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_x32(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_x32(
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 >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
35 const __m128i vx0 = _mm_loadu_si128((const __m128i*) x);
36 const __m128i vx1 = _mm_loadu_si128((const __m128i*) (x + 16));
37 x += 32;
38
39 __m128i vacc0 = _mm_unpacklo_epi8(vx0, vzero);
40 __m128i vacc1 = _mm_unpackhi_epi8(vx0, vzero);
41 __m128i vacc2 = _mm_unpacklo_epi8(vx1, vzero);
42 __m128i vacc3 = _mm_unpackhi_epi8(vx1, vzero);
43
44 vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
45 vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
46 vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
47 vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
48
49 vacc0 = _mm_slli_epi16(vacc0, 7);
50 vacc1 = _mm_slli_epi16(vacc1, 7);
51 vacc2 = _mm_slli_epi16(vacc2, 7);
52 vacc3 = _mm_slli_epi16(vacc3, 7);
53
54 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
55 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
56 vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier);
57 vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier);
58
59 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
60 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
61 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
62 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
63
64 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
65 const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
66
67 _mm_storeu_si128((__m128i*) y, vy0);
68 _mm_storeu_si128((__m128i*) (y + 16), vy1);
69 y += 32;
70 }
71 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
72 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
73 x += 16;
74
75 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
76 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
77 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
78 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
79 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
80 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
81 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
82 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
83 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
84 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
85
86 const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
87 _mm_storeu_si128((__m128i*) y, vy);
88 y += 16;
89 }
90 if XNN_UNLIKELY(n != 0) {
91 assert(n >= 1 * sizeof(uint8_t));
92 assert(n <= 15 * sizeof(uint8_t));
93
94 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
95
96 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
97 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
98 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
99 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
100 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
101 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
102 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
103 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
104 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
105 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
106
107 __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
108 if (n & (8 * sizeof(uint8_t))) {
109 _mm_storel_epi64((__m128i*) y, vy);
110 vy = _mm_unpackhi_epi64(vy, vy);
111 y += 8;
112 }
113 if (n & (4 * sizeof(uint8_t))) {
114 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
115 vy = _mm_srli_epi64(vy, 32);
116 y += 4;
117 }
118 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
119 if (n & (2 * sizeof(uint8_t))) {
120 unaligned_store_u16(y, (uint16_t) vy_lo);
121 vy_lo >>= 16;
122 y += 2;
123 }
124 if (n & (1 * sizeof(uint8_t))) {
125 *y = (uint8_t) vy_lo;
126 }
127 }
128 }
129