1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vlrelu/avx2.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__avx2_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__avx2_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 __m256i vinput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.input_zero_point);
31 const __m256i vpositive_multiplier = _mm256_load_si256((const __m256i*) params->avx2.positive_multiplier);
32 const __m256i vnegative_multiplier = _mm256_load_si256((const __m256i*) params->avx2.negative_multiplier);
33 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
34 for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
35 __m256i vacc0 = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) x));
36 __m256i vacc1 = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (x + 16)));
37 x += 32;
38
39 __m256i vmultiplier0 = _mm256_cmpgt_epi16(vacc0, vinput_zero_point);
40 vacc0 = _mm256_sub_epi16(vinput_zero_point, vacc0);
41 __m256i vmultiplier1 = _mm256_cmpgt_epi16(vacc1, vinput_zero_point);
42 vacc1 = _mm256_sub_epi16(vinput_zero_point, vacc1);
43
44 vmultiplier0 = _mm256_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier0);
45 vacc0 = _mm256_slli_epi16(vacc0, 7);
46 vmultiplier1 = _mm256_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier1);
47 vacc1 = _mm256_slli_epi16(vacc1, 7);
48
49 vacc0 = _mm256_mulhrs_epi16(vacc0, vmultiplier0);
50 vacc1 = _mm256_mulhrs_epi16(vacc1, vmultiplier1);
51
52 vacc0 = _mm256_adds_epi16(vacc0, voutput_zero_point);
53 vacc1 = _mm256_adds_epi16(vacc1, voutput_zero_point);
54
55 __m256i vy0 = _mm256_packs_epi16(vacc0, vacc1);
56
57 vy0 = _mm256_permute4x64_epi64(vy0, _MM_SHUFFLE(3, 1, 2, 0));
58
59 _mm256_storeu_si256((__m256i*) y, vy0);
60 y += 32;
61 }
62 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
63 __m256i vacc = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) x));
64 __m256i vmultiplier = _mm256_cmpgt_epi16(vacc, vinput_zero_point);
65 vacc = _mm256_sub_epi16(vinput_zero_point, vacc);
66 vmultiplier = _mm256_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier);
67 vacc = _mm256_slli_epi16(vacc, 7);
68 vacc = _mm256_mulhrs_epi16(vacc, vmultiplier);
69 vacc = _mm256_adds_epi16(vacc, voutput_zero_point);
70 x += 16;
71
72 const __m128i vacc_hi = _mm256_extracti128_si256(vacc, 1);
73 const __m128i vy = _mm_packs_epi16(_mm256_castsi256_si128(vacc), vacc_hi);
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 __m256i vacc = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) x));
82 __m256i vmultiplier = _mm256_cmpgt_epi16(vacc, vinput_zero_point);
83 vacc = _mm256_sub_epi16(vinput_zero_point, vacc);
84 vmultiplier = _mm256_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier);
85 vacc = _mm256_slli_epi16(vacc, 7);
86 vacc = _mm256_mulhrs_epi16(vacc, vmultiplier);
87 vacc = _mm256_adds_epi16(vacc, voutput_zero_point);
88
89 const __m128i vacc_hi = _mm256_extracti128_si256(vacc, 1);
90 __m128i vy = _mm_packs_epi16(_mm256_castsi256_si128(vacc), vacc_hi);
91 if (n & (8 * sizeof(int8_t))) {
92 _mm_storel_epi64((__m128i*) y, vy);
93 vy = _mm_unpackhi_epi64(vy, vy);
94 y += 8;
95 }
96 if (n & (4 * sizeof(int8_t))) {
97 _mm_storeu_si32(y, vy);
98 vy = _mm_srli_epi64(vy, 32);
99 y += 4;
100 }
101 if (n & (2 * sizeof(int8_t))) {
102 _mm_storeu_si16(y, vy);
103 vy = _mm_srli_epi32(vy, 16);
104 y += 2;
105 }
106 if (n & (1 * sizeof(int8_t))) {
107 *y = (int8_t) _mm_extract_epi8(vy, 0);
108 }
109 }
110 }
111