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 #include <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_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)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32(
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 vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
91 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t))));
92 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 4));
93 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(int8_t))));
94 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 8));
95 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t))));
96 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 12));
97 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(int8_t))));
98 i0 += 16;
99
100 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
101 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
102 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
103 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
104
105 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
106 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t))));
107 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 4));
108 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(int8_t))));
109 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 8));
110 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t))));
111 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 12));
112 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(int8_t))));
113 i1 += 16;
114
115 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
116 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
117 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
118 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
119
120 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
121 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t))));
122 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 4));
123 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(int8_t))));
124 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 8));
125 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t))));
126 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 12));
127 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(int8_t))));
128 i2 += 16;
129
130 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
131 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
132 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
133 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
134
135 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
136 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t))));
137 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 4));
138 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(int8_t))));
139 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 8));
140 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t))));
141 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 12));
142 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(int8_t))));
143 i3 += 16;
144
145 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
146 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
147 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
148 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
149
150 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
151 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t))));
152 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 4));
153 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(int8_t))));
154 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 8));
155 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t))));
156 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 12));
157 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(int8_t))));
158 i4 += 16;
159
160 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
161 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
162 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
163 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
164
165 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
166 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t))));
167 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 4));
168 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(int8_t))));
169 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 8));
170 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t))));
171 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 12));
172 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(int8_t))));
173 i5 += 16;
174
175 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
176 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
177 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
178 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
179
180 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
181 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t))));
182 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 4));
183 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(int8_t))));
184 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 8));
185 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t))));
186 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 12));
187 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(int8_t))));
188 i6 += 16;
189
190 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
191 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
192 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi6x89AB, vk6x89AB));
193 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi6xCDEF, vk6xCDEF));
194
195 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
196 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t))));
197 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 4));
198 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(int8_t))));
199 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 8));
200 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t))));
201 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 12));
202 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(int8_t))));
203 i7 += 16;
204
205 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
206 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
207 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi7x89AB, vk7x89AB));
208 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi7xCDEF, vk7xCDEF));
209
210 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
211 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t))));
212 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 4));
213 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(int8_t))));
214 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 8));
215 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t))));
216 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 12));
217 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(int8_t))));
218 i8 += 16;
219
220 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
221 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
222 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi8x89AB, vk8x89AB));
223 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi8xCDEF, vk8xCDEF));
224
225 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
226
227 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
228 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
229 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
230 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
231
232 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
233 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
234 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
235 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
236 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
237
238 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
239 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
240 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
241 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
242 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
243
244 vacc0123 = _mm_cvtps_epi32(vscaled0123);
245 vacc4567 = _mm_cvtps_epi32(vscaled4567);
246 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
247 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
248
249 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
250 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
251 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
252
253 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
254 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
255 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
256
257 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
258 output += 16;
259 }
260 if XNN_UNLIKELY(c != 0) {
261 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
262 do {
263 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
264
265 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
266 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(k));
267 i0 += 4;
268
269 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
270 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
271 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 16)));
272 i1 += 4;
273
274 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
275 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
276 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 32)));
277 i2 += 4;
278
279 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
280 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
281 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 48)));
282 i3 += 4;
283
284 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
285 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
286 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 64)));
287 i4 += 4;
288
289 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
290 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
291 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 80)));
292 i5 += 4;
293
294 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
295 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
296 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 96)));
297 i6 += 4;
298
299 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
300 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
301 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 112)));
302 i7 += 4;
303
304 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
305 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
306 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 128)));
307 i8 += 4;
308
309 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
310
311 k += 4;
312
313 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
314 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
315 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
316 vacc0123 = _mm_cvtps_epi32(vscaled0123);
317
318 w = (const void*) ((const int32_t*) w + 4);
319
320 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
321 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
322
323 vout0123 = _mm_packs_epi16(vout0123, vout0123);
324 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
325
326 if XNN_LIKELY(c >= 4) {
327 _mm_storeu_si32(output, vout0123);
328 output += 4;
329 c -= 4;
330 } else {
331 if (c & 2) {
332 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
333 vout0123 = _mm_srli_epi32(vout0123, 16);
334 output += 2;
335 }
336 if (c & 1) {
337 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
338 output += 1;
339 }
340 c = 0;
341 }
342 } while (c != 0);
343 }
344
345 output = (int8_t*) ((uintptr_t) output + output_increment);
346 } while (--output_width != 0);
347 }
348