1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vcvt/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/vcvt.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_vcvt_ukernel__ssse3_x16(size_t n,const uint8_t * x,uint8_t * y,const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_vcvt_ukernel__ssse3_x16(
20 size_t n,
21 const uint8_t* x,
22 uint8_t* y,
23 const union xnn_qu8_cvt_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->ssse3.input_zero_point);
31 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->ssse3.multiplier);
32 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.output_zero_point);
33 const __m128i vzero = _mm_setzero_si128();
34 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
35 const __m128i vx0 = _mm_loadu_si128((const __m128i*) x);
36 x += 16;
37
38 __m128i vacc0 = _mm_unpacklo_epi8(vx0, vzero);
39 __m128i vacc1 = _mm_unpackhi_epi8(vx0, vzero);
40
41 vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
42 vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
43
44 vacc0 = _mm_slli_epi16(vacc0, 7);
45 vacc1 = _mm_slli_epi16(vacc1, 7);
46
47 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
48 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
49
50 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
51 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
52
53 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
54
55 _mm_storeu_si128((__m128i*) y, vy0);
56 y += 16;
57 }
58 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
59 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
60 x += 16;
61
62 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
63 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
64 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
65 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
66 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
67 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
68 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
69 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
70 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
71 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
72
73 const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
74 _mm_storeu_si128((__m128i*) y, vy);
75 y += 16;
76 }
77 if XNN_UNLIKELY(n != 0) {
78 assert(n >= 1 * sizeof(uint8_t));
79 assert(n <= 15 * sizeof(uint8_t));
80
81 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
82
83 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
84 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
85 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
86 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
87 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
88 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
89 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
90 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
91 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
92 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
93
94 __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
95 if (n & (8 * sizeof(uint8_t))) {
96 _mm_storel_epi64((__m128i*) y, vy);
97 vy = _mm_unpackhi_epi64(vy, vy);
98 y += 8;
99 }
100 if (n & (4 * sizeof(uint8_t))) {
101 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
102 vy = _mm_srli_epi64(vy, 32);
103 y += 4;
104 }
105 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
106 if (n & (2 * sizeof(uint8_t))) {
107 unaligned_store_u16(y, (uint16_t) vy_lo);
108 vy_lo >>= 16;
109 y += 2;
110 }
111 if (n & (1 * sizeof(uint8_t))) {
112 *y = (uint8_t) vy_lo;
113 }
114 }
115 }
116