1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vcvt/sse2.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__sse2_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__sse2_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 vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
31 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
32 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
33 const __m128i vx0 = _mm_loadu_si128((const __m128i*) x);
34 x += 16;
35
36 const __m128i vm0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx0);
37 const __m128i vextx0 = _mm_unpacklo_epi8(vx0, vm0);
38 const __m128i vextx1 = _mm_unpackhi_epi8(vx0, vm0);
39
40 const __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier);
41 const __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier);
42 const __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier);
43 const __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier);
44
45 __m128i vacc0 = _mm_unpacklo_epi16(vprodlo0, vprodhi0);
46 __m128i vacc1 = _mm_unpackhi_epi16(vprodlo0, vprodhi0);
47 __m128i vacc2 = _mm_unpacklo_epi16(vprodlo1, vprodhi1);
48 __m128i vacc3 = _mm_unpackhi_epi16(vprodlo1, vprodhi1);
49
50 vacc0 = _mm_sub_epi32(vbias, vacc0);
51 vacc1 = _mm_sub_epi32(vbias, vacc1);
52 vacc2 = _mm_sub_epi32(vbias, vacc2);
53 vacc3 = _mm_sub_epi32(vbias, vacc3);
54
55 vacc0 = _mm_srai_epi32(vacc0, 8);
56 vacc1 = _mm_srai_epi32(vacc1, 8);
57 vacc2 = _mm_srai_epi32(vacc2, 8);
58 vacc3 = _mm_srai_epi32(vacc3, 8);
59
60 vacc0 = _mm_packs_epi32(vacc0, vacc1);
61 vacc1 = _mm_packs_epi32(vacc2, vacc3);
62
63 const __m128i vy0 = _mm_packs_epi16(vacc0, vacc1);
64
65 _mm_storeu_si128((__m128i*) y, vy0);
66 y += 16;
67 }
68 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
69 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
70 x += 16;
71
72 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
73 const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vm);
74 const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vm);
75
76 const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier);
77 const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier);
78 const __m128i vprodhi_lo = _mm_mulhi_epi16(vextx_lo, vmultiplier);
79 const __m128i vprodhi_hi = _mm_mulhi_epi16(vextx_hi, vmultiplier);
80
81 __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo);
82 __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo);
83 __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi);
84 __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi);
85
86 vacc_ll = _mm_sub_epi32(vbias, vacc_ll);
87 vacc_lh = _mm_sub_epi32(vbias, vacc_lh);
88 vacc_hl = _mm_sub_epi32(vbias, vacc_hl);
89 vacc_hh = _mm_sub_epi32(vbias, vacc_hh);
90
91 vacc_ll = _mm_srai_epi32(vacc_ll, 8);
92 vacc_lh = _mm_srai_epi32(vacc_lh, 8);
93 vacc_hl = _mm_srai_epi32(vacc_hl, 8);
94 vacc_hh = _mm_srai_epi32(vacc_hh, 8);
95
96 const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh);
97 const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh);
98
99 const __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
100 _mm_storeu_si128((__m128i*) y, vy);
101 y += 16;
102 }
103 if XNN_UNLIKELY(n != 0) {
104 assert(n >= 1 * sizeof(int8_t));
105 assert(n <= 15 * sizeof(int8_t));
106
107 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
108
109 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
110 const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vm);
111 const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vm);
112
113 const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier);
114 const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier);
115 const __m128i vprodhi_lo = _mm_mulhi_epi16(vextx_lo, vmultiplier);
116 const __m128i vprodhi_hi = _mm_mulhi_epi16(vextx_hi, vmultiplier);
117
118 __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo);
119 __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo);
120 __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi);
121 __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi);
122
123 vacc_ll = _mm_sub_epi32(vbias, vacc_ll);
124 vacc_lh = _mm_sub_epi32(vbias, vacc_lh);
125 vacc_hl = _mm_sub_epi32(vbias, vacc_hl);
126 vacc_hh = _mm_sub_epi32(vbias, vacc_hh);
127
128 vacc_ll = _mm_srai_epi32(vacc_ll, 8);
129 vacc_lh = _mm_srai_epi32(vacc_lh, 8);
130 vacc_hl = _mm_srai_epi32(vacc_hl, 8);
131 vacc_hh = _mm_srai_epi32(vacc_hh, 8);
132
133 const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh);
134 const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh);
135
136 __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
137 if (n & (8 * sizeof(int8_t))) {
138 _mm_storel_epi64((__m128i*) y, vy);
139 vy = _mm_unpackhi_epi64(vy, vy);
140 y += 8;
141 }
142 if (n & (4 * sizeof(int8_t))) {
143 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
144 vy = _mm_srli_epi64(vy, 32);
145 y += 4;
146 }
147 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
148 if (n & (2 * sizeof(int8_t))) {
149 unaligned_store_u16(y, (uint16_t) vy_lo);
150 vy_lo >>= 16;
151 y += 2;
152 }
153 if (n & (1 * sizeof(int8_t))) {
154 *y = (int8_t) vy_lo;
155 }
156 }
157 }
158