1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vlrelu/sse2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vlrelu.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_vlrelu_ukernel__sse2_x32(size_t n,const uint8_t * x,uint8_t * y,const union xnn_qu8_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_vlrelu_ukernel__sse2_x32(
20 size_t n,
21 const uint8_t* x,
22 uint8_t* y,
23 const union xnn_qu8_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(uint8_t) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->sse2.input_zero_point);
31 const __m128i vmultiplier_diff = _mm_load_si128((const __m128i*) params->sse2.multiplier_diff);
32 const __m128i vmultiplier_base = _mm_load_si128((const __m128i*) params->sse2.multiplier_base);
33 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
34 const __m128i vzero = _mm_setzero_si128();
35 for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
36 const __m128i vx0 = _mm_loadu_si128((const __m128i*) x);
37 const __m128i vx1 = _mm_loadu_si128((const __m128i*) (x + 16));
38 x += 32;
39
40 __m128i vextx0 = _mm_unpacklo_epi8(vx0, vzero);
41 __m128i vextx1 = _mm_unpackhi_epi8(vx0, vzero);
42 __m128i vextx2 = _mm_unpacklo_epi8(vx1, vzero);
43 __m128i vextx3 = _mm_unpackhi_epi8(vx1, vzero);
44
45 __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
46 vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
47 __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
48 vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);
49 __m128i vmultiplier2 = _mm_cmpgt_epi16(vextx2, vinput_zero_point);
50 vextx2 = _mm_sub_epi16(vinput_zero_point, vextx2);
51 __m128i vmultiplier3 = _mm_cmpgt_epi16(vextx3, vinput_zero_point);
52 vextx3 = _mm_sub_epi16(vinput_zero_point, vextx3);
53
54 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
55 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
56 vmultiplier2 = _mm_and_si128(vmultiplier2, vmultiplier_diff);
57 vmultiplier3 = _mm_and_si128(vmultiplier3, vmultiplier_diff);
58
59 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
60 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
61 vmultiplier2 = _mm_xor_si128(vmultiplier2, vmultiplier_base);
62 vmultiplier3 = _mm_xor_si128(vmultiplier3, vmultiplier_base);
63
64 __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
65 __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);
66 __m128i vprodlo2 = _mm_mullo_epi16(vextx2, vmultiplier2);
67 __m128i vprodlo3 = _mm_mullo_epi16(vextx3, vmultiplier3);
68
69 vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
70 __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
71 vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
72 __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);
73 vprodlo2 = _mm_srli_epi16(vprodlo2, 7);
74 __m128i vprodhi2 = _mm_mulhi_epi16(vextx2, vmultiplier2);
75 vprodlo3 = _mm_srli_epi16(vprodlo3, 7);
76 __m128i vprodhi3 = _mm_mulhi_epi16(vextx3, vmultiplier3);
77
78 vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
79 vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
80 vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
81 vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);
82 vprodhi2 = _mm_slli_epi16(vprodhi2, 8);
83 vprodlo2 = _mm_avg_epu16(vprodlo2, vzero);
84 vprodhi3 = _mm_slli_epi16(vprodhi3, 8);
85 vprodlo3 = _mm_avg_epu16(vprodlo3, vzero);
86
87 __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
88 __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);
89 __m128i vacc2 = _mm_add_epi16(vprodlo2, vprodhi2);
90 __m128i vacc3 = _mm_add_epi16(vprodlo3, vprodhi3);
91
92 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
93 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
94 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
95 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
96
97 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
98 const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
99
100 _mm_storeu_si128((__m128i*) y, vy0);
101 _mm_storeu_si128((__m128i*) (y + 16), vy1);
102 y += 32;
103 }
104 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
105 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
106 x += 16;
107
108 __m128i vextx0 = _mm_unpacklo_epi8(vx, vzero);
109 __m128i vextx1 = _mm_unpackhi_epi8(vx, vzero);
110
111 __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
112 __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
113 vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
114 vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);
115
116 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
117 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
118
119 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
120 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
121
122 __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
123 __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);
124
125 vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
126 vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
127 __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
128 __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);
129
130 vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
131 vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
132 vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
133 vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);
134
135 __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
136 __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);
137
138 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
139 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
140
141 const __m128i vy = _mm_packus_epi16(vacc0, vacc1);
142 _mm_storeu_si128((__m128i*) y, vy);
143 y += 16;
144 }
145 if XNN_UNLIKELY(n != 0) {
146 assert(n >= 1 * sizeof(uint8_t));
147 assert(n <= 15 * sizeof(uint8_t));
148
149 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
150
151 __m128i vextx0 = _mm_unpacklo_epi8(vx, vzero);
152 __m128i vextx1 = _mm_unpackhi_epi8(vx, vzero);
153
154 __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
155 __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
156 vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
157 vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);
158
159 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
160 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
161
162 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
163 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
164
165 __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
166 __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);
167
168 vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
169 vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
170 __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
171 __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);
172
173 vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
174 vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
175 vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
176 vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);
177
178 __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
179 __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);
180
181 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
182 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
183
184 __m128i vy = _mm_packus_epi16(vacc0, vacc1);
185 if (n & (8 * sizeof(uint8_t))) {
186 _mm_storel_epi64((__m128i*) y, vy);
187 vy = _mm_unpackhi_epi64(vy, vy);
188 y += 8;
189 }
190 if (n & (4 * sizeof(uint8_t))) {
191 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
192 vy = _mm_srli_epi64(vy, 32);
193 y += 4;
194 }
195 uint32_t vy0 = (uint32_t) _mm_cvtsi128_si32(vy);
196 if (n & (2 * sizeof(uint8_t))) {
197 unaligned_store_u16(y, (uint16_t) vy0);
198 vy0 >>= 16;
199 y += 2;
200 }
201 if (n & (1 * sizeof(uint8_t))) {
202 *y = (uint8_t) vy0;
203 }
204 }
205 }
206