1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_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)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32(
19 size_t channels,
20 size_t output_width,
21 const uint8_t** input,
22 const void* weights,
23 uint8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const uint8_t* zero,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m128i vk_zero_point = _mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) params->fp32_sse2.kernel_zero_point));
34 do {
35 const uint8_t* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
39 }
40 const uint8_t* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
44 }
45 const uint8_t* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
49 }
50 const uint8_t* i3 = input[3];
51 assert(i3 != NULL);
52 if XNN_UNPREDICTABLE(i3 != zero) {
53 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
54 }
55 const uint8_t* i4 = input[4];
56 assert(i4 != NULL);
57 if XNN_UNPREDICTABLE(i4 != zero) {
58 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
59 }
60 const uint8_t* i5 = input[5];
61 assert(i5 != NULL);
62 if XNN_UNPREDICTABLE(i5 != zero) {
63 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
64 }
65 const uint8_t* i6 = input[6];
66 assert(i6 != NULL);
67 if XNN_UNPREDICTABLE(i6 != zero) {
68 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
69 }
70 const uint8_t* i7 = input[7];
71 assert(i7 != NULL);
72 if XNN_UNPREDICTABLE(i7 != zero) {
73 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
74 }
75 const uint8_t* i8 = input[8];
76 assert(i8 != NULL);
77 if XNN_UNPREDICTABLE(i8 != zero) {
78 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
79 }
80 input = (const uint8_t**) ((uintptr_t) input + input_stride);
81
82 size_t c = channels;
83 const void* w = weights;
84 for (; c >= 16; c -= 16) {
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
90
91 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i0));
92 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)))), vk_zero_point);
93 const __m128i vi0x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i0 + 4));
94 const __m128i vk0x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(uint8_t)))), vk_zero_point);
95 const __m128i vi0x89AB = _mm_cvtepu8_epi32(_mm_loadu_si32(i0 + 8));
96 const __m128i vk0x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t)))), vk_zero_point);
97 const __m128i vi0xCDEF = _mm_cvtepu8_epi32(_mm_loadu_si32(i0 + 12));
98 const __m128i vk0xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(uint8_t)))), vk_zero_point);
99 i0 += 16;
100
101 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
102 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi0x4567, vk0x4567));
103 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi0x89AB, vk0x89AB));
104 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi0xCDEF, vk0xCDEF));
105
106 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i1));
107 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)))), vk_zero_point);
108 const __m128i vi1x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i1 + 4));
109 const __m128i vk1x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(uint8_t)))), vk_zero_point);
110 const __m128i vi1x89AB = _mm_cvtepu8_epi32(_mm_loadu_si32(i1 + 8));
111 const __m128i vk1x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t)))), vk_zero_point);
112 const __m128i vi1xCDEF = _mm_cvtepu8_epi32(_mm_loadu_si32(i1 + 12));
113 const __m128i vk1xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(uint8_t)))), vk_zero_point);
114 i1 += 16;
115
116 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
117 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi1x4567, vk1x4567));
118 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi1x89AB, vk1x89AB));
119 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi1xCDEF, vk1xCDEF));
120
121 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i2));
122 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)))), vk_zero_point);
123 const __m128i vi2x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i2 + 4));
124 const __m128i vk2x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(uint8_t)))), vk_zero_point);
125 const __m128i vi2x89AB = _mm_cvtepu8_epi32(_mm_loadu_si32(i2 + 8));
126 const __m128i vk2x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t)))), vk_zero_point);
127 const __m128i vi2xCDEF = _mm_cvtepu8_epi32(_mm_loadu_si32(i2 + 12));
128 const __m128i vk2xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(uint8_t)))), vk_zero_point);
129 i2 += 16;
130
131 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
132 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi2x4567, vk2x4567));
133 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi2x89AB, vk2x89AB));
134 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi2xCDEF, vk2xCDEF));
135
136 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i3));
137 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)))), vk_zero_point);
138 const __m128i vi3x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i3 + 4));
139 const __m128i vk3x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(uint8_t)))), vk_zero_point);
140 const __m128i vi3x89AB = _mm_cvtepu8_epi32(_mm_loadu_si32(i3 + 8));
141 const __m128i vk3x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t)))), vk_zero_point);
142 const __m128i vi3xCDEF = _mm_cvtepu8_epi32(_mm_loadu_si32(i3 + 12));
143 const __m128i vk3xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(uint8_t)))), vk_zero_point);
144 i3 += 16;
145
146 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
147 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi3x4567, vk3x4567));
148 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi3x89AB, vk3x89AB));
149 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi3xCDEF, vk3xCDEF));
150
151 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i4));
152 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)))), vk_zero_point);
153 const __m128i vi4x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i4 + 4));
154 const __m128i vk4x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(uint8_t)))), vk_zero_point);
155 const __m128i vi4x89AB = _mm_cvtepu8_epi32(_mm_loadu_si32(i4 + 8));
156 const __m128i vk4x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t)))), vk_zero_point);
157 const __m128i vi4xCDEF = _mm_cvtepu8_epi32(_mm_loadu_si32(i4 + 12));
158 const __m128i vk4xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(uint8_t)))), vk_zero_point);
159 i4 += 16;
160
161 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
162 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi4x4567, vk4x4567));
163 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi4x89AB, vk4x89AB));
164 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi4xCDEF, vk4xCDEF));
165
166 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i5));
167 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)))), vk_zero_point);
168 const __m128i vi5x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i5 + 4));
169 const __m128i vk5x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(uint8_t)))), vk_zero_point);
170 const __m128i vi5x89AB = _mm_cvtepu8_epi32(_mm_loadu_si32(i5 + 8));
171 const __m128i vk5x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t)))), vk_zero_point);
172 const __m128i vi5xCDEF = _mm_cvtepu8_epi32(_mm_loadu_si32(i5 + 12));
173 const __m128i vk5xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(uint8_t)))), vk_zero_point);
174 i5 += 16;
175
176 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
177 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi5x4567, vk5x4567));
178 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi5x89AB, vk5x89AB));
179 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi5xCDEF, vk5xCDEF));
180
181 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i6));
182 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)))), vk_zero_point);
183 const __m128i vi6x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i6 + 4));
184 const __m128i vk6x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(uint8_t)))), vk_zero_point);
185 const __m128i vi6x89AB = _mm_cvtepu8_epi32(_mm_loadu_si32(i6 + 8));
186 const __m128i vk6x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t)))), vk_zero_point);
187 const __m128i vi6xCDEF = _mm_cvtepu8_epi32(_mm_loadu_si32(i6 + 12));
188 const __m128i vk6xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(uint8_t)))), vk_zero_point);
189 i6 += 16;
190
191 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
192 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi6x4567, vk6x4567));
193 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi6x89AB, vk6x89AB));
194 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi6xCDEF, vk6xCDEF));
195
196 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i7));
197 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)))), vk_zero_point);
198 const __m128i vi7x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i7 + 4));
199 const __m128i vk7x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(uint8_t)))), vk_zero_point);
200 const __m128i vi7x89AB = _mm_cvtepu8_epi32(_mm_loadu_si32(i7 + 8));
201 const __m128i vk7x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t)))), vk_zero_point);
202 const __m128i vi7xCDEF = _mm_cvtepu8_epi32(_mm_loadu_si32(i7 + 12));
203 const __m128i vk7xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(uint8_t)))), vk_zero_point);
204 i7 += 16;
205
206 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
207 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi7x4567, vk7x4567));
208 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi7x89AB, vk7x89AB));
209 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi7xCDEF, vk7xCDEF));
210
211 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i8));
212 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)))), vk_zero_point);
213 const __m128i vi8x4567 = _mm_cvtepu8_epi32(_mm_loadu_si32(i8 + 4));
214 const __m128i vk8x4567 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(uint8_t)))), vk_zero_point);
215 const __m128i vi8x89AB = _mm_cvtepu8_epi32(_mm_loadu_si32(i8 + 8));
216 const __m128i vk8x89AB = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t)))), vk_zero_point);
217 const __m128i vi8xCDEF = _mm_cvtepu8_epi32(_mm_loadu_si32(i8 + 12));
218 const __m128i vk8xCDEF = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(uint8_t)))), vk_zero_point);
219 i8 += 16;
220
221 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
222 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vi8x4567, vk8x4567));
223 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vi8x89AB, vk8x89AB));
224 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vi8xCDEF, vk8xCDEF));
225
226 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t));
227
228 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
229 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
230 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
231 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
232
233 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
234 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
235 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
236 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
237 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
238
239 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
240 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
241 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
242 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
243 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
244
245 vacc0123 = _mm_cvtps_epi32(vscaled0123);
246 vacc4567 = _mm_cvtps_epi32(vscaled4567);
247 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
248 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
249
250 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
251 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
252 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
253
254 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
255 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
256 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
257
258 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
259 output += 16;
260 }
261 if XNN_UNLIKELY(c != 0) {
262 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
263 do {
264 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
265
266 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i0));
267 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32(k)), vk_zero_point);
268 i0 += 4;
269
270 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi0x0123, vk0x0123));
271 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i1));
272 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 16))), vk_zero_point);
273 i1 += 4;
274
275 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi1x0123, vk1x0123));
276 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i2));
277 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 32))), vk_zero_point);
278 i2 += 4;
279
280 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi2x0123, vk2x0123));
281 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i3));
282 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 48))), vk_zero_point);
283 i3 += 4;
284
285 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi3x0123, vk3x0123));
286 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i4));
287 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 64))), vk_zero_point);
288 i4 += 4;
289
290 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi4x0123, vk4x0123));
291 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i5));
292 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 80))), vk_zero_point);
293 i5 += 4;
294
295 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi5x0123, vk5x0123));
296 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i6));
297 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 96))), vk_zero_point);
298 i6 += 4;
299
300 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi6x0123, vk6x0123));
301 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i7));
302 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 112))), vk_zero_point);
303 i7 += 4;
304
305 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi7x0123, vk7x0123));
306 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_loadu_si32(i8));
307 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_loadu_si32((const void*) (k + 128))), vk_zero_point);
308 i8 += 4;
309
310 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vi8x0123, vk8x0123));
311
312 k += 4;
313
314 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
315 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.scale));
316 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.output_max_less_zero_point));
317 vacc0123 = _mm_cvtps_epi32(vscaled0123);
318
319 w = (const void*) ((const int32_t*) w + 4);
320
321 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
322 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
323
324 vout0123 = _mm_packus_epi16(vout0123, vout0123);
325 vout0123 = _mm_max_epu8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
326
327 if XNN_LIKELY(c >= 4) {
328 _mm_storeu_si32(output, vout0123);
329 output += 4;
330 c -= 4;
331 } else {
332 if (c & 2) {
333 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
334 vout0123 = _mm_srli_epi32(vout0123, 16);
335 output += 2;
336 }
337 if (c & 1) {
338 *output = (uint8_t) _mm_extract_epi8(vout0123, 0);
339 output += 1;
340 }
341 c = 0;
342 }
343 } while (c != 0);
344 }
345
346 output = (uint8_t*) ((uintptr_t) output + output_increment);
347 } while (--output_width != 0);
348 }
349