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_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_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__sse41_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 vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
92 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
93 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
94 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
95 const __m128i vxi0x89ABCDEF = _mm_cvtepu8_epi16(vi0x89ABCDEF);
96 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
97 const __m128i vxk0x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x89ABCDEF), vk_zero_point);
98 i0 += 16;
99
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 vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
113 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
114 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
115 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
116 const __m128i vxi1x89ABCDEF = _mm_cvtepu8_epi16(vi1x89ABCDEF);
117 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
118 const __m128i vxk1x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x89ABCDEF), vk_zero_point);
119 i1 += 16;
120
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 vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
134 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
135 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
136 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
137 const __m128i vxi2x89ABCDEF = _mm_cvtepu8_epi16(vi2x89ABCDEF);
138 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
139 const __m128i vxk2x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x89ABCDEF), vk_zero_point);
140 i2 += 16;
141
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 vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
155 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
156 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
157 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
158 const __m128i vxi3x89ABCDEF = _mm_cvtepu8_epi16(vi3x89ABCDEF);
159 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
160 const __m128i vxk3x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x89ABCDEF), vk_zero_point);
161 i3 += 16;
162
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 vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
176 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
177 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
178 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
179 const __m128i vxi4x89ABCDEF = _mm_cvtepu8_epi16(vi4x89ABCDEF);
180 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
181 const __m128i vxk4x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x89ABCDEF), vk_zero_point);
182 i4 += 16;
183
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 vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
197 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
198 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
199 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
200 const __m128i vxi5x89ABCDEF = _mm_cvtepu8_epi16(vi5x89ABCDEF);
201 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
202 const __m128i vxk5x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x89ABCDEF), vk_zero_point);
203 i5 += 16;
204
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 vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
218 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
219 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
220 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
221 const __m128i vxi6x89ABCDEF = _mm_cvtepu8_epi16(vi6x89ABCDEF);
222 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
223 const __m128i vxk6x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x89ABCDEF), vk_zero_point);
224 i6 += 16;
225
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 vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
239 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
240 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
241 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
242 const __m128i vxi7x89ABCDEF = _mm_cvtepu8_epi16(vi7x89ABCDEF);
243 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
244 const __m128i vxk7x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x89ABCDEF), vk_zero_point);
245 i7 += 16;
246
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 vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
260 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
261 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
262 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
263 const __m128i vxi8x89ABCDEF = _mm_cvtepu8_epi16(vi8x89ABCDEF);
264 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
265 const __m128i vxk8x89ABCDEF = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x89ABCDEF), vk_zero_point);
266 i8 += 16;
267
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(uint8_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 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
308
309 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
310 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
311
312 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
313 output += 16;
314 }
315 if XNN_UNLIKELY(c != 0) {
316 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
317 do {
318 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
319 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
320
321
322 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
323 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(vi0x01234567);
324 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
325 const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk0x01234567), vk_zero_point);
326 i0 += 8;
327
328
329 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
330 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
331
332 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
333 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
334
335 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
336 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(vi1x01234567);
337 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
338 const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk1x01234567), vk_zero_point);
339 i1 += 8;
340
341
342 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
343 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
344
345 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
346 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
347
348 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
349 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(vi2x01234567);
350 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
351 const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk2x01234567), vk_zero_point);
352 i2 += 8;
353
354
355 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
356 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
357
358 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
359 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
360
361 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
362 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(vi3x01234567);
363 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
364 const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk3x01234567), vk_zero_point);
365 i3 += 8;
366
367
368 const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
369 const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
370
371 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi));
372 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi));
373
374 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
375 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(vi4x01234567);
376 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
377 const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk4x01234567), vk_zero_point);
378 i4 += 8;
379
380
381 const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
382 const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
383
384 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi));
385 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi));
386
387 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
388 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(vi5x01234567);
389 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
390 const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk5x01234567), vk_zero_point);
391 i5 += 8;
392
393
394 const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
395 const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
396
397 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi));
398 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi));
399
400 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
401 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(vi6x01234567);
402 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
403 const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk6x01234567), vk_zero_point);
404 i6 += 8;
405
406
407 const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
408 const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
409
410 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi));
411 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi));
412
413 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
414 const __m128i vxi7x01234567 = _mm_cvtepu8_epi16(vi7x01234567);
415 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
416 const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk7x01234567), vk_zero_point);
417 i7 += 8;
418
419
420 const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
421 const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
422
423 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi));
424 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi));
425
426 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
427 const __m128i vxi8x01234567 = _mm_cvtepu8_epi16(vi8x01234567);
428 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
429 const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_cvtepu8_epi16(vk8x01234567), vk_zero_point);
430 i8 += 8;
431
432
433 const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
434 const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
435
436 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi));
437 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi));
438
439 k += 8;
440
441 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
442 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
443
444 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
445 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
446 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
447
448 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
449 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
450 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
451
452 vacc0123 = _mm_cvtps_epi32(vscaled0123);
453 vacc4567 = _mm_cvtps_epi32(vscaled4567);
454
455 w = (const void*) ((const int32_t*) w + 8);
456
457 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
458 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
459
460 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
461
462 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
463
464 if XNN_LIKELY(c >= 8) {
465 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
466 output += 8;
467 c -= 8;
468 } else {
469 if (c & 4) {
470 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
471 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
472 output += 4;
473 }
474 if (c & 2) {
475 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
476 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
477 output += 2;
478 }
479 if (c & 1) {
480 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
481 output += 1;
482 }
483 c = 0;
484 }
485 } while (c != 0);
486 }
487
488 output = (uint8_t*) ((uintptr_t) output + output_increment);
489 } while (--output_width != 0);
490 }
491