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_qu8_dwconv_minmax_fp32_ukernel_up16x9__xop_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)])24 void xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul32(
25 size_t channels,
26 size_t output_width,
27 const uint8_t** input,
28 const void* weights,
29 uint8_t* output,
30 size_t input_stride,
31 size_t output_increment,
32 size_t input_offset,
33 const uint8_t* zero,
34 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(channels != 0);
37 assert(output_width != 0);
38
39 const __m128i vk_zero_point = _mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i*) params->fp32_sse2.kernel_zero_point));
40 do {
41 const uint8_t* i0 = input[0];
42 assert(i0 != NULL);
43 if XNN_UNPREDICTABLE(i0 != zero) {
44 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
45 }
46 const uint8_t* i1 = input[1];
47 assert(i1 != NULL);
48 if XNN_UNPREDICTABLE(i1 != zero) {
49 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
50 }
51 const uint8_t* i2 = input[2];
52 assert(i2 != NULL);
53 if XNN_UNPREDICTABLE(i2 != zero) {
54 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
55 }
56 const uint8_t* i3 = input[3];
57 assert(i3 != NULL);
58 if XNN_UNPREDICTABLE(i3 != zero) {
59 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
60 }
61 const uint8_t* i4 = input[4];
62 assert(i4 != NULL);
63 if XNN_UNPREDICTABLE(i4 != zero) {
64 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
65 }
66 const uint8_t* i5 = input[5];
67 assert(i5 != NULL);
68 if XNN_UNPREDICTABLE(i5 != zero) {
69 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
70 }
71 const uint8_t* i6 = input[6];
72 assert(i6 != NULL);
73 if XNN_UNPREDICTABLE(i6 != zero) {
74 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
75 }
76 const uint8_t* i7 = input[7];
77 assert(i7 != NULL);
78 if XNN_UNPREDICTABLE(i7 != zero) {
79 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
80 }
81 const uint8_t* i8 = input[8];
82 assert(i8 != NULL);
83 if XNN_UNPREDICTABLE(i8 != zero) {
84 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
85 }
86 input = (const uint8_t**) ((uintptr_t) input + input_stride);
87
88 size_t c = channels;
89 const void* w = weights;
90 for (; c >= 16; c -= 16) {
91 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
92 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
93 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
94 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
95
96
97 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
98 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);
99 const __m128i vi0x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
100 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);
101 const __m128i vi0x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
102 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);
103 const __m128i vi0xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
104 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);
105 i0 += 16;
106
107 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
108 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
109 vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
110 vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
111
112 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
113 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);
114 const __m128i vi1x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
115 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);
116 const __m128i vi1x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
117 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);
118 const __m128i vi1xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
119 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);
120 i1 += 16;
121
122 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
123 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
124 vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
125 vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
126
127 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
128 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);
129 const __m128i vi2x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
130 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);
131 const __m128i vi2x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
132 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);
133 const __m128i vi2xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
134 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);
135 i2 += 16;
136
137 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
138 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
139 vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
140 vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
141
142 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
143 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);
144 const __m128i vi3x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
145 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);
146 const __m128i vi3x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
147 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);
148 const __m128i vi3xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
149 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);
150 i3 += 16;
151
152 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
153 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
154 vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
155 vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
156
157 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
158 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);
159 const __m128i vi4x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
160 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);
161 const __m128i vi4x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
162 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);
163 const __m128i vi4xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
164 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);
165 i4 += 16;
166
167 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
168 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
169 vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
170 vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
171
172 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
173 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);
174 const __m128i vi5x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
175 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);
176 const __m128i vi5x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
177 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);
178 const __m128i vi5xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
179 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);
180 i5 += 16;
181
182 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
183 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
184 vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
185 vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
186
187 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
188 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);
189 const __m128i vi6x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
190 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);
191 const __m128i vi6x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
192 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);
193 const __m128i vi6xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
194 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);
195 i6 += 16;
196
197 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
198 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
199 vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
200 vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
201
202 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
203 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);
204 const __m128i vi7x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
205 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);
206 const __m128i vi7x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
207 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);
208 const __m128i vi7xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
209 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);
210 i7 += 16;
211
212 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
213 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
214 vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
215 vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
216
217 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
218 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);
219 const __m128i vi8x4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
220 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);
221 const __m128i vi8x89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
222 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);
223 const __m128i vi8xCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
224 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);
225 i8 += 16;
226
227 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
228 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
229 vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
230 vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
231
232 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t));
233
234 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
235 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
236 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
237 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
238
239 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
240 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
241 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
242 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale);
243 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscale);
244
245 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
246 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
247 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
248 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
249 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
250
251 vacc0123 = _mm_cvtps_epi32(vscaled0123);
252 vacc4567 = _mm_cvtps_epi32(vscaled4567);
253 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
254 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
255
256 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
257 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
258 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
259
260 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
261 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
262 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
263
264 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
265 output += 16;
266 }
267 if XNN_UNLIKELY(c != 0) {
268 const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 16);
269 do {
270 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
271
272 const __m128i vi0x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
273 const __m128i vk0x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) k))), vk_zero_point);
274 i0 += 4;
275
276 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
277 const __m128i vi1x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
278 const __m128i vk1x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16)))), vk_zero_point);
279 i1 += 4;
280
281 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
282 const __m128i vi2x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
283 const __m128i vk2x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32)))), vk_zero_point);
284 i2 += 4;
285
286 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
287 const __m128i vi3x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
288 const __m128i vk3x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48)))), vk_zero_point);
289 i3 += 4;
290
291 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
292 const __m128i vi4x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
293 const __m128i vk4x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64)))), vk_zero_point);
294 i4 += 4;
295
296 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
297 const __m128i vi5x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
298 const __m128i vk5x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 80)))), vk_zero_point);
299 i5 += 4;
300
301 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
302 const __m128i vi6x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
303 const __m128i vk6x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96)))), vk_zero_point);
304 i6 += 4;
305
306 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
307 const __m128i vi7x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
308 const __m128i vk7x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 112)))), vk_zero_point);
309 i7 += 4;
310
311 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
312 const __m128i vi8x0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
313 const __m128i vk8x0123 = _mm_sub_epi32(_mm_cvtepu8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 128)))), vk_zero_point);
314 i8 += 4;
315
316 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
317
318 k += 4;
319
320 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
321 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.scale));
322 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse2.output_max_less_zero_point));
323 vacc0123 = _mm_cvtps_epi32(vscaled0123);
324
325 w = (const void*) ((const int32_t*) w + 4);
326
327 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
328 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
329
330 vout0123 = _mm_packus_epi16(vout0123, vout0123);
331 vout0123 = _mm_max_epu8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
332
333 if XNN_LIKELY(c >= 4) {
334 _mm_storeu_si32(output, vout0123);
335 output += 4;
336 c -= 4;
337 } else {
338 if (c & 2) {
339 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
340 vout0123 = _mm_srli_epi32(vout0123, 16);
341 output += 2;
342 }
343 if (c & 1) {
344 *output = (uint8_t) _mm_extract_epi8(vout0123, 0);
345 output += 1;
346 }
347 c = 0;
348 }
349 } while (c != 0);
350 }
351
352 output = (uint8_t*) ((uintptr_t) output + output_increment);
353 } while (--output_width != 0);
354 }
355