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_qs8_vlrelu_ukernel__sse2_x32(size_t n,const int8_t * x,int8_t * y,const union xnn_qs8_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_vlrelu_ukernel__sse2_x32(
20 size_t n,
21 const int8_t* x,
22 int8_t* y,
23 const union xnn_qs8_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(int8_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(int8_t); n -= 32 * sizeof(int8_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 const __m128i vm0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx0);
41 __m128i vextx0 = _mm_unpacklo_epi8(vx0, vm0);
42 __m128i vextx1 = _mm_unpackhi_epi8(vx0, vm0);
43 const __m128i vm1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx1);
44 __m128i vextx2 = _mm_unpacklo_epi8(vx1, vm1);
45 __m128i vextx3 = _mm_unpackhi_epi8(vx1, vm1);
46
47 __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
48 vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
49 __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
50 vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);
51 __m128i vmultiplier2 = _mm_cmpgt_epi16(vextx2, vinput_zero_point);
52 vextx2 = _mm_sub_epi16(vinput_zero_point, vextx2);
53 __m128i vmultiplier3 = _mm_cmpgt_epi16(vextx3, vinput_zero_point);
54 vextx3 = _mm_sub_epi16(vinput_zero_point, vextx3);
55
56 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
57 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
58 vmultiplier2 = _mm_and_si128(vmultiplier2, vmultiplier_diff);
59 vmultiplier3 = _mm_and_si128(vmultiplier3, vmultiplier_diff);
60
61 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
62 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
63 vmultiplier2 = _mm_xor_si128(vmultiplier2, vmultiplier_base);
64 vmultiplier3 = _mm_xor_si128(vmultiplier3, vmultiplier_base);
65
66 __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
67 __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);
68 __m128i vprodlo2 = _mm_mullo_epi16(vextx2, vmultiplier2);
69 __m128i vprodlo3 = _mm_mullo_epi16(vextx3, vmultiplier3);
70
71 vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
72 __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
73 vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
74 __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);
75 vprodlo2 = _mm_srli_epi16(vprodlo2, 7);
76 __m128i vprodhi2 = _mm_mulhi_epi16(vextx2, vmultiplier2);
77 vprodlo3 = _mm_srli_epi16(vprodlo3, 7);
78 __m128i vprodhi3 = _mm_mulhi_epi16(vextx3, vmultiplier3);
79
80 vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
81 vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
82 vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
83 vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);
84 vprodhi2 = _mm_slli_epi16(vprodhi2, 8);
85 vprodlo2 = _mm_avg_epu16(vprodlo2, vzero);
86 vprodhi3 = _mm_slli_epi16(vprodhi3, 8);
87 vprodlo3 = _mm_avg_epu16(vprodlo3, vzero);
88
89 __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
90 __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);
91 __m128i vacc2 = _mm_add_epi16(vprodlo2, vprodhi2);
92 __m128i vacc3 = _mm_add_epi16(vprodlo3, vprodhi3);
93
94 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
95 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
96 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
97 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
98
99 const __m128i vy0 = _mm_packs_epi16(vacc0, vacc1);
100 const __m128i vy1 = _mm_packs_epi16(vacc2, vacc3);
101
102 _mm_storeu_si128((__m128i*) y, vy0);
103 _mm_storeu_si128((__m128i*) (y + 16), vy1);
104 y += 32;
105 }
106 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
107 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
108 x += 16;
109
110 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
111 __m128i vextx0 = _mm_unpacklo_epi8(vx, vm);
112 __m128i vextx1 = _mm_unpackhi_epi8(vx, vm);
113
114 __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
115 __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
116 vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
117 vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);
118
119 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
120 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
121
122 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
123 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
124
125 __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
126 __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);
127
128 vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
129 vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
130 __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
131 __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);
132
133 vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
134 vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
135 vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
136 vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);
137
138 __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
139 __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);
140
141 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
142 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
143
144 const __m128i vy = _mm_packs_epi16(vacc0, vacc1);
145 _mm_storeu_si128((__m128i*) y, vy);
146 y += 16;
147 }
148 if XNN_UNLIKELY(n != 0) {
149 assert(n >= 1 * sizeof(int8_t));
150 assert(n <= 15 * sizeof(int8_t));
151
152 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
153
154 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
155 __m128i vextx0 = _mm_unpacklo_epi8(vx, vm);
156 __m128i vextx1 = _mm_unpackhi_epi8(vx, vm);
157
158 __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
159 __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
160 vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
161 vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);
162
163 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
164 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
165
166 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
167 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
168
169 __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
170 __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);
171
172 vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
173 vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
174 __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
175 __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);
176
177 vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
178 vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
179 vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
180 vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);
181
182 __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
183 __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);
184
185 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
186 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
187
188 __m128i vy = _mm_packs_epi16(vacc0, vacc1);
189 if (n & (8 * sizeof(int8_t))) {
190 _mm_storel_epi64((__m128i*) y, vy);
191 vy = _mm_unpackhi_epi64(vy, vy);
192 y += 8;
193 }
194 if (n & (4 * sizeof(int8_t))) {
195 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
196 vy = _mm_srli_epi64(vy, 32);
197 y += 4;
198 }
199 uint32_t vy0 = (uint32_t) _mm_cvtsi128_si32(vy);
200 if (n & (2 * sizeof(int8_t))) {
201 unaligned_store_u16(y, (uint16_t) vy0);
202 vy0 >>= 16;
203 y += 2;
204 }
205 if (n & (1 * sizeof(int8_t))) {
206 *y = (int8_t) vy0;
207 }
208 }
209 }
210