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_x32(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_x32(
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 >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
34 const __m128i vx0 = _mm_loadu_si128((const __m128i*) x);
35 const __m128i vx1 = _mm_loadu_si128((const __m128i*) (x + 16));
36 x += 32;
37
38 const __m128i vm0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx0);
39 __m128i vacc0 = _mm_unpacklo_epi8(vx0, vm0);
40 __m128i vacc1 = _mm_unpackhi_epi8(vx0, vm0);
41 const __m128i vm1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx1);
42 __m128i vacc2 = _mm_unpacklo_epi8(vx1, vm1);
43 __m128i vacc3 = _mm_unpackhi_epi8(vx1, vm1);
44
45 vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
46 vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
47 vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
48 vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
49
50 vacc0 = _mm_slli_epi16(vacc0, 7);
51 vacc1 = _mm_slli_epi16(vacc1, 7);
52 vacc2 = _mm_slli_epi16(vacc2, 7);
53 vacc3 = _mm_slli_epi16(vacc3, 7);
54
55 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
56 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
57 vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier);
58 vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier);
59
60 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
61 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
62 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
63 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
64
65 const __m128i vy0 = _mm_packs_epi16(vacc0, vacc1);
66 const __m128i vy1 = _mm_packs_epi16(vacc2, vacc3);
67
68 _mm_storeu_si128((__m128i*) y, vy0);
69 _mm_storeu_si128((__m128i*) (y + 16), vy1);
70 y += 32;
71 }
72 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
73 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
74 x += 16;
75
76 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
77 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
78 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
79 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
80 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
81 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
82 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
83 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
84 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
85 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
86 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
87
88 const __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
89 _mm_storeu_si128((__m128i*) y, vy);
90 y += 16;
91 }
92 if XNN_UNLIKELY(n != 0) {
93 assert(n >= 1 * sizeof(int8_t));
94 assert(n <= 15 * sizeof(int8_t));
95
96 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
97
98 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
99 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
100 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
101 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
102 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
103 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
104 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
105 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
106 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
107 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
108 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
109
110 __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
111 if (n & (8 * sizeof(int8_t))) {
112 _mm_storel_epi64((__m128i*) y, vy);
113 vy = _mm_unpackhi_epi64(vy, vy);
114 y += 8;
115 }
116 if (n & (4 * sizeof(int8_t))) {
117 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
118 vy = _mm_srli_epi64(vy, 32);
119 y += 4;
120 }
121 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
122 if (n & (2 * sizeof(int8_t))) {
123 unaligned_store_u16(y, (uint16_t) vy_lo);
124 vy_lo >>= 16;
125 y += 2;
126 }
127 if (n & (1 * sizeof(int8_t))) {
128 *y = (int8_t) vy_lo;
129 }
130 }
131 }
132