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_up24x9__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_up24x9__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 >= 24; c -= 24) {
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 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
89 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
90
91
92 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
93 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t))));
94 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 4));
95 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 4 * sizeof(int8_t))));
96 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 8));
97 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t))));
98 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 12));
99 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 12 * sizeof(int8_t))));
100 const __m128i vi0xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 16));
101 const __m128i vk0xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t))));
102 const __m128i vi0xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 20));
103 const __m128i vk0xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 20 * sizeof(int8_t))));
104 i0 += 24;
105
106 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
107 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
108 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
109 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
110 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ));
111 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN));
112
113 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
114 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t))));
115 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 4));
116 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 28 * sizeof(int8_t))));
117 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 8));
118 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t))));
119 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 12));
120 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 36 * sizeof(int8_t))));
121 const __m128i vi1xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 16));
122 const __m128i vk1xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t))));
123 const __m128i vi1xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 20));
124 const __m128i vk1xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 44 * sizeof(int8_t))));
125 i1 += 24;
126
127 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
128 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
129 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
130 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
131 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi1xGHIJ, vk1xGHIJ));
132 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi1xKLMN, vk1xKLMN));
133
134 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
135 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t))));
136 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 4));
137 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 52 * sizeof(int8_t))));
138 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 8));
139 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t))));
140 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 12));
141 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 60 * sizeof(int8_t))));
142 const __m128i vi2xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 16));
143 const __m128i vk2xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t))));
144 const __m128i vi2xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 20));
145 const __m128i vk2xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 68 * sizeof(int8_t))));
146 i2 += 24;
147
148 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
149 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
150 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
151 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
152 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi2xGHIJ, vk2xGHIJ));
153 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi2xKLMN, vk2xKLMN));
154
155 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
156 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t))));
157 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 4));
158 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 76 * sizeof(int8_t))));
159 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 8));
160 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t))));
161 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 12));
162 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 84 * sizeof(int8_t))));
163 const __m128i vi3xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 16));
164 const __m128i vk3xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t))));
165 const __m128i vi3xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 20));
166 const __m128i vk3xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 92 * sizeof(int8_t))));
167 i3 += 24;
168
169 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
170 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
171 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
172 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
173 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi3xGHIJ, vk3xGHIJ));
174 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi3xKLMN, vk3xKLMN));
175
176 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
177 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t))));
178 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 4));
179 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 100 * sizeof(int8_t))));
180 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 8));
181 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t))));
182 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 12));
183 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 108 * sizeof(int8_t))));
184 const __m128i vi4xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 16));
185 const __m128i vk4xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t))));
186 const __m128i vi4xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 20));
187 const __m128i vk4xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 116 * sizeof(int8_t))));
188 i4 += 24;
189
190 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
191 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
192 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
193 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
194 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi4xGHIJ, vk4xGHIJ));
195 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi4xKLMN, vk4xKLMN));
196
197 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
198 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t))));
199 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 4));
200 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 124 * sizeof(int8_t))));
201 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 8));
202 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t))));
203 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 12));
204 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 132 * sizeof(int8_t))));
205 const __m128i vi5xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 16));
206 const __m128i vk5xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t))));
207 const __m128i vi5xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 20));
208 const __m128i vk5xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 140 * sizeof(int8_t))));
209 i5 += 24;
210
211 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
212 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
213 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
214 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
215 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi5xGHIJ, vk5xGHIJ));
216 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi5xKLMN, vk5xKLMN));
217
218 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
219 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t))));
220 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 4));
221 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 148 * sizeof(int8_t))));
222 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 8));
223 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t))));
224 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 12));
225 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 156 * sizeof(int8_t))));
226 const __m128i vi6xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 16));
227 const __m128i vk6xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t))));
228 const __m128i vi6xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 20));
229 const __m128i vk6xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 164 * sizeof(int8_t))));
230 i6 += 24;
231
232 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
233 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
234 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi6x89AB, vk6x89AB));
235 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi6xCDEF, vk6xCDEF));
236 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi6xGHIJ, vk6xGHIJ));
237 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi6xKLMN, vk6xKLMN));
238
239 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
240 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t))));
241 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 4));
242 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 172 * sizeof(int8_t))));
243 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 8));
244 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t))));
245 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 12));
246 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 180 * sizeof(int8_t))));
247 const __m128i vi7xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 16));
248 const __m128i vk7xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t))));
249 const __m128i vi7xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 20));
250 const __m128i vk7xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 188 * sizeof(int8_t))));
251 i7 += 24;
252
253 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
254 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
255 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi7x89AB, vk7x89AB));
256 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi7xCDEF, vk7xCDEF));
257 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi7xGHIJ, vk7xGHIJ));
258 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi7xKLMN, vk7xKLMN));
259
260 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
261 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t))));
262 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 4));
263 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 196 * sizeof(int8_t))));
264 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 8));
265 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t))));
266 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 12));
267 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 204 * sizeof(int8_t))));
268 const __m128i vi8xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 16));
269 const __m128i vk8xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t))));
270 const __m128i vi8xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 20));
271 const __m128i vk8xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 212 * sizeof(int8_t))));
272 i8 += 24;
273
274 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
275 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
276 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi8x89AB, vk8x89AB));
277 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi8xCDEF, vk8xCDEF));
278 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi8xGHIJ, vk8xGHIJ));
279 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi8xKLMN, vk8xKLMN));
280
281 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t));
282
283 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
284 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
285 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
286 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
287 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
288 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
289
290 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
291 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
292 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
293 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
294 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
295 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale);
296 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale);
297
298 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
299 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
300 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
301 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
302 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
303 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
304 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
305
306 vacc0123 = _mm_cvtps_epi32(vscaled0123);
307 vacc4567 = _mm_cvtps_epi32(vscaled4567);
308 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
309 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
310 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
311 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
312
313 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
314 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
315 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
316 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
317
318 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
319 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
320 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
321 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
322 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
323
324 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
325 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
326 output += 24;
327 }
328 if XNN_UNLIKELY(c != 0) {
329 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
330 do {
331 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
332
333 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
334 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(k));
335 i0 += 4;
336
337 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
338 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
339 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 24)));
340 i1 += 4;
341
342 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
343 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
344 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 48)));
345 i2 += 4;
346
347 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
348 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
349 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 72)));
350 i3 += 4;
351
352 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
353 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
354 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 96)));
355 i4 += 4;
356
357 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
358 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
359 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 120)));
360 i5 += 4;
361
362 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
363 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
364 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 144)));
365 i6 += 4;
366
367 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
368 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
369 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 168)));
370 i7 += 4;
371
372 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
373 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
374 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 192)));
375 i8 += 4;
376
377 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
378
379 k += 4;
380
381 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
382 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
383 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
384 vacc0123 = _mm_cvtps_epi32(vscaled0123);
385
386 w = (const void*) ((const int32_t*) w + 4);
387
388 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
389 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
390
391 vout0123 = _mm_packs_epi16(vout0123, vout0123);
392 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
393
394 if XNN_LIKELY(c >= 4) {
395 _mm_storeu_si32(output, vout0123);
396 output += 4;
397 c -= 4;
398 } else {
399 if (c & 2) {
400 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
401 vout0123 = _mm_srli_epi32(vout0123, 16);
402 output += 2;
403 }
404 if (c & 1) {
405 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
406 output += 1;
407 }
408 c = 0;
409 }
410 } while (c != 0);
411 }
412
413 output = (int8_t*) ((uintptr_t) output + output_increment);
414 } while (--output_width != 0);
415 }
416