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_qu8_vlrelu_ukernel__avx_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__avx_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->avx.input_zero_point);
31 const __m128i vpositive_multiplier = _mm_load_si128((const __m128i*) params->avx.positive_multiplier);
32 const __m128i vnegative_multiplier = _mm_load_si128((const __m128i*) params->avx.negative_multiplier);
33 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx.output_zero_point);
34 for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
35 __m128i vacc0 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
36 __m128i vacc1 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (x + 8)));
37 __m128i vacc2 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (x + 16)));
38 __m128i vacc3 = _mm_cvtepu8_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_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier0);
51 vacc0 = _mm_slli_epi16(vacc0, 7);
52 vmultiplier1 = _mm_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier1);
53 vacc1 = _mm_slli_epi16(vacc1, 7);
54 vmultiplier2 = _mm_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier2);
55 vacc2 = _mm_slli_epi16(vacc2, 7);
56 vmultiplier3 = _mm_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier3);
57 vacc3 = _mm_slli_epi16(vacc3, 7);
58
59 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier0);
60 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier1);
61 vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier2);
62 vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier3);
63
64 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
65 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
66 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
67 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
68
69 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
70 const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
71
72 _mm_storeu_si128((__m128i*) y, vy0);
73 _mm_storeu_si128((__m128i*) (y + 16), vy1);
74 y += 32;
75 }
76 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
77 __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
78 __m128i vmultiplier = _mm_cmpgt_epi16(vacc, vinput_zero_point);
79 vacc = _mm_sub_epi16(vinput_zero_point, vacc);
80 vmultiplier = _mm_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier);
81 vacc = _mm_slli_epi16(vacc, 7);
82 vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
83 vacc = _mm_adds_epi16(vacc, voutput_zero_point);
84 x += 8;
85
86 const __m128i vy = _mm_packus_epi16(vacc, vacc);
87 _mm_storel_epi64((__m128i*) y, vy);
88 y += 8;
89 }
90 if XNN_UNLIKELY(n != 0) {
91 assert(n >= 1 * sizeof(uint8_t));
92 assert(n <= 7 * sizeof(uint8_t));
93
94 __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
95 __m128i vmultiplier = _mm_cmpgt_epi16(vacc, vinput_zero_point);
96 vacc = _mm_sub_epi16(vinput_zero_point, vacc);
97 vmultiplier = _mm_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier);
98 vacc = _mm_slli_epi16(vacc, 7);
99 vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
100 vacc = _mm_adds_epi16(vacc, voutput_zero_point);
101
102 __m128i vy = _mm_packus_epi16(vacc, vacc);
103 if (n & (4 * sizeof(uint8_t))) {
104 _mm_storeu_si32(y, vy);
105 vy = _mm_srli_epi64(vy, 32);
106 y += 4;
107 }
108 if (n & (2 * sizeof(uint8_t))) {
109 _mm_storeu_si16(y, vy);
110 vy = _mm_srli_epi32(vy, 16);
111 y += 2;
112 }
113 if (n & (1 * sizeof(uint8_t))) {
114 *y = (uint8_t) _mm_extract_epi8(vy, 0);
115 }
116 }
117 }
118