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 #include <xnnpack/unaligned.h>
22
23
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)])24 void xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__xop_mul32(
25 size_t channels,
26 size_t output_width,
27 const int8_t** input,
28 const void* weights,
29 int8_t* output,
30 size_t input_stride,
31 size_t output_increment,
32 size_t input_offset,
33 const int8_t* zero,
34 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(channels != 0);
37 assert(output_width != 0);
38
39 do {
40 const int8_t* i0 = input[0];
41 assert(i0 != NULL);
42 if XNN_UNPREDICTABLE(i0 != zero) {
43 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
44 }
45 const int8_t* i1 = input[1];
46 assert(i1 != NULL);
47 if XNN_UNPREDICTABLE(i1 != zero) {
48 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
49 }
50 const int8_t* i2 = input[2];
51 assert(i2 != NULL);
52 if XNN_UNPREDICTABLE(i2 != zero) {
53 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
54 }
55 const int8_t* i3 = input[3];
56 assert(i3 != NULL);
57 if XNN_UNPREDICTABLE(i3 != zero) {
58 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
59 }
60 const int8_t* i4 = input[4];
61 assert(i4 != NULL);
62 if XNN_UNPREDICTABLE(i4 != zero) {
63 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
64 }
65 const int8_t* i5 = input[5];
66 assert(i5 != NULL);
67 if XNN_UNPREDICTABLE(i5 != zero) {
68 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
69 }
70 const int8_t* i6 = input[6];
71 assert(i6 != NULL);
72 if XNN_UNPREDICTABLE(i6 != zero) {
73 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
74 }
75 const int8_t* i7 = input[7];
76 assert(i7 != NULL);
77 if XNN_UNPREDICTABLE(i7 != zero) {
78 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
79 }
80 const int8_t* i8 = input[8];
81 assert(i8 != NULL);
82 if XNN_UNPREDICTABLE(i8 != zero) {
83 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
84 }
85 input = (const int8_t**) ((uintptr_t) input + input_stride);
86
87 size_t c = channels;
88 const void* w = weights;
89 for (; c >= 24; c -= 24) {
90 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
91 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
92 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
93 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
94 __m128i vaccGHIJ = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 16));
95 __m128i vaccKLMN = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 20));
96
97
98 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
99 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
100 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
101 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
102 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
103 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
104 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
105 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
106 const __m128i vi0xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 16)));
107 const __m128i vk0xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
108 const __m128i vi0xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 20)));
109 const __m128i vk0xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
110 i0 += 24;
111
112 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
113 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
114 vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
115 vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
116 vaccGHIJ = _mm_macc_epi32(vi0xGHIJ, vk0xGHIJ, vaccGHIJ);
117 vaccKLMN = _mm_macc_epi32(vi0xKLMN, vk0xKLMN, vaccKLMN);
118
119 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
120 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
121 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
122 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
123 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
124 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
125 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
126 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
127 const __m128i vi1xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 16)));
128 const __m128i vk1xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
129 const __m128i vi1xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 20)));
130 const __m128i vk1xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
131 i1 += 24;
132
133 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
134 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
135 vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
136 vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
137 vaccGHIJ = _mm_macc_epi32(vi1xGHIJ, vk1xGHIJ, vaccGHIJ);
138 vaccKLMN = _mm_macc_epi32(vi1xKLMN, vk1xKLMN, vaccKLMN);
139
140 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
141 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
142 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
143 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
144 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
145 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
146 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
147 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
148 const __m128i vi2xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 16)));
149 const __m128i vk2xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)))));
150 const __m128i vi2xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 20)));
151 const __m128i vk2xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 68 * sizeof(int8_t)))));
152 i2 += 24;
153
154 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
155 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
156 vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
157 vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
158 vaccGHIJ = _mm_macc_epi32(vi2xGHIJ, vk2xGHIJ, vaccGHIJ);
159 vaccKLMN = _mm_macc_epi32(vi2xKLMN, vk2xKLMN, vaccKLMN);
160
161 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
162 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)))));
163 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
164 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 76 * sizeof(int8_t)))));
165 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
166 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)))));
167 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
168 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 84 * sizeof(int8_t)))));
169 const __m128i vi3xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 16)));
170 const __m128i vk3xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)))));
171 const __m128i vi3xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 20)));
172 const __m128i vk3xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 92 * sizeof(int8_t)))));
173 i3 += 24;
174
175 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
176 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
177 vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
178 vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
179 vaccGHIJ = _mm_macc_epi32(vi3xGHIJ, vk3xGHIJ, vaccGHIJ);
180 vaccKLMN = _mm_macc_epi32(vi3xKLMN, vk3xKLMN, vaccKLMN);
181
182 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
183 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)))));
184 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
185 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 100 * sizeof(int8_t)))));
186 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
187 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)))));
188 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
189 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 108 * sizeof(int8_t)))));
190 const __m128i vi4xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 16)));
191 const __m128i vk4xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)))));
192 const __m128i vi4xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 20)));
193 const __m128i vk4xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 116 * sizeof(int8_t)))));
194 i4 += 24;
195
196 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
197 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
198 vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
199 vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
200 vaccGHIJ = _mm_macc_epi32(vi4xGHIJ, vk4xGHIJ, vaccGHIJ);
201 vaccKLMN = _mm_macc_epi32(vi4xKLMN, vk4xKLMN, vaccKLMN);
202
203 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
204 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)))));
205 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
206 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 124 * sizeof(int8_t)))));
207 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
208 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)))));
209 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
210 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 132 * sizeof(int8_t)))));
211 const __m128i vi5xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 16)));
212 const __m128i vk5xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)))));
213 const __m128i vi5xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 20)));
214 const __m128i vk5xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 140 * sizeof(int8_t)))));
215 i5 += 24;
216
217 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
218 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
219 vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
220 vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
221 vaccGHIJ = _mm_macc_epi32(vi5xGHIJ, vk5xGHIJ, vaccGHIJ);
222 vaccKLMN = _mm_macc_epi32(vi5xKLMN, vk5xKLMN, vaccKLMN);
223
224 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
225 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)))));
226 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
227 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 148 * sizeof(int8_t)))));
228 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
229 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)))));
230 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
231 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 156 * sizeof(int8_t)))));
232 const __m128i vi6xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 16)));
233 const __m128i vk6xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)))));
234 const __m128i vi6xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 20)));
235 const __m128i vk6xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 164 * sizeof(int8_t)))));
236 i6 += 24;
237
238 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
239 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
240 vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
241 vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
242 vaccGHIJ = _mm_macc_epi32(vi6xGHIJ, vk6xGHIJ, vaccGHIJ);
243 vaccKLMN = _mm_macc_epi32(vi6xKLMN, vk6xKLMN, vaccKLMN);
244
245 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
246 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)))));
247 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
248 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 172 * sizeof(int8_t)))));
249 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
250 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)))));
251 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
252 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 180 * sizeof(int8_t)))));
253 const __m128i vi7xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 16)));
254 const __m128i vk7xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)))));
255 const __m128i vi7xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 20)));
256 const __m128i vk7xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 188 * sizeof(int8_t)))));
257 i7 += 24;
258
259 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
260 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
261 vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
262 vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
263 vaccGHIJ = _mm_macc_epi32(vi7xGHIJ, vk7xGHIJ, vaccGHIJ);
264 vaccKLMN = _mm_macc_epi32(vi7xKLMN, vk7xKLMN, vaccKLMN);
265
266 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
267 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)))));
268 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
269 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 196 * sizeof(int8_t)))));
270 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
271 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)))));
272 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
273 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 204 * sizeof(int8_t)))));
274 const __m128i vi8xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 16)));
275 const __m128i vk8xGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)))));
276 const __m128i vi8xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 20)));
277 const __m128i vk8xKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 24 * sizeof(int32_t) + 212 * sizeof(int8_t)))));
278 i8 += 24;
279
280 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
281 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
282 vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
283 vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
284 vaccGHIJ = _mm_macc_epi32(vi8xGHIJ, vk8xGHIJ, vaccGHIJ);
285 vaccKLMN = _mm_macc_epi32(vi8xKLMN, vk8xKLMN, vaccKLMN);
286
287 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t));
288
289 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
290 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
291 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
292 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
293 __m128 vscaledGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
294 __m128 vscaledKLMN = _mm_cvtepi32_ps(vaccKLMN);
295
296 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
297 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
298 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
299 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
300 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
301 vscaledGHIJ = _mm_mul_ps(vscaledGHIJ, vscale);
302 vscaledKLMN = _mm_mul_ps(vscaledKLMN, vscale);
303
304 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
305 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
306 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
307 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
308 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
309 vscaledGHIJ = _mm_min_ps(vscaledGHIJ, voutput_max_less_zero_point);
310 vscaledKLMN = _mm_min_ps(vscaledKLMN, voutput_max_less_zero_point);
311
312 vacc0123 = _mm_cvtps_epi32(vscaled0123);
313 vacc4567 = _mm_cvtps_epi32(vscaled4567);
314 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
315 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
316 vaccGHIJ = _mm_cvtps_epi32(vscaledGHIJ);
317 vaccKLMN = _mm_cvtps_epi32(vscaledKLMN);
318
319 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
320 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
321 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
322 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
323
324 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
325 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
326 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
327 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
328 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
329
330 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
331 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
332 output += 24;
333 }
334 if XNN_UNLIKELY(c != 0) {
335 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
336 do {
337 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
338
339 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
340 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) k)));
341 i0 += 4;
342
343 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
344 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
345 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 24))));
346 i1 += 4;
347
348 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
349 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
350 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48))));
351 i2 += 4;
352
353 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
354 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
355 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 72))));
356 i3 += 4;
357
358 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
359 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
360 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96))));
361 i4 += 4;
362
363 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
364 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
365 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 120))));
366 i5 += 4;
367
368 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
369 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
370 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 144))));
371 i6 += 4;
372
373 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
374 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
375 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 168))));
376 i7 += 4;
377
378 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
379 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
380 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 192))));
381 i8 += 4;
382
383 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
384
385 k += 4;
386
387 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
388 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
389 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
390 vacc0123 = _mm_cvtps_epi32(vscaled0123);
391
392 w = (const void*) ((const int32_t*) w + 4);
393
394 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
395 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
396
397 vout0123 = _mm_packs_epi16(vout0123, vout0123);
398 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
399
400 if XNN_LIKELY(c >= 4) {
401 _mm_storeu_si32(output, vout0123);
402 output += 4;
403 c -= 4;
404 } else {
405 if (c & 2) {
406 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
407 vout0123 = _mm_srli_epi32(vout0123, 16);
408 output += 2;
409 }
410 if (c & 1) {
411 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
412 output += 1;
413 }
414 c = 0;
415 }
416 } while (c != 0);
417 }
418
419 output = (int8_t*) ((uintptr_t) output + output_increment);
420 } while (--output_width != 0);
421 }
422