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_qs8_vcvt_ukernel__ssse3_x16(size_t n,const int8_t * x,int8_t * y,const union xnn_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_vcvt_ukernel__ssse3_x16(
20 size_t n,
21 const int8_t* x,
22 int8_t* y,
23 const union xnn_qs8_cvt_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->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 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
34 const __m128i vx0 = _mm_loadu_si128((const __m128i*) x);
35 x += 16;
36
37 const __m128i vm0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx0);
38 __m128i vacc0 = _mm_unpacklo_epi8(vx0, vm0);
39 __m128i vacc1 = _mm_unpackhi_epi8(vx0, vm0);
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_packs_epi16(vacc0, vacc1);
54
55 _mm_storeu_si128((__m128i*) y, vy0);
56 y += 16;
57 }
58 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
59 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
60 x += 16;
61
62 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
63 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
64 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
65 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
66 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
67 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
68 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
69 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
70 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
71 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
72 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
73
74 const __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
75 _mm_storeu_si128((__m128i*) y, vy);
76 y += 16;
77 }
78 if XNN_UNLIKELY(n != 0) {
79 assert(n >= 1 * sizeof(int8_t));
80 assert(n <= 15 * sizeof(int8_t));
81
82 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
83
84 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
85 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
86 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
87 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
88 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
89 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
90 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
91 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
92 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
93 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
94 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
95
96 __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
97 if (n & (8 * sizeof(int8_t))) {
98 _mm_storel_epi64((__m128i*) y, vy);
99 vy = _mm_unpackhi_epi64(vy, vy);
100 y += 8;
101 }
102 if (n & (4 * sizeof(int8_t))) {
103 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
104 vy = _mm_srli_epi64(vy, 32);
105 y += 4;
106 }
107 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
108 if (n & (2 * sizeof(int8_t))) {
109 unaligned_store_u16(y, (uint16_t) vy_lo);
110 vy_lo >>= 16;
111 y += 2;
112 }
113 if (n & (1 * sizeof(int8_t))) {
114 *y = (int8_t) vy_lo;
115 }
116 }
117 }
118