1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vlrelu/ssse3.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 <tmmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vlrelu.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_vlrelu_ukernel__ssse3_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__ssse3_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 vacc0 = _mm_unpacklo_epi8(vx0, vzero);
41 __m128i vacc1 = _mm_unpackhi_epi8(vx0, vzero);
42 __m128i vacc2 = _mm_unpacklo_epi8(vx1, vzero);
43 __m128i vacc3 = _mm_unpackhi_epi8(vx1, vzero);
44
45 __m128i vmultiplier0 = _mm_cmpgt_epi16(vacc0, vinput_zero_point);
46 vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
47 __m128i vmultiplier1 = _mm_cmpgt_epi16(vacc1, vinput_zero_point);
48 vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
49 __m128i vmultiplier2 = _mm_cmpgt_epi16(vacc2, vinput_zero_point);
50 vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
51 __m128i vmultiplier3 = _mm_cmpgt_epi16(vacc3, vinput_zero_point);
52 vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
53
54 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
55 vacc0 = _mm_slli_epi16(vacc0, 7);
56 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
57 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
58 vacc1 = _mm_slli_epi16(vacc1, 7);
59 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
60 vmultiplier2 = _mm_and_si128(vmultiplier2, vmultiplier_diff);
61 vacc2 = _mm_slli_epi16(vacc2, 7);
62 vmultiplier2 = _mm_xor_si128(vmultiplier2, vmultiplier_base);
63 vmultiplier3 = _mm_and_si128(vmultiplier3, vmultiplier_diff);
64 vacc3 = _mm_slli_epi16(vacc3, 7);
65 vmultiplier3 = _mm_xor_si128(vmultiplier3, vmultiplier_base);
66
67 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier0);
68 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier1);
69 vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier2);
70 vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier3);
71
72 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
73 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
74 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
75 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
76
77 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
78 const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
79
80 _mm_storeu_si128((__m128i*) y, vy0);
81 _mm_storeu_si128((__m128i*) (y + 16), vy1);
82 y += 32;
83 }
84 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
85 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
86 x += 16;
87
88 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
89 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
90 __m128i vmultiplier_lo = _mm_cmpgt_epi16(vacc_lo, vinput_zero_point);
91 __m128i vmultiplier_hi = _mm_cmpgt_epi16(vacc_hi, vinput_zero_point);
92 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
93 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
94 vmultiplier_lo = _mm_and_si128(vmultiplier_lo, vmultiplier_diff);
95 vmultiplier_hi = _mm_and_si128(vmultiplier_hi, vmultiplier_diff);
96 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
97 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
98 vmultiplier_lo = _mm_xor_si128(vmultiplier_lo, vmultiplier_base);
99 vmultiplier_hi = _mm_xor_si128(vmultiplier_hi, vmultiplier_base);
100 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier_lo);
101 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier_hi);
102 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
103 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
104
105 const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
106 _mm_storeu_si128((__m128i*) y, vy);
107 y += 16;
108 }
109 if XNN_UNLIKELY(n != 0) {
110 assert(n >= 1 * sizeof(uint8_t));
111 assert(n <= 15 * sizeof(uint8_t));
112
113 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
114
115 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
116 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
117 __m128i vmultiplier_lo = _mm_cmpgt_epi16(vacc_lo, vinput_zero_point);
118 __m128i vmultiplier_hi = _mm_cmpgt_epi16(vacc_hi, vinput_zero_point);
119 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
120 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
121 vmultiplier_lo = _mm_and_si128(vmultiplier_lo, vmultiplier_diff);
122 vmultiplier_hi = _mm_and_si128(vmultiplier_hi, vmultiplier_diff);
123 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
124 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
125 vmultiplier_lo = _mm_xor_si128(vmultiplier_lo, vmultiplier_base);
126 vmultiplier_hi = _mm_xor_si128(vmultiplier_hi, vmultiplier_base);
127 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier_lo);
128 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier_hi);
129 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
130 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
131
132 __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
133 if (n & (8 * sizeof(uint8_t))) {
134 _mm_storel_epi64((__m128i*) y, vy);
135 vy = _mm_unpackhi_epi64(vy, vy);
136 y += 8;
137 }
138 if (n & (4 * sizeof(uint8_t))) {
139 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
140 vy = _mm_srli_epi64(vy, 32);
141 y += 4;
142 }
143 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
144 if (n & (2 * sizeof(uint8_t))) {
145 unaligned_store_u16(y, (uint16_t) vy_lo);
146 vy_lo >>= 16;
147 y += 2;
148 }
149 if (n & (1 * sizeof(uint8_t))) {
150 *y = (uint8_t) vy_lo;
151 }
152 }
153 }
154