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_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16(
19 size_t channels,
20 size_t output_width,
21 const uint8_t** input,
22 const void* weights,
23 uint8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_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 uint8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const uint8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const uint8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const uint8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const uint8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const uint8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const uint8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const uint8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const uint8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 input = (const uint8_t**) ((uintptr_t) input + input_stride);
80
81 size_t c = channels;
82 const void* w = weights;
83 const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
84 for (; c >= 16; c -= 16) {
85 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
86 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
87 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
88 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
89
90
91 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
92 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
93 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
94 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
95 i0 += 16;
96
97 const __m128i vzero = _mm_setzero_si128();
98 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
99 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
100 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
101 const __m128i vxk0x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vzero), vk_zero_point);
102
103 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
104 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
105 const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
106 const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
107
108 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
109 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
110 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
111 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
112
113 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
114 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
115 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
116 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
117 i1 += 16;
118
119 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
120 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
121 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
122 const __m128i vxk1x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vzero), vk_zero_point);
123
124 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
125 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
126 const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
127 const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
128
129 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
130 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
131 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
132 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
133
134 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
135 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
136 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
137 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
138 i2 += 16;
139
140 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
141 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
142 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
143 const __m128i vxk2x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vzero), vk_zero_point);
144
145 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
146 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
147 const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
148 const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
149
150 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
151 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
152 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
153 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
154
155 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
156 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
157 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
158 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
159 i3 += 16;
160
161 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
162 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
163 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
164 const __m128i vxk3x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vzero), vk_zero_point);
165
166 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
167 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
168 const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
169 const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
170
171 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
172 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
173 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
174 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
175
176 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
177 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
178 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
179 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
180 i4 += 16;
181
182 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
183 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
184 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
185 const __m128i vxk4x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vzero), vk_zero_point);
186
187 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
188 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
189 const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
190 const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
191
192 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
193 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
194 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
195 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
196
197 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
198 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
199 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
200 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
201 i5 += 16;
202
203 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
204 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
205 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
206 const __m128i vxk5x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vzero), vk_zero_point);
207
208 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
209 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
210 const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
211 const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
212
213 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
214 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
215 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
216 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
217
218 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
219 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
220 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
221 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
222 i6 += 16;
223
224 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
225 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
226 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
227 const __m128i vxk6x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vzero), vk_zero_point);
228
229 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
230 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
231 const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
232 const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
233
234 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
235 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
236 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
237 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
238
239 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
240 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
241 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
242 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
243 i7 += 16;
244
245 const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
246 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
247 const __m128i vxi7x89ABCDEF = _mm_unpacklo_epi8(vi7x89ABCDEF, vzero);
248 const __m128i vxk7x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vzero), vk_zero_point);
249
250 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
251 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
252 const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
253 const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
254
255 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
256 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
257 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
258 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
259
260 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
261 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
262 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
263 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
264 i8 += 16;
265
266 const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
267 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
268 const __m128i vxi8x89ABCDEF = _mm_unpacklo_epi8(vi8x89ABCDEF, vzero);
269 const __m128i vxk8x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vzero), vk_zero_point);
270
271 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
272 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
273 const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
274 const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
275
276 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
277 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
278 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
279 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
280
281 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t));
282
283 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
284 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
285 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
286 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
287
288 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
289 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
290 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
291 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
292 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
293
294 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
295 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
296 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
297 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
298 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
299
300 vacc0123 = _mm_cvtps_epi32(vscaled0123);
301 vacc4567 = _mm_cvtps_epi32(vscaled4567);
302 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
303 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
304
305 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
306 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
307 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
308
309 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
310
311 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
312 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
313
314 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
315 output += 16;
316 }
317 if XNN_UNLIKELY(c != 0) {
318 const uint8_t* k = (const uint8_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 vzero = _mm_setzero_si128();
329 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
330 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
331
332 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
333 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
334
335 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
336 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
337
338 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
339 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
340 i1 += 8;
341
342 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
343 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
344
345 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
346 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
347
348 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
349 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
350
351 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
352 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
353 i2 += 8;
354
355 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
356 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
357
358 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
359 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
360
361 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
362 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
363
364 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
365 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
366 i3 += 8;
367
368 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
369 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
370
371 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
372 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
373
374 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
375 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
376
377 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
378 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
379 i4 += 8;
380
381 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
382 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
383
384 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
385 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
386
387 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
388 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
389
390 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
391 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
392 i5 += 8;
393
394 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
395 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
396
397 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
398 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
399
400 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
401 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
402
403 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
404 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
405 i6 += 8;
406
407 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
408 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
409
410 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
411 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
412
413 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
414 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
415
416 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
417 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
418 i7 += 8;
419
420 const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
421 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
422
423 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
424 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
425
426 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
427 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
428
429 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
430 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
431 i8 += 8;
432
433 const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
434 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
435
436 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
437 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
438
439 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
440 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
441
442 k += 8;
443
444 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
445 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
446
447 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
448 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
449 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
450
451 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
452 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
453 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
454
455 vacc0123 = _mm_cvtps_epi32(vscaled0123);
456 vacc4567 = _mm_cvtps_epi32(vscaled4567);
457
458 w = (const void*) ((const int32_t*) w + 8);
459
460 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
461 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
462
463 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
464
465 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
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 = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
484 output += 1;
485 }
486 c = 0;
487 }
488 } while (c != 0);
489 }
490
491 output = (uint8_t*) ((uintptr_t) output + output_increment);
492 } while (--output_width != 0);
493 }
494