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_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)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_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_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
105 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
106 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
107 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
108 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
109 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
110 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
111 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
112 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
113 i1 += 16;
114
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 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
120 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
121 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
122 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
123
124 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
125 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
126 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
127 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
128 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
129 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
130 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
131 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
132 i2 += 16;
133
134
135 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
136 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
137
138
139 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
140 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
141 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
142 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
143 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
144 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(vi3x89ABCDEF);
145 const __m128i vk3x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
146 const __m128i vxk3x89ABCDEF = _mm_cvtepi8_epi16(vk3x89ABCDEF);
147 i3 += 16;
148
149
150 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
151 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi3x89ABCDEF, vxk3x89ABCDEF));
152
153 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
154 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
155 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
156 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
157
158 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
159 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
160 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
161 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
162 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
163 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(vi4x89ABCDEF);
164 const __m128i vk4x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
165 const __m128i vxk4x89ABCDEF = _mm_cvtepi8_epi16(vk4x89ABCDEF);
166 i4 += 16;
167
168
169 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
170 vprod89ABCDEF = _mm_mullo_epi16(vxi4x89ABCDEF, vxk4x89ABCDEF);
171
172
173 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
174 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
175 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
176 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
177 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
178 const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(vi5x89ABCDEF);
179 const __m128i vk5x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
180 const __m128i vxk5x89ABCDEF = _mm_cvtepi8_epi16(vk5x89ABCDEF);
181 i5 += 16;
182
183
184 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
185 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi5x89ABCDEF, vxk5x89ABCDEF));
186
187 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
188 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
189 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
190 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
191
192 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
193 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
194 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
195 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
196 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
197 const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(vi6x89ABCDEF);
198 const __m128i vk6x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
199 const __m128i vxk6x89ABCDEF = _mm_cvtepi8_epi16(vk6x89ABCDEF);
200 i6 += 16;
201
202
203 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
204 vprod89ABCDEF = _mm_mullo_epi16(vxi6x89ABCDEF, vxk6x89ABCDEF);
205
206
207 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
208 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
209 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
210 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
211 const __m128i vi7x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i7 + 8));
212 const __m128i vxi7x89ABCDEF = _mm_cvtepi8_epi16(vi7x89ABCDEF);
213 const __m128i vk7x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
214 const __m128i vxk7x89ABCDEF = _mm_cvtepi8_epi16(vk7x89ABCDEF);
215 i7 += 16;
216
217
218 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
219 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi7x89ABCDEF, vxk7x89ABCDEF));
220
221 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
222 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
223 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
224 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
225
226 const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
227 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
228 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
229 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
230 const __m128i vi8x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i8 + 8));
231 const __m128i vxi8x89ABCDEF = _mm_cvtepi8_epi16(vi8x89ABCDEF);
232 const __m128i vk8x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
233 const __m128i vxk8x89ABCDEF = _mm_cvtepi8_epi16(vk8x89ABCDEF);
234 i8 += 16;
235
236
237 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
238 vprod89ABCDEF = _mm_mullo_epi16(vxi8x89ABCDEF, vxk8x89ABCDEF);
239
240 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
241 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
242 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
243 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
244
245 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
246
247 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
248 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
249 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
250 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
251
252 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
253 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
254 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
255 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
256 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
257
258 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
259 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
260 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
261 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
262 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
263
264 vacc0123 = _mm_cvtps_epi32(vscaled0123);
265 vacc4567 = _mm_cvtps_epi32(vscaled4567);
266 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
267 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
268
269 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
270 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
271 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
272
273
274 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
275
276 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
277 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
278
279 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
280 output += 16;
281 }
282 if XNN_UNLIKELY(c != 0) {
283 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
284 do {
285 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
286 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
287
288
289 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
290 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
291 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
292 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
293 i0 += 8;
294
295
296 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
297
298
299 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
300 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
301 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
302 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
303 i1 += 8;
304
305
306 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
307
308 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
309 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
310
311 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
312 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
313 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
314 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
315 i2 += 8;
316
317
318 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
319
320
321 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
322 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
323 const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) (k + 48));
324 const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
325 i3 += 8;
326
327
328 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567));
329
330 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
331 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
332
333 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
334 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
335 const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) (k + 64));
336 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
337 i4 += 8;
338
339
340 vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
341
342
343 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
344 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
345 const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) (k + 80));
346 const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
347 i5 += 8;
348
349
350 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567));
351
352 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
353 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
354
355 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
356 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
357 const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) (k + 96));
358 const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
359 i6 += 8;
360
361
362 vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
363
364
365 const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
366 const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
367 const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) (k + 112));
368 const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
369 i7 += 8;
370
371
372 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567));
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 vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
378 const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
379 const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) (k + 128));
380 const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
381 i8 += 8;
382
383
384 vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
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 k += 8;
390
391 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
392 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
393
394 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
395 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
396 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
397
398 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
399 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
400 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
401
402 vacc0123 = _mm_cvtps_epi32(vscaled0123);
403 vacc4567 = _mm_cvtps_epi32(vscaled4567);
404
405 w = (const void*) ((const int32_t*) w + 8);
406
407 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
408 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
409
410
411 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
412
413 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
414
415 if XNN_LIKELY(c >= 8) {
416 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
417 output += 8;
418 c -= 8;
419 } else {
420 if (c & 4) {
421 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
422 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
423 output += 4;
424 }
425 if (c & 2) {
426 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
427 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
428 output += 2;
429 }
430 if (c & 1) {
431 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
432 output += 1;
433 }
434 c = 0;
435 }
436 } while (c != 0);
437 }
438
439 output = (int8_t*) ((uintptr_t) output + output_increment);
440 } while (--output_width != 0);
441 }
442