ABCDEFhi));
108       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp0x89ABCDEFlo, vp0x89ABCDEFhi));
109 
110       const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
111       const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
112       const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
113       const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
114       const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
115       const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
116       const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
117       const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
118       i1 += 16;
119 
120 
121       const __m128i vp1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
122       const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
123       const __m128i vp1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
124       const __m128i vp1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
125 
126       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi));
127       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi));
128       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp1x89ABCDEFlo, vp1x89ABCDEFhi));
129       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp1x89ABCDEFlo, vp1x89ABCDEFhi));
130 
131       const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
132       const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
133       const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
134       const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
135       const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
136       const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
137       const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
138       const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
139       i2 += 16;
140 
141 
142       const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
143       const __m128i vp2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
144       const __m128i vp2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
145       const __m128i vp2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
146 
147       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi));
148       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi));
149       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp2x89ABCDEFlo, vp2x89ABCDEFhi));
150       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp2x89ABCDEFlo, vp2x89ABCDEFhi));
151 
152       const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
153       const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
154       const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
155       const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
156       const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
157       const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
158       const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
159       const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
160       i3 += 16;
161 
162 
163       const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
164       const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
165       const __m128i vp3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
166       const __m128i vp3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
167 
168       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi));
169       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi));
170       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp3x89ABCDEFlo, vp3x89ABCDEFhi));
171       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp3x89ABCDEFlo, vp3x89ABCDEFhi));
172 
173       const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
174       const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
175       const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
176       const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
177       const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
178       const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
179       const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
180       const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
181       i4 += 16;
182 
183 
184       const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
185       const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
186       const __m128i vp4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
187       const __m128i vp4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
188 
189       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi));
190       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi));
191       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp4x89ABCDEFlo, vp4x89ABCDEFhi));
192       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp4x89ABCDEFlo, vp4x89ABCDEFhi));
193 
194       const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
195       const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
196       const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
197       const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
198       const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
199       const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
200       const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
201       const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
202       i5 += 16;
203 
204 
205       const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
206       const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
207       const __m128i vp5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
208       const __m128i vp5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
209 
210       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp5x01234567lo, vp5x01234567hi));
211       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp5x01234567lo, vp5x01234567hi));
212       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp5x89ABCDEFlo, vp5x89ABCDEFhi));
213       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp5x89ABCDEFlo, vp5x89ABCDEFhi));
214 
215       const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
216       const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
217       const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
218       const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
219       const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
220       const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
221       const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
222       const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
223       i6 += 16;
224 
225 
226       const __m128i vp6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
227       const __m128i vp6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
228       const __m128i vp6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
229       const __m128i vp6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
230 
231       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp6x01234567lo, vp6x01234567hi));
232       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp6x01234567lo, vp6x01234567hi));
233       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp6x89ABCDEFlo, vp6x89ABCDEFhi));
234       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp6x89ABCDEFlo, vp6x89ABCDEFhi));
235 
236       const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
237       const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
238       const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
239       const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
240       const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
241       const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
242       const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
243       const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
244       i7 += 16;
245 
246 
247       const __m128i vp7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
248       const __m128i vp7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
249       const __m128i vp7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
250       const __m128i vp7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
251 
252       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp7x01234567lo, vp7x01234567hi));
253       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp7x01234567lo, vp7x01234567hi));
254       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp7x89ABCDEFlo, vp7x89ABCDEFhi));
255       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp7x89ABCDEFlo, vp7x89ABCDEFhi));
256 
257       const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
258       const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
259       const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
260       const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
261       const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
262       const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
263       const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
264       const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
265       i8 += 16;
266 
267 
268       const __m128i vp8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
269       const __m128i vp8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
270       const __m128i vp8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
271       const __m128i vp8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
272 
273       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp8x01234567lo, vp8x01234567hi));
274       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp8x01234567lo, vp8x01234567hi));
275       vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vp8x89ABCDEFlo, vp8x89ABCDEFhi));
276       vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vp8x89ABCDEFlo, vp8x89ABCDEFhi));
277 
278       w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
279 
280       const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
281       const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
282 
283       const __m128i vacc13 = _mm_shuffle_epi32(vacc0123, _MM_SHUFFLE(3, 3, 1, 1));
284       const __m128i vprod02 = _mm_add_epi64(_mm_mul_epi32(vacc0123, vmultiplier), vrounding);
285       const __m128i vprod13 = _mm_add_epi64(_mm_mul_epi32(vacc13, vmultiplier), vrounding);
286       const __m128i vacc57 = _mm_shuffle_epi32(vacc4567, _MM_SHUFFLE(3, 3, 1, 1));
287       const __m128i vprod46 = _mm_add_epi64(_mm_mul_epi32(vacc4567, vmultiplier), vrounding);
288       const __m128i vprod57 = _mm_add_epi64(_mm_mul_epi32(vacc57, vmultiplier), vrounding);
289       const __m128i vacc9B = _mm_shuffle_epi32(vacc89AB, _MM_SHUFFLE(3, 3, 1, 1));
290       const __m128i vprod8A = _mm_add_epi64(_mm_mul_epi32(vacc89AB, vmultiplier), vrounding);
291       const __m128i vprod9B = _mm_add_epi64(_mm_mul_epi32(vacc9B, vmultiplier), vrounding);
292       const __m128i vaccDF = _mm_shuffle_epi32(vaccCDEF, _MM_SHUFFLE(3, 3, 1, 1));
293       const __m128i vprodCE = _mm_add_epi64(_mm_mul_epi32(vaccCDEF, vmultiplier), vrounding);
294       const __m128i vprodDF = _mm_add_epi64(_mm_mul_epi32(vaccDF, vmultiplier), vrounding);
295 
296       const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31);
297       const __m128i vq31prod13 = _mm_add_epi64(vprod13, vprod13);
298       const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31);
299       const __m128i vq31prod57 = _mm_add_epi64(vprod57, vprod57);
300       const __m128i vq31prod8A = _mm_srli_epi64(vprod8A, 31);
301       const __m128i vq31prod9B = _mm_add_epi64(vprod9B, vprod9B);
302       const __m128i vq31prodCE = _mm_srli_epi64(vprodCE, 31);
303       const __m128i vq31prodDF = _mm_add_epi64(vprodDF, vprodDF);
304 
305       const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC);
306       const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC);
307       const __m128i vq31prod89AB = _mm_blend_epi16(vq31prod8A, vq31prod9B, 0xCC);
308       const __m128i vq31prodCDEF = _mm_blend_epi16(vq31prodCE, vq31prodDF, 0xCC);
309 
310       const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
311       const __m128i vrem0123 =
312         _mm_add_epi32(_mm_and_si128(vq31prod0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0123));
313       const __m128i vrem4567 =
314         _mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod4567));
315       const __m128i vrem89AB =
316         _mm_add_epi32(_mm_and_si128(vq31prod89AB, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod89AB));
317       const __m128i vremCDEF =
318         _mm_add_epi32(_mm_and_si128(vq31prodCDEF, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prodCDEF));
319 
320       const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
321       const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
322       vacc0123 =
323         _mm_sub_epi32(_mm_sra_epi32(vq31prod0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
324       vacc4567 =
325         _mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
326       vacc89AB =
327         _mm_sub_epi32(_mm_sra_epi32(vq31prod89AB, vshift), _mm_cmpgt_epi32(vrem89AB, vremainder_threshold));
328       vaccCDEF =
329         _mm_sub_epi32(_mm_sra_epi32(vq31prodCDEF, vshift), _mm_cmpgt_epi32(vremCDEF, vremainder_threshold));
330 
331       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
332       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
333       __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
334 
335       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
336       const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
337       vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
338       vout89ABCDEF = _mm_min_epi16(_mm_max_epi16(vout89ABCDEF, voutput_min), voutput_max);
339 
340       __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
341 
342       _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
343       output += 16;
344     }
345     if XNN_UNLIKELY(c != 0) {
346       const int8_t* k = (const int8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
347       do {
348         __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
349         __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 4 * sizeof(int32_t)));
350 
351 
352         const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
353         const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
354         const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
355         const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
356         i0 += 8;
357 
358 
359         const __m128i vp0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
360         const __m128i vp0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
361 
362         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp0x01234567lo, vp0x01234567hi));
363         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp0x01234567lo, vp0x01234567hi));
364 
365         const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
366         const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
367         const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
368         const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
369         i1 += 8;
370 
371 
372         const __m128i vp1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
373         const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
374 
375         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi));
376         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi));
377 
378         const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
379         const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
380         const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
381         const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
382         i2 += 8;
383 
384 
385         const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
386         const __m128i vp2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
387 
388         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi));
389         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi));
390 
391         const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
392         const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
393         const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
394         const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
395         i3 += 8;
396 
397 
398         const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
399         const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
400 
401         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi));
402         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi));
403 
404         const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
405         const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
406         const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
407         const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
408         i4 += 8;
409 
410 
411         const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
412         const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
413 
414         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi));
415         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi));
416 
417         const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
418         const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
419         const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
420         const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
421         i5 += 8;
422 
423 
424         const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
425         const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
426 
427         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp5x01234567lo, vp5x01234567hi));
428         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp5x01234567lo, vp5x01234567hi));
429 
430         const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
431         const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
432         const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
433         const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
434         i6 += 8;
435 
436 
437         const __m128i vp6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
438         const __m128i vp6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
439 
440         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp6x01234567lo, vp6x01234567hi));
441         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp6x01234567lo, vp6x01234567hi));
442 
443         const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
444         const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
445         const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
446         const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
447         i7 += 8;
448 
449 
450         const __m128i vp7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
451         const __m128i vp7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
452 
453         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp7x01234567lo, vp7x01234567hi));
454         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp7x01234567lo, vp7x01234567hi));
455 
456         const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
457         const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
458         const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
459         const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
460         i8 += 8;
461 
462 
463         const __m128i vp8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
464         const __m128i vp8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
465 
466         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp8x01234567lo, vp8x01234567hi));
467         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp8x01234567lo, vp8x01234567hi));
468 
469         w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
470         k += 8;
471 
472         const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
473         const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
474 
475         const __m128i vacc13 = _mm_shuffle_epi32(vacc0123, _MM_SHUFFLE(3, 3, 1, 1));
476         const __m128i vacc57 = _mm_shuffle_epi32(vacc4567, _MM_SHUFFLE(3, 3, 1, 1));
477 
478         const __m128i vprod02 = _mm_add_epi64(_mm_mul_epi32(vacc0123, vmultiplier), vrounding);
479         const __m128i vprod46 = _mm_add_epi64(_mm_mul_epi32(vacc4567, vmultiplier), vrounding);
480 
481         const __m128i vprod13 = _mm_add_epi64(_mm_mul_epi32(vacc13, vmultiplier), vrounding);
482         const __m128i vprod57 = _mm_add_epi64(_mm_mul_epi32(vacc57, vmultiplier), vrounding);
483 
484         const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31);
485         const __m128i vq31prod13 = _mm_add_epi64(vprod13, vprod13);
486         const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31);
487         const __m128i vq31prod57 = _mm_add_epi64(vprod57, vprod57);
488 
489         const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC);
490         const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC);
491 
492         const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
493         const __m128i vrem0123 =
494           _mm_add_epi32(_mm_and_si128(vq31prod0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0123));
495         const __m128i vrem4567 =
496           _mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod4567));
497 
498         const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
499         const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
500         vacc0123 =
501           _mm_sub_epi32(_mm_sra_epi32(vq31prod0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
502         vacc4567 =
503           _mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
504 
505         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
506         __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
507 
508         const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
509         const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
510         vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
511 
512         __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
513 
514         if XNN_LIKELY(c >= 8) {
515           _mm_storel_epi64((__m128i*) output, vout0123456701234567);
516           output += 8;
517           c -= 8;
518         } else {
519           if (c & 4) {
520             *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
521             vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
522             output += 4;
523           }
524           if (c & 2) {
525             *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
526             vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
527             output += 2;
528           }
529           if (c & 1) {
530             *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
531             output += 1;
532           }
533           c = 0;
534         }
535       } while (c != 0);
536     }
537 
538     output = (int8_t*) ((uintptr_t) output + output_increment);
539   } while (--output_width != 0);
540 }
541