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 #include <xnnpack/unaligned.h>
17
18
xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__avx_mul32(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 size_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(channels != 0);
32 assert(output_width != 0);
33
34 do {
35 const int8_t* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
39 }
40 const int8_t* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
44 }
45 const int8_t* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
49 }
50 const int8_t* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
54 }
55 const int8_t* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
59 }
60 const int8_t* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
64 }
65 const int8_t* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
69 }
70 const int8_t* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
74 }
75 const int8_t* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
79 }
80 input = (const int8_t**) ((uintptr_t) input + input_stride);
81
82 size_t c = channels;
83 const void* w = weights;
84 for (; c >= 24; c -= 24) {
85 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
86 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
87 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
88 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
89 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
90 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
91
92
93 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
94 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
95 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
96 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
97 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
98 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
99 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
100 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
101 const __m128i vi0xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 16)));
102 const __m128i vk0xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
103 const __m128i vi0xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 20)));
104 const __m128i vk0xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
105 i0 += 24;
106
107 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
108 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
109 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
110 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
111 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi0xGHIJ, vk0xGHIJ));
112 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi0xKLMN, vk0xKLMN));
113
114 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
115 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
116 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
117 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
118 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
119 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
120 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
121 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
122 const __m128i vi1xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 16)));
123 const __m128i vk1xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
124 const __m128i vi1xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 20)));
125 const __m128i vk1xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
126 i1 += 24;
127
128 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
129 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
130 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
131 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
132 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi1xGHIJ, vk1xGHIJ));
133 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi1xKLMN, vk1xKLMN));
134
135 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
136 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
137 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
138 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
139 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
140 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
141 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
142 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
143 const __m128i vi2xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 16)));
144 const __m128i vk2xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)))));
145 const __m128i vi2xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 20)));
146 const __m128i vk2xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 68 * sizeof(int8_t)))));
147 i2 += 24;
148
149 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
150 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
151 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
152 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
153 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi2xGHIJ, vk2xGHIJ));
154 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi2xKLMN, vk2xKLMN));
155
156 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
157 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)))));
158 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
159 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 76 * sizeof(int8_t)))));
160 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
161 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)))));
162 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
163 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 84 * sizeof(int8_t)))));
164 const __m128i vi3xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 16)));
165 const __m128i vk3xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)))));
166 const __m128i vi3xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 20)));
167 const __m128i vk3xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 92 * sizeof(int8_t)))));
168 i3 += 24;
169
170 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
171 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
172 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
173 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
174 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi3xGHIJ, vk3xGHIJ));
175 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi3xKLMN, vk3xKLMN));
176
177 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
178 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)))));
179 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
180 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 100 * sizeof(int8_t)))));
181 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
182 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)))));
183 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
184 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 108 * sizeof(int8_t)))));
185 const __m128i vi4xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 16)));
186 const __m128i vk4xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)))));
187 const __m128i vi4xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 20)));
188 const __m128i vk4xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 116 * sizeof(int8_t)))));
189 i4 += 24;
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
193 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
194 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
195 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi4xGHIJ, vk4xGHIJ));
196 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi4xKLMN, vk4xKLMN));
197
198 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
199 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)))));
200 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
201 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 124 * sizeof(int8_t)))));
202 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
203 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)))));
204 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
205 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 132 * sizeof(int8_t)))));
206 const __m128i vi5xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 16)));
207 const __m128i vk5xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)))));
208 const __m128i vi5xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 20)));
209 const __m128i vk5xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 140 * sizeof(int8_t)))));
210 i5 += 24;
211
212 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
213 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
214 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
215 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
216 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi5xGHIJ, vk5xGHIJ));
217 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi5xKLMN, vk5xKLMN));
218
219 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
220 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)))));
221 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
222 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 148 * sizeof(int8_t)))));
223 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
224 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)))));
225 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
226 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 156 * sizeof(int8_t)))));
227 const __m128i vi6xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 16)));
228 const __m128i vk6xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)))));
229 const __m128i vi6xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 20)));
230 const __m128i vk6xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 164 * sizeof(int8_t)))));
231 i6 += 24;
232
233 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
234 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
235 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi6x89AB, vk6x89AB));
236 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi6xCDEF, vk6xCDEF));
237 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi6xGHIJ, vk6xGHIJ));
238 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi6xKLMN, vk6xKLMN));
239
240 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
241 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)))));
242 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
243 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 172 * sizeof(int8_t)))));
244 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
245 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)))));
246 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
247 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 180 * sizeof(int8_t)))));
248 const __m128i vi7xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 16)));
249 const __m128i vk7xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)))));
250 const __m128i vi7xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 20)));
251 const __m128i vk7xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 188 * sizeof(int8_t)))));
252 i7 += 24;
253
254 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
255 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
256 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi7x89AB, vk7x89AB));
257 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi7xCDEF, vk7xCDEF));
258 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi7xGHIJ, vk7xGHIJ));
259 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi7xKLMN, vk7xKLMN));
260
261 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
262 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)))));
263 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
264 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 196 * sizeof(int8_t)))));
265 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
266 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)))));
267 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
268 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 204 * sizeof(int8_t)))));
269 const __m128i vi8xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 16)));
270 const __m128i vk8xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)))));
271 const __m128i vi8xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 20)));
272 const __m128i vk8xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 212 * sizeof(int8_t)))));
273 i8 += 24;
274
275 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
276 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
277 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi8x89AB, vk8x89AB));
278 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi8xCDEF, vk8xCDEF));
279 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_mullo_epi32(vi8xGHIJ, vk8xGHIJ));
280 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_mullo_epi32(vi8xKLMN, vk8xKLMN));
281
282 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t));
283
284 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
285 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
286 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
287 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
288 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
289 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
290
291 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
292 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
293 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
294 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
295 const __m128 vscaleGHIJ = _mm_loadu_ps((const float*) w + 16);
296 const __m128 vscaleKLMN = _mm_loadu_ps((const float*) w + 20);
297 w = (const void*) ((const float*) w + 24);
298 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
299 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
300 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
301 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
302 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscaleGHIJ);
303 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscaleKLMN);
304
305 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
306 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
307 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
308 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
309 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
310 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
311 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
312
313 vacc0123 = _mm_cvtps_epi32(vscaled0123);
314 vacc4567 = _mm_cvtps_epi32(vscaled4567);
315 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
316 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
317 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
318 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
319
320 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
321 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
322 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
323 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
324
325 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
326 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
327 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
328 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
329 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
330
331 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
332 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
333 output += 24;
334 }
335 if XNN_UNLIKELY(c != 0) {
336 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
337 do {
338 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
339
340 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
341 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) k)));
342 i0 += 4;
343
344 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
345 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
346 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 24))));
347 i1 += 4;
348
349 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
350 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
351 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48))));
352 i2 += 4;
353
354 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
355 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
356 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 72))));
357 i3 += 4;
358
359 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
360 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
361 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96))));
362 i4 += 4;
363
364 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
365 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
366 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 120))));
367 i5 += 4;
368
369 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
370 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
371 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144))));
372 i6 += 4;
373
374 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
375 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
376 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 168))));
377 i7 += 4;
378
379 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
380 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
381 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192))));
382 i8 += 4;
383
384 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
385
386 k += 4;
387
388 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
389 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t)));
390 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
391 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
392 vacc0123 = _mm_cvtps_epi32(vscaled0123);
393
394 w = (const void*) ((const int32_t*) w + 4);
395
396 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
397 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
398
399 vout0123 = _mm_packs_epi16(vout0123, vout0123);
400 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
401
402 if XNN_LIKELY(c >= 4) {
403 _mm_storeu_si32(output, vout0123);
404 output += 4;
405 c -= 4;
406 } else {
407 if (c & 2) {
408 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
409 vout0123 = _mm_srli_epi32(vout0123, 16);
410 output += 2;
411 }
412 if (c & 1) {
413 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
414 output += 1;
415 }
416 c = 0;
417 }
418 } while (c != 0);
419 }
420
421 output = (int8_t*) ((uintptr_t) output + output_increment);
422 } while (--output_width != 0);
423 }
424