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 #include <xnnpack/unaligned.h>
21
22
xnn_qc8_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul16_add16(
24 size_t channels,
25 size_t output_width,
26 const int8_t** input,
27 const void* weights,
28 int8_t* output,
29 size_t input_stride,
30 size_t output_increment,
31 size_t input_offset,
32 const int8_t* zero,
33 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37
38 do {
39 const int8_t* i0 = input[0];
40 assert(i0 != NULL);
41 if XNN_UNPREDICTABLE(i0 != zero) {
42 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43 }
44 const int8_t* i1 = input[1];
45 assert(i1 != NULL);
46 if XNN_UNPREDICTABLE(i1 != zero) {
47 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48 }
49 const int8_t* i2 = input[2];
50 assert(i2 != NULL);
51 if XNN_UNPREDICTABLE(i2 != zero) {
52 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53 }
54 const int8_t* i3 = input[3];
55 assert(i3 != NULL);
56 if XNN_UNPREDICTABLE(i3 != zero) {
57 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58 }
59 const int8_t* i4 = input[4];
60 assert(i4 != NULL);
61 if XNN_UNPREDICTABLE(i4 != zero) {
62 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63 }
64 const int8_t* i5 = input[5];
65 assert(i5 != NULL);
66 if XNN_UNPREDICTABLE(i5 != zero) {
67 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68 }
69 const int8_t* i6 = input[6];
70 assert(i6 != NULL);
71 if XNN_UNPREDICTABLE(i6 != zero) {
72 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73 }
74 const int8_t* i7 = input[7];
75 assert(i7 != NULL);
76 if XNN_UNPREDICTABLE(i7 != zero) {
77 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78 }
79 const int8_t* i8 = input[8];
80 assert(i8 != NULL);
81 if XNN_UNPREDICTABLE(i8 != zero) {
82 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83 }
84 input = (const int8_t**) ((uintptr_t) input + input_stride);
85
86 size_t c = channels;
87 const void* w = weights;
88 for (; c >= 16; c -= 16) {
89 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
90 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
91 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
92 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
93
94
95 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
96 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
97 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
98 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
99 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
100 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
101 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
102 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
103 i0 += 16;
104
105
106 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
107 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
108
109
110 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
111 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
112 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
113 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
114 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
115 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
116 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
117 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
118 i1 += 16;
119
120
121 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
122 vprod89ABCDEF = _mm_macc_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF, vprod89ABCDEF);
123
124 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
125 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
126 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
127 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
128
129 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
130 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
131 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
132 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
133 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
134 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
135 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
136 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
137 i2 += 16;
138
139
140 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
141 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
142
143
144 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
145 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
146 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
147 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
148 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
149 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
150 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
151 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
152 i3 += 16;
153
154
155 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
156 vprod89ABCDEF = _mm_macc_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF, vprod89ABCDEF);
157
158 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
159 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
160 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
161 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
162
163 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
164 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
165 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
166 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
167 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
168 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
169 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
170 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
171 i4 += 16;
172
173
174 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
175 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
176
177
178 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
179 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
180 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
181 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
182 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
183 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
184 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
185 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
186 i5 += 16;
187
188
189 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
190 vprod89ABCDEF = _mm_macc_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF, vprod89ABCDEF);
191
192 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
193 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
194 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
195 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
196
197 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
198 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
199 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
200 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
201 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
202 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
203 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
204 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
205 i6 += 16;
206
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 vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
214 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
215 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
216 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
217 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
218 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
219 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
220 i7 += 16;
221
222
223 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
224 vprod89ABCDEF = _mm_macc_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF, vprod89ABCDEF);
225
226 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
227 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
228 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
229 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
230
231 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
232 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
233 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
234 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
235 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
236 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
237 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
238 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
239 i8 += 16;
240
241
242 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
243 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
244
245 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
246 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
247 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
248 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
249
250 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
251
252 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
253 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
254 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
255 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
256
257 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
258 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
259 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
260 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
261 w = (const void*) ((const float*) w + 16);
262 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
263 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
264 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
265 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
266
267 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.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_sse4.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
283 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
284
285 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
286 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
287
288 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
289 output += 16;
290 }
291 if XNN_UNLIKELY(c != 0) {
292 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
293 do {
294 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
295 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
296
297
298 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
299 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
300 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
301 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
302 i0 += 8;
303
304
305 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
306
307
308 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
309 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
310 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
311 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
312 i1 += 8;
313
314
315 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
316
317 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
318 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
319
320 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
321 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
322 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
323 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
324 i2 += 8;
325
326
327 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
328
329
330 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
331 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
332 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
333 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
334 i3 += 8;
335
336
337 vprod01234567 = _mm_macc_epi16(vxi3x01234567, vxk3x01234567, vprod01234567);
338
339 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
340 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
341
342 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
343 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
344 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
345 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
346 i4 += 8;
347
348
349 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
350
351
352 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
353 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
354 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
355 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
356 i5 += 8;
357
358
359 vprod01234567 = _mm_macc_epi16(vxi5x01234567, vxk5x01234567, vprod01234567);
360
361 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
362 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
363
364 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
365 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
366 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
367 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
368 i6 += 8;
369
370
371 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
372
373
374 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
375 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
376 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
377 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
378 i7 += 8;
379
380
381 vprod01234567 = _mm_macc_epi16(vxi7x01234567, vxk7x01234567, vprod01234567);
382
383 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
384 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
385
386 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
387 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
388 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
389 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
390 i8 += 8;
391
392
393 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
394
395 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
396 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
397
398 k += 8;
399
400 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
401 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
402
403 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
404 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t) + 4 * sizeof(float)));
405 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
406 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
407
408 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
409 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
410 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
411
412 vacc0123 = _mm_cvtps_epi32(vscaled0123);
413 vacc4567 = _mm_cvtps_epi32(vscaled4567);
414
415 w = (const void*) ((const int32_t*) w + 8);
416
417 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
418 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
419
420
421 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
422
423 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
424
425 if XNN_LIKELY(c >= 8) {
426 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
427 output += 8;
428 c -= 8;
429 } else {
430 if (c & 4) {
431 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
432 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
433 output += 4;
434 }
435 if (c & 2) {
436 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
437 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
438 output += 2;
439 }
440 if (c & 1) {
441 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
442 output += 1;
443 }
444 c = 0;
445 }
446 } while (c != 0);
447 }
448
449 output = (int8_t*) ((uintptr_t) output + output_increment);
450 } while (--output_width != 0);
451 }
452