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_x16(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_x16(
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 >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
36 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
37 x += 16;
38
39 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
40 __m128i vextx0 = _mm_unpacklo_epi8(vx, vm);
41 __m128i vextx1 = _mm_unpackhi_epi8(vx, vm);
42
43 __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
44 __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
45 vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
46 vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);
47
48 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
49 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
50
51 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
52 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
53
54 __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
55 __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);
56
57 vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
58 vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
59 __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
60 __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);
61
62 vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
63 vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
64 vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
65 vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);
66
67 __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
68 __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);
69
70 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
71 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
72
73 const __m128i vy = _mm_packs_epi16(vacc0, vacc1);
74 _mm_storeu_si128((__m128i*) y, vy);
75 y += 16;
76 }
77 if XNN_UNLIKELY(n != 0) {
78 assert(n >= 1 * sizeof(int8_t));
79 assert(n <= 15 * sizeof(int8_t));
80
81 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
82
83 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
84 __m128i vextx0 = _mm_unpacklo_epi8(vx, vm);
85 __m128i vextx1 = _mm_unpackhi_epi8(vx, vm);
86
87 __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point);
88 __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point);
89 vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0);
90 vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1);
91
92 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
93 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
94
95 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
96 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
97
98 __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0);
99 __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1);
100
101 vprodlo0 = _mm_srli_epi16(vprodlo0, 7);
102 vprodlo1 = _mm_srli_epi16(vprodlo1, 7);
103 __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0);
104 __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1);
105
106 vprodhi0 = _mm_slli_epi16(vprodhi0, 8);
107 vprodhi1 = _mm_slli_epi16(vprodhi1, 8);
108 vprodlo0 = _mm_avg_epu16(vprodlo0, vzero);
109 vprodlo1 = _mm_avg_epu16(vprodlo1, vzero);
110
111 __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0);
112 __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1);
113
114 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
115 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
116
117 __m128i vy = _mm_packs_epi16(vacc0, vacc1);
118 if (n & (8 * sizeof(int8_t))) {
119 _mm_storel_epi64((__m128i*) y, vy);
120 vy = _mm_unpackhi_epi64(vy, vy);
121 y += 8;
122 }
123 if (n & (4 * sizeof(int8_t))) {
124 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
125 vy = _mm_srli_epi64(vy, 32);
126 y += 4;
127 }
128 uint32_t vy0 = (uint32_t) _mm_cvtsi128_si32(vy);
129 if (n & (2 * sizeof(int8_t))) {
130 unaligned_store_u16(y, (uint16_t) vy0);
131 vy0 >>= 16;
132 y += 2;
133 }
134 if (n & (1 * sizeof(int8_t))) {
135 *y = (int8_t) vy0;
136 }
137 }
138 }
139