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_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__sse2_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 vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
31 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
32 for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
33 const __m128i vx0 = _mm_loadu_si128((const __m128i*) x);
34 const __m128i vx1 = _mm_loadu_si128((const __m128i*) (x + 16));
35 x += 32;
36
37 const __m128i vm0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx0);
38 const __m128i vextx0 = _mm_unpacklo_epi8(vx0, vm0);
39 const __m128i vextx1 = _mm_unpackhi_epi8(vx0, vm0);
40 const __m128i vm1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx1);
41 const __m128i vextx2 = _mm_unpacklo_epi8(vx1, vm1);
42 const __m128i vextx3 = _mm_unpackhi_epi8(vx1, vm1);
43
44 const __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier);
45 const __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier);
46 const __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier);
47 const __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier);
48 const __m128i vprodlo2 = _mm_mullo_epi16(vextx2, vmultiplier);
49 const __m128i vprodhi2 = _mm_mulhi_epi16(vextx2, vmultiplier);
50 const __m128i vprodlo3 = _mm_mullo_epi16(vextx3, vmultiplier);
51 const __m128i vprodhi3 = _mm_mulhi_epi16(vextx3, vmultiplier);
52
53 __m128i vacc0 = _mm_unpacklo_epi16(vprodlo0, vprodhi0);
54 __m128i vacc1 = _mm_unpackhi_epi16(vprodlo0, vprodhi0);
55 __m128i vacc2 = _mm_unpacklo_epi16(vprodlo1, vprodhi1);
56 __m128i vacc3 = _mm_unpackhi_epi16(vprodlo1, vprodhi1);
57 __m128i vacc4 = _mm_unpacklo_epi16(vprodlo2, vprodhi2);
58 __m128i vacc5 = _mm_unpackhi_epi16(vprodlo2, vprodhi2);
59 __m128i vacc6 = _mm_unpacklo_epi16(vprodlo3, vprodhi3);
60 __m128i vacc7 = _mm_unpackhi_epi16(vprodlo3, vprodhi3);
61
62 vacc0 = _mm_sub_epi32(vbias, vacc0);
63 vacc1 = _mm_sub_epi32(vbias, vacc1);
64 vacc2 = _mm_sub_epi32(vbias, vacc2);
65 vacc3 = _mm_sub_epi32(vbias, vacc3);
66 vacc4 = _mm_sub_epi32(vbias, vacc4);
67 vacc5 = _mm_sub_epi32(vbias, vacc5);
68 vacc6 = _mm_sub_epi32(vbias, vacc6);
69 vacc7 = _mm_sub_epi32(vbias, vacc7);
70
71 vacc0 = _mm_srai_epi32(vacc0, 8);
72 vacc1 = _mm_srai_epi32(vacc1, 8);
73 vacc2 = _mm_srai_epi32(vacc2, 8);
74 vacc3 = _mm_srai_epi32(vacc3, 8);
75 vacc4 = _mm_srai_epi32(vacc4, 8);
76 vacc5 = _mm_srai_epi32(vacc5, 8);
77 vacc6 = _mm_srai_epi32(vacc6, 8);
78 vacc7 = _mm_srai_epi32(vacc7, 8);
79
80 vacc0 = _mm_packs_epi32(vacc0, vacc1);
81 vacc1 = _mm_packs_epi32(vacc2, vacc3);
82 vacc2 = _mm_packs_epi32(vacc4, vacc5);
83 vacc3 = _mm_packs_epi32(vacc6, vacc7);
84
85 const __m128i vy0 = _mm_packs_epi16(vacc0, vacc1);
86 const __m128i vy1 = _mm_packs_epi16(vacc2, vacc3);
87
88 _mm_storeu_si128((__m128i*) y, vy0);
89 _mm_storeu_si128((__m128i*) (y + 16), vy1);
90 y += 32;
91 }
92 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
93 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
94 x += 16;
95
96 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
97 const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vm);
98 const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vm);
99
100 const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier);
101 const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier);
102 const __m128i vprodhi_lo = _mm_mulhi_epi16(vextx_lo, vmultiplier);
103 const __m128i vprodhi_hi = _mm_mulhi_epi16(vextx_hi, vmultiplier);
104
105 __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo);
106 __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo);
107 __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi);
108 __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi);
109
110 vacc_ll = _mm_sub_epi32(vbias, vacc_ll);
111 vacc_lh = _mm_sub_epi32(vbias, vacc_lh);
112 vacc_hl = _mm_sub_epi32(vbias, vacc_hl);
113 vacc_hh = _mm_sub_epi32(vbias, vacc_hh);
114
115 vacc_ll = _mm_srai_epi32(vacc_ll, 8);
116 vacc_lh = _mm_srai_epi32(vacc_lh, 8);
117 vacc_hl = _mm_srai_epi32(vacc_hl, 8);
118 vacc_hh = _mm_srai_epi32(vacc_hh, 8);
119
120 const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh);
121 const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh);
122
123 const __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
124 _mm_storeu_si128((__m128i*) y, vy);
125 y += 16;
126 }
127 if XNN_UNLIKELY(n != 0) {
128 assert(n >= 1 * sizeof(int8_t));
129 assert(n <= 15 * sizeof(int8_t));
130
131 const __m128i vx = _mm_loadu_si128((const __m128i*) x);
132
133 const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
134 const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vm);
135 const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vm);
136
137 const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier);
138 const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier);
139 const __m128i vprodhi_lo = _mm_mulhi_epi16(vextx_lo, vmultiplier);
140 const __m128i vprodhi_hi = _mm_mulhi_epi16(vextx_hi, vmultiplier);
141
142 __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo);
143 __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo);
144 __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi);
145 __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi);
146
147 vacc_ll = _mm_sub_epi32(vbias, vacc_ll);
148 vacc_lh = _mm_sub_epi32(vbias, vacc_lh);
149 vacc_hl = _mm_sub_epi32(vbias, vacc_hl);
150 vacc_hh = _mm_sub_epi32(vbias, vacc_hh);
151
152 vacc_ll = _mm_srai_epi32(vacc_ll, 8);
153 vacc_lh = _mm_srai_epi32(vacc_lh, 8);
154 vacc_hl = _mm_srai_epi32(vacc_hl, 8);
155 vacc_hh = _mm_srai_epi32(vacc_hh, 8);
156
157 const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh);
158 const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh);
159
160 __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
161 if (n & (8 * sizeof(int8_t))) {
162 _mm_storel_epi64((__m128i*) y, vy);
163 vy = _mm_unpackhi_epi64(vy, vy);
164 y += 8;
165 }
166 if (n & (4 * sizeof(int8_t))) {
167 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
168 vy = _mm_srli_epi64(vy, 32);
169 y += 4;
170 }
171 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
172 if (n & (2 * sizeof(int8_t))) {
173 unaligned_store_u16(y, (uint16_t) vy_lo);
174 vy_lo >>= 16;
175 y += 2;
176 }
177 if (n & (1 * sizeof(int8_t))) {
178 *y = (int8_t) vy_lo;
179 }
180 }
181 }
182