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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/dwconv.h>
20
21
xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__xop_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)])22 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul16_add16(
23 size_t channels,
24 size_t output_width,
25 const int8_t** input,
26 const void* weights,
27 int8_t* output,
28 size_t input_stride,
29 size_t output_increment,
30 size_t input_offset,
31 const int8_t* zero,
32 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(channels != 0);
35 assert(output_width != 0);
36
37 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 input = (const int8_t**) ((uintptr_t) input + input_stride);
84
85 size_t c = channels;
86 const void* w = weights;
87 for (; c >= 16; c -= 16) {
88 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
89 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
90 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
91 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
92
93
94 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
95 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
96 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
97 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
98 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
99 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
100 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
101 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
102 i0 += 16;
103
104
105 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
106 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
107
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_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
121 vprod89ABCDEF = _mm_macc_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF, vprod89ABCDEF);
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
143 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
144 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
145 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
146 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
147 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
148 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
149 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
150 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
151 i3 += 16;
152
153
154 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
155 vprod89ABCDEF = _mm_macc_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF, vprod89ABCDEF);
156
157 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
158 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
159 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
160 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
161
162 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
163 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
164 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
165 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
166 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
167 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
168 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
169 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
170 i4 += 16;
171
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 vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
179 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
180 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
181 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
182 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
183 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
184 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
185 i5 += 16;
186
187
188 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
189 vprod89ABCDEF = _mm_macc_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF, vprod89ABCDEF);
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
193 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
194 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
195
196 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
197 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
198 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
199 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
200 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
201 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
202 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
203 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
204 i6 += 16;
205
206
207 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
208 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
209
210
211 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
212 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
213 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
214 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
215 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
216 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
217 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
218 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
219 i7 += 16;
220
221
222 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
223 vprod89ABCDEF = _mm_macc_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF, vprod89ABCDEF);
224
225 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
226 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
227 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
228 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
229
230 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
231 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
232 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
233 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
234 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
235 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
236 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
237 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
238 i8 += 16;
239
240
241 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
242 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
243
244 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
245 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
246 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
247 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
248
249 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
250
251 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
252 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
253 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
254 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
255
256 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
257 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
258 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
259 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
260 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
261
262 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
263 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
264 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
265 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
266 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
267
268 vacc0123 = _mm_cvtps_epi32(vscaled0123);
269 vacc4567 = _mm_cvtps_epi32(vscaled4567);
270 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
271 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
272
273 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
274 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
275 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
276
277
278 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
279
280 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
281 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
282
283 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
284 output += 16;
285 }
286 if XNN_UNLIKELY(c != 0) {
287 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
288 do {
289 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
290 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
291
292
293 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
294 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
295 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
296 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
297 i0 += 8;
298
299
300 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
301
302
303 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
304 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
305 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
306 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
307 i1 += 8;
308
309
310 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
311
312 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
313 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
314
315 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
316 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
317 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
318 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
319 i2 += 8;
320
321
322 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
323
324
325 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
326 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
327 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
328 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
329 i3 += 8;
330
331
332 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
333
334 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
335 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
336
337 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
338 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
339 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
340 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
341 i4 += 8;
342
343
344 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
345
346
347 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
348 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
349 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
350 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
351 i5 += 8;
352
353
354 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
355
356 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
357 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
358
359 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
360 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
361 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
362 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
363 i6 += 8;
364
365
366 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
367
368
369 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
370 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
371 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
372 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
373 i7 += 8;
374
375
376 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
377
378 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
379 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
380
381 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
382 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
383 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
384 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
385 i8 += 8;
386
387
388 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
389
390 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
391 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
392
393 k += 8;
394
395 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
396 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
397
398 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
399 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
400 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
401
402 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
403 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
404 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
405
406 vacc0123 = _mm_cvtps_epi32(vscaled0123);
407 vacc4567 = _mm_cvtps_epi32(vscaled4567);
408
409 w = (const void*) ((const int32_t*) w + 8);
410
411 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
412 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
413
414
415 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
416
417 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
418
419 if XNN_LIKELY(c >= 8) {
420 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
421 output += 8;
422 c -= 8;
423 } else {
424 if (c & 4) {
425 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
426 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
427 output += 4;
428 }
429 if (c & 2) {
430 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
431 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
432 output += 2;
433 }
434 if (c & 1) {
435 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
436 output += 1;
437 }
438 c = 0;
439 }
440 } while (c != 0);
441 }
442
443 output = (int8_t*) ((uintptr_t) output + output_increment);
444 } while (--output_width != 0);
445 }
446