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_qc8_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16_add16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qc8_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 int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const int8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const int8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const int8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const int8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const int8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const int8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 input = (const int8_t**) ((uintptr_t) input + input_stride);
80
81 size_t c = channels;
82 const void* w = weights;
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(int8_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(int8_t)));
94 i0 += 16;
95
96 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
97 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
98 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
99 const __m128i vxk0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x89ABCDEF, vk0x89ABCDEF), 8);
100
101 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
102 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
103
104
105 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
106 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
107 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
108 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
109 i1 += 16;
110
111 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
112 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
113 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
114 const __m128i vxk1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x89ABCDEF, vk1x89ABCDEF), 8);
115
116 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
117 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
118
119 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
120 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
121 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
122 const __m128i vsignprod1x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
123 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
124 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod1x89ABCDEF));
125
126 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
127 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
128 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
129 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
130 i2 += 16;
131
132 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
133 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
134 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
135 const __m128i vxk2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x89ABCDEF, vk2x89ABCDEF), 8);
136
137 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
138 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
139
140
141 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
142 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
143 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
144 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
145 i3 += 16;
146
147 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
148 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
149 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
150 const __m128i vxk3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x89ABCDEF, vk3x89ABCDEF), 8);
151
152 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
153 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
154
155 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
156 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
157 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
158 const __m128i vsignprod3x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
159 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
160 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod3x89ABCDEF));
161
162 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
163 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
164 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
165 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
166 i4 += 16;
167
168 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
169 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
170 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
171 const __m128i vxk4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x89ABCDEF, vk4x89ABCDEF), 8);
172
173 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
174 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
175
176
177 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
178 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
179 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
180 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
181 i5 += 16;
182
183 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
184 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
185 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
186 const __m128i vxk5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x89ABCDEF, vk5x89ABCDEF), 8);
187
188 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
189 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
190
191 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
192 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
193 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
194 const __m128i vsignprod5x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
195 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
196 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod5x89ABCDEF));
197
198 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
199 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
200 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
201 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
202 i6 += 16;
203
204 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
205 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
206 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
207 const __m128i vxk6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x89ABCDEF, vk6x89ABCDEF), 8);
208
209 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
210 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
211
212
213 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
214 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
215 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
216 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
217 i7 += 16;
218
219 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
220 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
221 const __m128i vxi7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x89ABCDEF, vi7x89ABCDEF), 8);
222 const __m128i vxk7x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x89ABCDEF, vk7x89ABCDEF), 8);
223
224 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
225 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
226
227 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
228 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
229 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
230 const __m128i vsignprod7x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
231 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
232 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod7x89ABCDEF));
233
234 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
235 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
236 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
237 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
238 i8 += 16;
239
240 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
241 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
242 const __m128i vxi8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x89ABCDEF, vi8x89ABCDEF), 8);
243 const __m128i vxk8x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x89ABCDEF, vk8x89ABCDEF), 8);
244
245 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
246 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
247
248 const __m128i vsignprod8x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
249 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod8x01234567));
250 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod8x01234567));
251 const __m128i vsignprod8x89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod89ABCDEF);
252 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vprod89ABCDEF, vsignprod8x89ABCDEF));
253 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vprod89ABCDEF, vsignprod8x89ABCDEF));
254
255 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
256
257 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
258 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
259 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
260 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
261
262 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
263 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
264 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
265 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
266 w = (const void*) ((const float*) w + 16);
267 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
268 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
269 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
270 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
271
272 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
273 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
274 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
275 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
276 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
277
278 vacc0123 = _mm_cvtps_epi32(vscaled0123);
279 vacc4567 = _mm_cvtps_epi32(vscaled4567);
280 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
281 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
282
283 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
284 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
285 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
286
287 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
288 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
289 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
290
291 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
292
293
294 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
295 output += 16;
296 }
297 if XNN_UNLIKELY(c != 0) {
298 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
299 do {
300 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
301 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
302
303
304 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
305 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
306 i0 += 8;
307
308 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
309 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
310
311 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
312
313
314 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
315 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
316 i1 += 8;
317
318 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
319 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
320
321 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
322
323 const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
324 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567));
325 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567));
326
327 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
328 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
329 i2 += 8;
330
331 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
332 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
333
334 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
335
336
337 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
338 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
339 i3 += 8;
340
341 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
342 const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8);
343
344 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
345
346 const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
347 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567));
348 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567));
349
350 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
351 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
352 i4 += 8;
353
354 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
355 const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8);
356
357 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
358
359
360 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
361 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
362 i5 += 8;
363
364 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
365 const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8);
366
367 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
368
369 const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
370 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567));
371 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567));
372
373 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
374 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
375 i6 += 8;
376
377 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
378 const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8);
379
380 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
381
382
383 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
384 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
385 i7 += 8;
386
387 const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8);
388 const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8);
389
390 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
391
392 const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
393 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567));
394 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567));
395
396 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
397 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
398 i8 += 8;
399
400 const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8);
401 const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8);
402
403 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
404
405 const __m128i vsignprod8x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567);
406 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod8x01234567));
407 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod8x01234567));
408
409 k += 8;
410
411 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
412 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
413
414 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
415 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t) + 4 * sizeof(float)));
416 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
417 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
418
419 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
420 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
421 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
422
423 vacc0123 = _mm_cvtps_epi32(vscaled0123);
424 vacc4567 = _mm_cvtps_epi32(vscaled4567);
425
426 w = (const void*) ((const int32_t*) w + 8);
427
428 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
429 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
430
431 vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
432
433 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
434
435
436 if XNN_LIKELY(c >= 8) {
437 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
438 output += 8;
439 c -= 8;
440 } else {
441 if (c & 4) {
442 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
443 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
444 output += 4;
445 }
446 if (c & 2) {
447 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
448 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
449 output += 2;
450 }
451 if (c & 1) {
452 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
453 output += 1;
454 }
455 c = 0;
456 }
457 } while (c != 0);
458 }
459
460 output = (int8_t*) ((uintptr_t) output + output_increment);
461 } while (--output_width != 0);
462 }
463