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_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)])17 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16(
18 size_t channels,
19 size_t output_width,
20 const uint8_t** input,
21 const void* weights,
22 uint8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const uint8_t* zero,
27 const union xnn_qu8_conv_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 uint8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const uint8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const uint8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const uint8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const uint8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const uint8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const uint8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const uint8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const uint8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 input = (const uint8_t**) ((uintptr_t) input + input_stride);
79
80 size_t c = channels;
81 const void* w = weights;
82 const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
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(uint8_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(uint8_t)));
94 i0 += 16;
95
96 const __m128i vzero = _mm_setzero_si128();
97 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
98 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
99 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
100 const __m128i vxk0x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vzero), vk_zero_point);
101
102 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
103 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
104 const __m128i vprod0x89ABCDEFlo = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
105 const __m128i vprod0x89ABCDEFhi = _mm_mulhi_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
106
107 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
108 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
109 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
110 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod0x89ABCDEFlo, vprod0x89ABCDEFhi));
111
112 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
113 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
114 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
115 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
116 i1 += 16;
117
118 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
119 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
120 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
121 const __m128i vxk1x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vzero), vk_zero_point);
122
123 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
124 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
125 const __m128i vprod1x89ABCDEFlo = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
126 const __m128i vprod1x89ABCDEFhi = _mm_mulhi_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
127
128 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
129 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
130 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
131 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod1x89ABCDEFlo, vprod1x89ABCDEFhi));
132
133 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
134 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
135 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
136 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
137 i2 += 16;
138
139 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
140 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
141 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
142 const __m128i vxk2x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vzero), vk_zero_point);
143
144 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
145 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
146 const __m128i vprod2x89ABCDEFlo = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
147 const __m128i vprod2x89ABCDEFhi = _mm_mulhi_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
148
149 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
150 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
151 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
152 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod2x89ABCDEFlo, vprod2x89ABCDEFhi));
153
154 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
155 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
156 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
157 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
158 i3 += 16;
159
160 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
161 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
162 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
163 const __m128i vxk3x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vzero), vk_zero_point);
164
165 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
166 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
167 const __m128i vprod3x89ABCDEFlo = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
168 const __m128i vprod3x89ABCDEFhi = _mm_mulhi_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
169
170 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
171 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
172 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
173 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod3x89ABCDEFlo, vprod3x89ABCDEFhi));
174
175 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
176 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
177 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
178 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
179 i4 += 16;
180
181 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
182 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
183 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
184 const __m128i vxk4x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vzero), vk_zero_point);
185
186 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
187 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
188 const __m128i vprod4x89ABCDEFlo = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
189 const __m128i vprod4x89ABCDEFhi = _mm_mulhi_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
193 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
194 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod4x89ABCDEFlo, vprod4x89ABCDEFhi));
195
196 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
197 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
198 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
199 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
200 i5 += 16;
201
202 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
203 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
204 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
205 const __m128i vxk5x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vzero), vk_zero_point);
206
207 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
208 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
209 const __m128i vprod5x89ABCDEFlo = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
210 const __m128i vprod5x89ABCDEFhi = _mm_mulhi_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
211
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
214 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
215 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod5x89ABCDEFlo, vprod5x89ABCDEFhi));
216
217 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
218 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
219 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
220 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
221 i6 += 16;
222
223 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
224 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
225 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
226 const __m128i vxk6x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vzero), vk_zero_point);
227
228 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
229 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
230 const __m128i vprod6x89ABCDEFlo = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
231 const __m128i vprod6x89ABCDEFhi = _mm_mulhi_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
232
233 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
234 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
235 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
236 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod6x89ABCDEFlo, vprod6x89ABCDEFhi));
237
238 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
239 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
240 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
241 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
242 i7 += 16;
243
244 const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero);
245 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
246 const __m128i vxi7x89ABCDEF = _mm_unpacklo_epi8(vi7x89ABCDEF, vzero);
247 const __m128i vxk7x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vzero), vk_zero_point);
248
249 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
250 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
251 const __m128i vprod7x89ABCDEFlo = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
252 const __m128i vprod7x89ABCDEFhi = _mm_mulhi_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
253
254 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
255 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
256 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
257 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod7x89ABCDEFlo, vprod7x89ABCDEFhi));
258
259 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
260 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
261 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
262 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
263 i8 += 16;
264
265 const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero);
266 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
267 const __m128i vxi8x89ABCDEF = _mm_unpacklo_epi8(vi8x89ABCDEF, vzero);
268 const __m128i vxk8x89ABCDEF = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vzero), vk_zero_point);
269
270 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
271 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
272 const __m128i vprod8x89ABCDEFlo = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
273 const __m128i vprod8x89ABCDEFhi = _mm_mulhi_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
274
275 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
276 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
277 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
278 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod8x89ABCDEFlo, vprod8x89ABCDEFhi));
279
280 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t));
281
282 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
283 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
284 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
285 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
286
287 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
288 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
289 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
290 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
291 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
292
293 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
294 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
295 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
296 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
297 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
298
299 vacc0123 = _mm_cvtps_epi32(vscaled0123);
300 vacc4567 = _mm_cvtps_epi32(vscaled4567);
301 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
302 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
303
304 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
305 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
306 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
307
308 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
309
310 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
311 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
312
313 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
314 output += 16;
315 }
316 if XNN_UNLIKELY(c != 0) {
317 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
318 do {
319 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
320 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
321
322
323 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
324 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
325 i0 += 8;
326
327 const __m128i vzero = _mm_setzero_si128();
328 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
329 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point);
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_unpacklo_epi8(vi1x01234567, vzero);
342 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point);
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_unpacklo_epi8(vi2x01234567, vzero);
355 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point);
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_unpacklo_epi8(vi3x01234567, vzero);
368 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point);
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_unpacklo_epi8(vi4x01234567, vzero);
381 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point);
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_unpacklo_epi8(vi5x01234567, vzero);
394 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point);
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_unpacklo_epi8(vi6x01234567, vzero);
407 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point);
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_unpacklo_epi8(vi7x01234567, vzero);
420 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point);
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_unpacklo_epi8(vi8x01234567, vzero);
433 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point);
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 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
463
464 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
465
466 if XNN_LIKELY(c >= 8) {
467 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
468 output += 8;
469 c -= 8;
470 } else {
471 if (c & 4) {
472 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
473 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
474 output += 4;
475 }
476 if (c & 2) {
477 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
478 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
479 output += 2;
480 }
481 if (c & 1) {
482 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
483 output += 1;
484 }
485 c = 0;
486 }
487 } while (c != 0);
488 }
489
490 output = (uint8_t*) ((uintptr_t) output + output_increment);
491 } while (--output_width != 0);
492 }
493