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_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16_add16(
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_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 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*) ((const int32_t*) w + 4));
85 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
86 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
87
88
89 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
90 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
91 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
92 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
93 i0 += 16;
94
95 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
96 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
97 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
98 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
99
100 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
101 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
102
103
104 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
105 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
106 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
107 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
108 i1 += 16;
109
110 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
111 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
112 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
113 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
114
115 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
116 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
117
118 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
119 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
120 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
121 const __m128i vsignprod1x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
122 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
123 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
124
125 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
126 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
127 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
128 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
129 i2 += 16;
130
131 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
132 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
133 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
134 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
135
136 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
137 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
138
139
140 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
141 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
142 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
143 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
144 i3 += 16;
145
146 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
147 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
148 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
149 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
150
151 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
152 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
153
154 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
155 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
156 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
157 const __m128i vsignprod3x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
158 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
159 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
160
161 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
162 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
163 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
164 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
165 i4 += 16;
166
167 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
168 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
169 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
170 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
171
172 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
173 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
174
175
176 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
177 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
178 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
179 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
180 i5 += 16;
181
182 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
183 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
184 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
185 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
186
187 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
188 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
189
190 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
193 const __m128i vsignprod5x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
194 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
195 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
196
197 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
198 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
199 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
200 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
201 i6 += 16;
202
203 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
204 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
205 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
206 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
207
208 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
209 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
210
211
212 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
213 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
214 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
215 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
216 i7 += 16;
217
218 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
219 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
220 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
221 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
222
223 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
224 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
225
226 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
227 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
228 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
229 const __m128i vsignprod7x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
230 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
231 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
232
233 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
234 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
235 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
236 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
237 i8 += 16;
238
239 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
240 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
241 const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
242 const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
243
244 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
245 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
246
247 const __m128i vsignprod8x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
248 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod8x01234567));
249 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod8x01234567));
250 const __m128i vsignprod8x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
251 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod8x89ABCDEF));
252 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod8x89ABCDEF));
253
254 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
255
256 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
257 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
258 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
259 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
260
261 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
262 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
263 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
264 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
265 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
266
267 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
268 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
269 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
270 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
271 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
272
273 vacc0123 = _mm_cvtps_epi32(vscaled0123);
274 vacc4567 = _mm_cvtps_epi32(vscaled4567);
275 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
276 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
277
278 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
279 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
280 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
281
282 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
283 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
284 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
285
286 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
287
288
289 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
290 output += 16;
291 }
292 if XNN_UNLIKELY(c != 0) {
293 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
294 do {
295 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
296 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
297
298
299 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
300 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
301 i0 += 8;
302
303 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
304 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
305
306 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
307
308
309 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
310 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
311 i1 += 8;
312
313 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
314 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
315
316 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
317
318 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
319 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
320 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
321
322 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
323 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
324 i2 += 8;
325
326 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
327 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
328
329 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
330
331
332 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
333 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
334 i3 += 8;
335
336 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
337 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
338
339 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
340
341 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
342 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
343 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
344
345 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
346 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
347 i4 += 8;
348
349 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
350 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
351
352 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
353
354
355 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
356 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
357 i5 += 8;
358
359 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
360 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
361
362 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
363
364 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
365 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
366 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
367
368 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
369 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
370 i6 += 8;
371
372 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
373 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
374
375 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
376
377
378 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
379 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
380 i7 += 8;
381
382 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
383 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
384
385 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
386
387 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
388 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
389 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
390
391 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
392 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
393 i8 += 8;
394
395 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
396 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
397
398 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
399
400 const __m128i vsignprod8x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
401 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod8x01234567));
402 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod8x01234567));
403
404 k += 8;
405
406 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
407 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
408
409 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
410 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
411 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
412
413 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
414 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
415 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
416
417 vacc0123 = _mm_cvtps_epi32(vscaled0123);
418 vacc4567 = _mm_cvtps_epi32(vscaled4567);
419
420 w = (const void*) ((const int32_t*) w + 8);
421
422 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
423 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
424
425 vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
426
427 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
428
429
430 if XNN_LIKELY(c >= 8) {
431 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
432 output += 8;
433 c -= 8;
434 } else {
435 if (c & 4) {
436 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
437 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
438 output += 4;
439 }
440 if (c & 2) {
441 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
442 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
443 output += 2;
444 }
445 if (c & 1) {
446 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
447 output += 1;
448 }
449 c = 0;
450 }
451 } while (c != 0);
452 }
453
454 output = (int8_t*) ((uintptr_t) output + output_increment);
455 } while (--output_width != 0);
456 }
457