1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vlrelu/sse4.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 <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vlrelu.h>
17
18
xnn_qs8_vlrelu_ukernel__sse41_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__sse41_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 for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
35 __m128i vacc0 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) x));
36 __m128i vacc1 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (x + 8)));
37 __m128i vacc2 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (x + 16)));
38 __m128i vacc3 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (x + 24)));
39 x += 32;
40
41 __m128i vmultiplier0 = _mm_cmpgt_epi16(vacc0, vinput_zero_point);
42 vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
43 __m128i vmultiplier1 = _mm_cmpgt_epi16(vacc1, vinput_zero_point);
44 vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
45 __m128i vmultiplier2 = _mm_cmpgt_epi16(vacc2, vinput_zero_point);
46 vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
47 __m128i vmultiplier3 = _mm_cmpgt_epi16(vacc3, vinput_zero_point);
48 vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
49
50 vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
51 vacc0 = _mm_slli_epi16(vacc0, 7);
52 vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
53 vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
54 vacc1 = _mm_slli_epi16(vacc1, 7);
55 vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
56 vmultiplier2 = _mm_and_si128(vmultiplier2, vmultiplier_diff);
57 vacc2 = _mm_slli_epi16(vacc2, 7);
58 vmultiplier2 = _mm_xor_si128(vmultiplier2, vmultiplier_base);
59 vmultiplier3 = _mm_and_si128(vmultiplier3, vmultiplier_diff);
60 vacc3 = _mm_slli_epi16(vacc3, 7);
61 vmultiplier3 = _mm_xor_si128(vmultiplier3, vmultiplier_base);
62
63 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier0);
64 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier1);
65 vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier2);
66 vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier3);
67
68 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
69 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
70 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
71 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
72
73 const __m128i vy0 = _mm_packs_epi16(vacc0, vacc1);
74 const __m128i vy1 = _mm_packs_epi16(vacc2, vacc3);
75
76 _mm_storeu_si128((__m128i*) y, vy0);
77 _mm_storeu_si128((__m128i*) (y + 16), vy1);
78 y += 32;
79 }
80 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
81 __m128i vacc = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) x));
82 __m128i vmultiplier = _mm_cmpgt_epi16(vacc, vinput_zero_point);
83 vacc = _mm_sub_epi16(vinput_zero_point, vacc);
84 vmultiplier = _mm_and_si128(vmultiplier, vmultiplier_diff);
85 vacc = _mm_slli_epi16(vacc, 7);
86 vmultiplier = _mm_xor_si128(vmultiplier, vmultiplier_base);
87 vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
88 vacc = _mm_adds_epi16(vacc, voutput_zero_point);
89 x += 8;
90
91 const __m128i vy = _mm_packs_epi16(vacc, vacc);
92 _mm_storel_epi64((__m128i*) y, vy);
93 y += 8;
94 }
95 if XNN_UNLIKELY(n != 0) {
96 assert(n >= 1 * sizeof(int8_t));
97 assert(n <= 7 * sizeof(int8_t));
98
99 __m128i vacc = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) x));
100 __m128i vmultiplier = _mm_cmpgt_epi16(vacc, vinput_zero_point);
101 vacc = _mm_sub_epi16(vinput_zero_point, vacc);
102 vmultiplier = _mm_and_si128(vmultiplier, vmultiplier_diff);
103 vacc = _mm_slli_epi16(vacc, 7);
104 vmultiplier = _mm_xor_si128(vmultiplier, vmultiplier_base);
105 vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
106 vacc = _mm_adds_epi16(vacc, voutput_zero_point);
107
108 __m128i vy = _mm_packs_epi16(vacc, vacc);
109 if (n & (4 * sizeof(int8_t))) {
110 _mm_storeu_si32(y, vy);
111 vy = _mm_srli_epi64(vy, 32);
112 y += 4;
113 }
114 if (n & (2 * sizeof(int8_t))) {
115 _mm_storeu_si16(y, vy);
116 vy = _mm_srli_epi32(vy, 16);
117 y += 2;
118 }
119 if (n & (1 * sizeof(int8_t))) {
120 *y = (int8_t) _mm_extract_epi8(vy, 0);
121 }
122 }
123 }
124