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_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32(
20 size_t channels,
21 size_t output_width,
22 const uint8_t** input,
23 const void* weights,
24 uint8_t* output,
25 size_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const uint8_t* zero,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(channels != 0);
32 assert(output_width != 0);
33
34 const __m128i vk_zero_point = _mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) params->fp32_sse2.kernel_zero_point));
35 do {
36 const uint8_t* i0 = input[0];
37 assert(i0 != NULL);
38 if XNN_UNPREDICTABLE(i0 != zero) {
39 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
40 }
41 const uint8_t* i1 = input[1];
42 assert(i1 != NULL);
43 if XNN_UNPREDICTABLE(i1 != zero) {
44 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
45 }
46 const uint8_t* i2 = input[2];
47 assert(i2 != NULL);
48 if XNN_UNPREDICTABLE(i2 != zero) {
49 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
50 }
51 const uint8_t* i3 = input[3];
52 assert(i3 != NULL);
53 if XNN_UNPREDICTABLE(i3 != zero) {
54 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
55 }
56 const uint8_t* i4 = input[4];
57 assert(i4 != NULL);
58 if XNN_UNPREDICTABLE(i4 != zero) {
59 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
60 }
61 const uint8_t* i5 = input[5];
62 assert(i5 != NULL);
63 if XNN_UNPREDICTABLE(i5 != zero) {
64 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
65 }
66 const uint8_t* i6 = input[6];
67 assert(i6 != NULL);
68 if XNN_UNPREDICTABLE(i6 != zero) {
69 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
70 }
71 const uint8_t* i7 = input[7];
72 assert(i7 != NULL);
73 if XNN_UNPREDICTABLE(i7 != zero) {
74 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
75 }
76 const uint8_t* i8 = input[8];
77 assert(i8 != NULL);
78 if XNN_UNPREDICTABLE(i8 != zero) {
79 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
80 }
81 input = (const uint8_t**) ((uintptr_t) input + input_stride);
82
83 size_t c = channels;
84 const void* w = weights;
85 for (; c >= 16; c -= 16) {
86 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
87 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
88 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
89 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
90
91
92 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
93 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t))))), vk_zero_point);
94 const __m128i vi0x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
95 const __m128i vk0x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(uint8_t))))), vk_zero_point);
96 const __m128i vi0x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
97 const __m128i vk0x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t))))), vk_zero_point);
98 const __m128i vi0xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
99 const __m128i vk0xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(uint8_t))))), vk_zero_point);
100 i0 += 16;
101
102 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
103 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
104 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
105 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
106
107 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
108 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t))))), vk_zero_point);
109 const __m128i vi1x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
110 const __m128i vk1x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(uint8_t))))), vk_zero_point);
111 const __m128i vi1x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
112 const __m128i vk1x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t))))), vk_zero_point);
113 const __m128i vi1xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
114 const __m128i vk1xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(uint8_t))))), vk_zero_point);
115 i1 += 16;
116
117 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
118 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
119 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
120 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
121
122 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
123 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t))))), vk_zero_point);
124 const __m128i vi2x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
125 const __m128i vk2x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(uint8_t))))), vk_zero_point);
126 const __m128i vi2x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
127 const __m128i vk2x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t))))), vk_zero_point);
128 const __m128i vi2xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
129 const __m128i vk2xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(uint8_t))))), vk_zero_point);
130 i2 += 16;
131
132 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
133 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
134 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
135 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
136
137 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
138 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t))))), vk_zero_point);
139 const __m128i vi3x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
140 const __m128i vk3x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(uint8_t))))), vk_zero_point);
141 const __m128i vi3x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
142 const __m128i vk3x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t))))), vk_zero_point);
143 const __m128i vi3xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
144 const __m128i vk3xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(uint8_t))))), vk_zero_point);
145 i3 += 16;
146
147 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
148 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
149 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
150 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
151
152 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
153 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t))))), vk_zero_point);
154 const __m128i vi4x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
155 const __m128i vk4x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(uint8_t))))), vk_zero_point);
156 const __m128i vi4x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
157 const __m128i vk4x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t))))), vk_zero_point);
158 const __m128i vi4xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
159 const __m128i vk4xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(uint8_t))))), vk_zero_point);
160 i4 += 16;
161
162 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
163 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
164 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
165 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
166
167 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
168 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t))))), vk_zero_point);
169 const __m128i vi5x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
170 const __m128i vk5x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(uint8_t))))), vk_zero_point);
171 const __m128i vi5x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
172 const __m128i vk5x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t))))), vk_zero_point);
173 const __m128i vi5xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
174 const __m128i vk5xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(uint8_t))))), vk_zero_point);
175 i5 += 16;
176
177 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
178 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
179 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
180 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
181
182 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
183 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t))))), vk_zero_point);
184 const __m128i vi6x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
185 const __m128i vk6x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(uint8_t))))), vk_zero_point);
186 const __m128i vi6x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
187 const __m128i vk6x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t))))), vk_zero_point);
188 const __m128i vi6xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
189 const __m128i vk6xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(uint8_t))))), vk_zero_point);
190 i6 += 16;
191
192 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
193 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
194 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi6x89AB, vk6x89AB));
195 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi6xCDEF, vk6xCDEF));
196
197 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
198 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t))))), vk_zero_point);
199 const __m128i vi7x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
200 const __m128i vk7x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(uint8_t))))), vk_zero_point);
201 const __m128i vi7x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
202 const __m128i vk7x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t))))), vk_zero_point);
203 const __m128i vi7xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
204 const __m128i vk7xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(uint8_t))))), vk_zero_point);
205 i7 += 16;
206
207 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
208 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
209 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi7x89AB, vk7x89AB));
210 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi7xCDEF, vk7xCDEF));
211
212 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
213 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t))))), vk_zero_point);
214 const __m128i vi8x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
215 const __m128i vk8x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(uint8_t))))), vk_zero_point);
216 const __m128i vi8x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
217 const __m128i vk8x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t))))), vk_zero_point);
218 const __m128i vi8xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
219 const __m128i vk8xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(uint8_t))))), vk_zero_point);
220 i8 += 16;
221
222 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
223 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
224 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi8x89AB, vk8x89AB));
225 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi8xCDEF, vk8xCDEF));
226
227 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t));
228
229 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
230 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
231 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
232 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
233
234 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
235 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
236 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
237 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
238 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
239
240 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
241 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
242 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
243 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
244 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
245
246 vacc0123 = _mm_cvtps_epi32(vscaled0123);
247 vacc4567 = _mm_cvtps_epi32(vscaled4567);
248 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
249 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
250
251 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
252 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
253 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
254
255 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
256 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
257 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
258
259 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
260 output += 16;
261 }
262 if XNN_UNLIKELY(c != 0) {
263 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
264 do {
265 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
266
267 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
268 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) k))), vk_zero_point);
269 i0 += 4;
270
271 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
272 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
273 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16)))), vk_zero_point);
274 i1 += 4;
275
276 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
277 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
278 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32)))), vk_zero_point);
279 i2 += 4;
280
281 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
282 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
283 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48)))), vk_zero_point);
284 i3 += 4;
285
286 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
287 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
288 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64)))), vk_zero_point);
289 i4 += 4;
290
291 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
292 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
293 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 80)))), vk_zero_point);
294 i5 += 4;
295
296 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
297 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
298 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96)))), vk_zero_point);
299 i6 += 4;
300
301 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
302 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
303 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 112)))), vk_zero_point);
304 i7 += 4;
305
306 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
307 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
308 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 128)))), vk_zero_point);
309 i8 += 4;
310
311 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
312
313 k += 4;
314
315 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
316 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.scale));
317 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.output_max_less_zero_point));
318 vacc0123 = _mm_cvtps_epi32(vscaled0123);
319
320 w = (const void*) ((const int32_t*) w + 4);
321
322 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
323 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
324
325 vout0123 = _mm_packus_epi16(vout0123, vout0123);
326 vout0123 = _mm_max_epu8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
327
328 if XNN_LIKELY(c >= 4) {
329 _mm_storeu_si32(output, vout0123);
330 output += 4;
331 c -= 4;
332 } else {
333 if (c & 2) {
334 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
335 vout0123 = _mm_srli_epi32(vout0123, 16);
336 output += 2;
337 }
338 if (c & 1) {
339 *output = (uint8_t) _mm_extract_epi8(vout0123, 0);
340 output += 1;
341 }
342 c = 0;
343 }
344 } while (c != 0);
345 }
346
347 output = (uint8_t*) ((uintptr_t) output + output_increment);
348 } while (--output_width != 0);
349 }
350