1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul32.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/intrinsics-polyfill.h>
21
22
xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul32(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)])23 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul32(
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_qs8_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 vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
96 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
97 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 4));
98 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(int8_t))));
99 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 8));
100 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t))));
101 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 12));
102 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(int8_t))));
103 i0 += 16;
104
105 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
106 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
107 vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
108 vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
109
110 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
111 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
112 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 4));
113 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(int8_t))));
114 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 8));
115 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t))));
116 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 12));
117 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(int8_t))));
118 i1 += 16;
119
120 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
121 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
122 vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
123 vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
124
125 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
126 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
127 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 4));
128 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(int8_t))));
129 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 8));
130 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t))));
131 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 12));
132 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(int8_t))));
133 i2 += 16;
134
135 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
136 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
137 vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
138 vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
139
140 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
141 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
142 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 4));
143 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(int8_t))));
144 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 8));
145 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t))));
146 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 12));
147 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(int8_t))));
148 i3 += 16;
149
150 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
151 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
152 vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
153 vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
154
155 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
156 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
157 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 4));
158 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(int8_t))));
159 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 8));
160 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t))));
161 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 12));
162 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(int8_t))));
163 i4 += 16;
164
165 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
166 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
167 vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
168 vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
169
170 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
171 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
172 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 4));
173 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(int8_t))));
174 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 8));
175 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t))));
176 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 12));
177 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(int8_t))));
178 i5 += 16;
179
180 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
181 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
182 vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
183 vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
184
185 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
186 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
187 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 4));
188 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(int8_t))));
189 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 8));
190 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t))));
191 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 12));
192 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(int8_t))));
193 i6 += 16;
194
195 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
196 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
197 vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
198 vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
199
200 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
201 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
202 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 4));
203 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(int8_t))));
204 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 8));
205 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t))));
206 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 12));
207 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(int8_t))));
208 i7 += 16;
209
210 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
211 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
212 vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
213 vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
214
215 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
216 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
217 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 4));
218 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(int8_t))));
219 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 8));
220 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t))));
221 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 12));
222 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(int8_t))));
223 i8 += 16;
224
225 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
226 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
227 vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
228 vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
229
230 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
231
232 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
233 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
234 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
235 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
236
237 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
238 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
239 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
240 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
241 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
242
243 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
244 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
245 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
246 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
247 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
248
249 vacc0123 = _mm_cvtps_epi32(vscaled0123);
250 vacc4567 = _mm_cvtps_epi32(vscaled4567);
251 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
252 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
253
254 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
255 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
256 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
257
258 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
259 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
260 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
261
262 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
263 output += 16;
264 }
265 if XNN_UNLIKELY(c != 0) {
266 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
267 do {
268 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
269
270 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
271 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(k));
272 i0 += 4;
273
274 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
275 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
276 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 16)));
277 i1 += 4;
278
279 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
280 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
281 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 32)));
282 i2 += 4;
283
284 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
285 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
286 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 48)));
287 i3 += 4;
288
289 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
290 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
291 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 64)));
292 i4 += 4;
293
294 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
295 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
296 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 80)));
297 i5 += 4;
298
299 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
300 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
301 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 96)));
302 i6 += 4;
303
304 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
305 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
306 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 112)));
307 i7 += 4;
308
309 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
310 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
311 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 128)));
312 i8 += 4;
313
314 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
315
316 k += 4;
317
318 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
319 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
320 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
321 vacc0123 = _mm_cvtps_epi32(vscaled0123);
322
323 w = (const void*) ((const int32_t*) w + 4);
324
325 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
326 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
327
328 vout0123 = _mm_packs_epi16(vout0123, vout0123);
329 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
330
331 if XNN_LIKELY(c >= 4) {
332 _mm_storeu_si32(output, vout0123);
333 output += 4;
334 c -= 4;
335 } else {
336 if (c & 2) {
337 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
338 vout0123 = _mm_srli_epi32(vout0123, 16);
339 output += 2;
340 }
341 if (c & 1) {
342 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
343 output += 1;
344 }
345 c = 0;
346 }
347 } while (c != 0);
348 }
349
350 output = (int8_t*) ((uintptr_t) output + output_increment);
351 } while (--output_width != 0);
352 }
353