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_up24x9__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_up24x9__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 >= 24; c -= 24) {
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 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
94 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
95
96
97 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
98 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t))));
99 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 4));
100 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 4 * sizeof(int8_t))));
101 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 8));
102 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t))));
103 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 12));
104 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 12 * sizeof(int8_t))));
105 const __m128i vi0xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 16));
106 const __m128i vk0xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t))));
107 const __m128i vi0xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 20));
108 const __m128i vk0xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 20 * sizeof(int8_t))));
109 i0 += 24;
110
111 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
112 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
113 vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
114 vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
115 vaccGHIJ = _mm_macc_epi32(vi0xGHIJ, vk0xGHIJ, vaccGHIJ);
116 vaccKLMN = _mm_macc_epi32(vi0xKLMN, vk0xKLMN, vaccKLMN);
117
118 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
119 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t))));
120 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 4));
121 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 28 * sizeof(int8_t))));
122 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 8));
123 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t))));
124 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 12));
125 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 36 * sizeof(int8_t))));
126 const __m128i vi1xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 16));
127 const __m128i vk1xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t))));
128 const __m128i vi1xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 20));
129 const __m128i vk1xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 44 * sizeof(int8_t))));
130 i1 += 24;
131
132 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
133 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
134 vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
135 vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
136 vaccGHIJ = _mm_macc_epi32(vi1xGHIJ, vk1xGHIJ, vaccGHIJ);
137 vaccKLMN = _mm_macc_epi32(vi1xKLMN, vk1xKLMN, vaccKLMN);
138
139 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
140 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t))));
141 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 4));
142 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 52 * sizeof(int8_t))));
143 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 8));
144 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t))));
145 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 12));
146 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 60 * sizeof(int8_t))));
147 const __m128i vi2xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 16));
148 const __m128i vk2xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t))));
149 const __m128i vi2xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 20));
150 const __m128i vk2xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 68 * sizeof(int8_t))));
151 i2 += 24;
152
153 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
154 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
155 vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
156 vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
157 vaccGHIJ = _mm_macc_epi32(vi2xGHIJ, vk2xGHIJ, vaccGHIJ);
158 vaccKLMN = _mm_macc_epi32(vi2xKLMN, vk2xKLMN, vaccKLMN);
159
160 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
161 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t))));
162 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 4));
163 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 76 * sizeof(int8_t))));
164 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 8));
165 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t))));
166 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 12));
167 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 84 * sizeof(int8_t))));
168 const __m128i vi3xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 16));
169 const __m128i vk3xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t))));
170 const __m128i vi3xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 20));
171 const __m128i vk3xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 92 * sizeof(int8_t))));
172 i3 += 24;
173
174 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
175 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
176 vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
177 vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
178 vaccGHIJ = _mm_macc_epi32(vi3xGHIJ, vk3xGHIJ, vaccGHIJ);
179 vaccKLMN = _mm_macc_epi32(vi3xKLMN, vk3xKLMN, vaccKLMN);
180
181 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
182 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t))));
183 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 4));
184 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 100 * sizeof(int8_t))));
185 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 8));
186 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t))));
187 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 12));
188 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 108 * sizeof(int8_t))));
189 const __m128i vi4xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 16));
190 const __m128i vk4xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t))));
191 const __m128i vi4xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 20));
192 const __m128i vk4xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 116 * sizeof(int8_t))));
193 i4 += 24;
194
195 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
196 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
197 vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
198 vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
199 vaccGHIJ = _mm_macc_epi32(vi4xGHIJ, vk4xGHIJ, vaccGHIJ);
200 vaccKLMN = _mm_macc_epi32(vi4xKLMN, vk4xKLMN, vaccKLMN);
201
202 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
203 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t))));
204 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 4));
205 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 124 * sizeof(int8_t))));
206 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 8));
207 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t))));
208 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 12));
209 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 132 * sizeof(int8_t))));
210 const __m128i vi5xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 16));
211 const __m128i vk5xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t))));
212 const __m128i vi5xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 20));
213 const __m128i vk5xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 140 * sizeof(int8_t))));
214 i5 += 24;
215
216 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
217 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
218 vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
219 vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
220 vaccGHIJ = _mm_macc_epi32(vi5xGHIJ, vk5xGHIJ, vaccGHIJ);
221 vaccKLMN = _mm_macc_epi32(vi5xKLMN, vk5xKLMN, vaccKLMN);
222
223 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
224 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t))));
225 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 4));
226 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 148 * sizeof(int8_t))));
227 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 8));
228 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t))));
229 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 12));
230 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 156 * sizeof(int8_t))));
231 const __m128i vi6xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 16));
232 const __m128i vk6xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t))));
233 const __m128i vi6xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 20));
234 const __m128i vk6xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 164 * sizeof(int8_t))));
235 i6 += 24;
236
237 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
238 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
239 vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
240 vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
241 vaccGHIJ = _mm_macc_epi32(vi6xGHIJ, vk6xGHIJ, vaccGHIJ);
242 vaccKLMN = _mm_macc_epi32(vi6xKLMN, vk6xKLMN, vaccKLMN);
243
244 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
245 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t))));
246 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 4));
247 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 172 * sizeof(int8_t))));
248 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 8));
249 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t))));
250 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 12));
251 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 180 * sizeof(int8_t))));
252 const __m128i vi7xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 16));
253 const __m128i vk7xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t))));
254 const __m128i vi7xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 20));
255 const __m128i vk7xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 188 * sizeof(int8_t))));
256 i7 += 24;
257
258 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
259 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
260 vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
261 vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
262 vaccGHIJ = _mm_macc_epi32(vi7xGHIJ, vk7xGHIJ, vaccGHIJ);
263 vaccKLMN = _mm_macc_epi32(vi7xKLMN, vk7xKLMN, vaccKLMN);
264
265 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
266 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t))));
267 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 4));
268 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 196 * sizeof(int8_t))));
269 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 8));
270 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t))));
271 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 12));
272 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 204 * sizeof(int8_t))));
273 const __m128i vi8xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 16));
274 const __m128i vk8xGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t))));
275 const __m128i vi8xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 20));
276 const __m128i vk8xKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 212 * sizeof(int8_t))));
277 i8 += 24;
278
279 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
280 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
281 vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
282 vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
283 vaccGHIJ = _mm_macc_epi32(vi8xGHIJ, vk8xGHIJ, vaccGHIJ);
284 vaccKLMN = _mm_macc_epi32(vi8xKLMN, vk8xKLMN, vaccKLMN);
285
286 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t));
287
288 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
289 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
290 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
291 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
292 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
293 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
294
295 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
296 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
297 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
298 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
299 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
300 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale);
301 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale);
302
303 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
304 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
305 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
306 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
307 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
308 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
309 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
310
311 vacc0123 = _mm_cvtps_epi32(vscaled0123);
312 vacc4567 = _mm_cvtps_epi32(vscaled4567);
313 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
314 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
315 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
316 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
317
318 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
319 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
320 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
321 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
322
323 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
324 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
325 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
326 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
327 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
328
329 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
330 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
331 output += 24;
332 }
333 if XNN_UNLIKELY(c != 0) {
334 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
335 do {
336 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
337
338 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
339 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(k));
340 i0 += 4;
341
342 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
343 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
344 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 24)));
345 i1 += 4;
346
347 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
348 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
349 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 48)));
350 i2 += 4;
351
352 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
353 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
354 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 72)));
355 i3 += 4;
356
357 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
358 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
359 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 96)));
360 i4 += 4;
361
362 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
363 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
364 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 120)));
365 i5 += 4;
366
367 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
368 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
369 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 144)));
370 i6 += 4;
371
372 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
373 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
374 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 168)));
375 i7 += 4;
376
377 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
378 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
379 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 192)));
380 i8 += 4;
381
382 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
383
384 k += 4;
385
386 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
387 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
388 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
389 vacc0123 = _mm_cvtps_epi32(vscaled0123);
390
391 w = (const void*) ((const int32_t*) w + 4);
392
393 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
394 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
395
396 vout0123 = _mm_packs_epi16(vout0123, vout0123);
397 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
398
399 if XNN_LIKELY(c >= 4) {
400 _mm_storeu_si32(output, vout0123);
401 output += 4;
402 c -= 4;
403 } else {
404 if (c & 2) {
405 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
406 vout0123 = _mm_srli_epi32(vout0123, 16);
407 output += 2;
408 }
409 if (c & 1) {
410 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
411 output += 1;
412 }
413 c = 0;
414 }
415 } while (c != 0);
416 }
417
418 output = (int8_t*) ((uintptr_t) output + output_increment);
419 } while (--output_width != 0);
420 }
421