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 #include <xnnpack/unaligned.h>
16
17
xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16(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)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16(
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_qs8_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 vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
92 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
93 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
94 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
95 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
96 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
97 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
98 i0 += 16;
99
100
101 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
102 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
103
104 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
105 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
106 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
107 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
108
109 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
110 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
111 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
112 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
113 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
114 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
115 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
116 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
117 i1 += 16;
118
119
120 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
121 vprod89ABCDEF = _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF);
122
123 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
124 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
125 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
126 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
127
128 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
129 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
130 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
131 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
132 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
133 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
134 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
135 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
136 i2 += 16;
137
138
139 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
140 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
141
142 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
143 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
144 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
145 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
146
147 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
148 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
149 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
150 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
151 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
152 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
153 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
154 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
155 i3 += 16;
156
157
158 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
159 vprod89ABCDEF = _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF);
160
161 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
162 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
163 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
164 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
165
166 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
167 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
168 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
169 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
170 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
171 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
172 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
173 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
174 i4 += 16;
175
176
177 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
178 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
179
180 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
181 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
182 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
183 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
184
185 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
186 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
187 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
188 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
189 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
190 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
191 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
192 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
193 i5 += 16;
194
195
196 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
197 vprod89ABCDEF = _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF);
198
199 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
200 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
201 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
202 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
203
204 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
205 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
206 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
207 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
208 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
209 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
210 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
211 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
212 i6 += 16;
213
214
215 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
216 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
217
218 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
219 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
220 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
221 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
222
223 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
224 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
225 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
226 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
227 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
228 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
229 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
230 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
231 i7 += 16;
232
233
234 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
235 vprod89ABCDEF = _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF);
236
237 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
238 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
239 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
240 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
241
242 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
243 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
244 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
245 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
246 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
247 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
248 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
249 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
250 i8 += 16;
251
252
253 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
254 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
255
256 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
257 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
258 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
259 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
260
261 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
262
263 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
264 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
265 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
266 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
267
268 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
269 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
270 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
271 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
272 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
273
274 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
275 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
276 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
277 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
278 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
279
280 vacc0123 = _mm_cvtps_epi32(vscaled0123);
281 vacc4567 = _mm_cvtps_epi32(vscaled4567);
282 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
283 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
284
285 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
286 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
287 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
288
289
290 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
291
292 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
293 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
294
295 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
296 output += 16;
297 }
298 if XNN_UNLIKELY(c != 0) {
299 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
300 do {
301 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
302 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
303
304
305 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
306 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
307 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
308 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
309 i0 += 8;
310
311
312 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
313
314 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
315 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
316
317 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
318 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
319 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
320 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
321 i1 += 8;
322
323
324 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
325
326 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
327 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
328
329 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
330 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
331 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
332 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
333 i2 += 8;
334
335
336 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
337
338 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
339 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
340
341 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
342 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
343 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
344 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
345 i3 += 8;
346
347
348 vprod01234567 = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
349
350 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
351 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
352
353 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
354 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
355 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
356 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
357 i4 += 8;
358
359
360 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
361
362 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
363 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
364
365 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
366 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
367 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
368 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
369 i5 += 8;
370
371
372 vprod01234567 = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
373
374 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
375 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
376
377 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
378 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
379 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
380 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
381 i6 += 8;
382
383
384 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
385
386 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
387 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
388
389 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
390 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
391 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
392 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
393 i7 += 8;
394
395
396 vprod01234567 = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
397
398 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
399 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
400
401 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
402 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
403 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
404 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
405 i8 += 8;
406
407
408 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
409
410 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
411 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
412
413 k += 8;
414
415 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
416 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
417
418 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
419 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
420 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
421
422 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
423 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
424 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
425
426 vacc0123 = _mm_cvtps_epi32(vscaled0123);
427 vacc4567 = _mm_cvtps_epi32(vscaled4567);
428
429 w = (const void*) ((const int32_t*) w + 8);
430
431 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
432 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
433
434
435 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
436
437 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
438
439 if XNN_LIKELY(c >= 8) {
440 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
441 output += 8;
442 c -= 8;
443 } else {
444 if (c & 4) {
445 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
446 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
447 output += 4;
448 }
449 if (c & 2) {
450 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
451 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
452 output += 2;
453 }
454 if (c & 1) {
455 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
456 output += 1;
457 }
458 c = 0;
459 }
460 } while (c != 0);
461 }
462
463 output = (int8_t*) ((uintptr_t) output + output_increment);
464 } while (--output_width != 0);
465 }
466