Debian Codegraph @

src/qu8-vcvt/gen/qu8-vcvt-sse2-u16.c

raw
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
19 void xnn_qu8_vcvt_ukernel__sse2_u16(
20 size_t batch,
21 const uint8_t* input,
22 uint8_t* output,
23 const struct xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 const __m128i vmultiplier = _mm_set1_epi16(params->scalar.multiplier);
31 const __m128i vbias = _mm_set1_epi32(
32 (int32_t) (((uint32_t) (int32_t) params->scalar.output_zero_point) << 8) -
33 (int32_t) params->scalar.multiplier * (int32_t) params->scalar.input_zero_point +
34 INT32_C(0x80));
35 XNN_FORCE_REALIZATION(vmultiplier);
36 XNN_FORCE_REALIZATION(vbias);
37 const __m128i vzero = _mm_setzero_si128();
38 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
39 const __m128i vx0 = _mm_loadu_si128((const __m128i*) input);
40 input += 16;
41
42 const __m128i vextx0 = _mm_unpacklo_epi8(vx0, vzero);
43 const __m128i vextx1 = _mm_unpackhi_epi8(vx0, vzero);
44
45 const __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier);
46 const __m128i vprodhi0 = _mm_mulhi_epu16(vextx0, vmultiplier);
47 const __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier);
48 const __m128i vprodhi1 = _mm_mulhi_epu16(vextx1, vmultiplier);
49
50 __m128i vacc0 = _mm_unpacklo_epi16(vprodlo0, vprodhi0);
51 __m128i vacc1 = _mm_unpackhi_epi16(vprodlo0, vprodhi0);
52 __m128i vacc2 = _mm_unpacklo_epi16(vprodlo1, vprodhi1);
53 __m128i vacc3 = _mm_unpackhi_epi16(vprodlo1, vprodhi1);
54
55 vacc0 = _mm_add_epi32(vacc0, vbias);
56 vacc1 = _mm_add_epi32(vacc1, vbias);
57 vacc2 = _mm_add_epi32(vacc2, vbias);
58 vacc3 = _mm_add_epi32(vacc3, vbias);
59
60 vacc0 = _mm_srai_epi32(vacc0, 8);
61 vacc1 = _mm_srai_epi32(vacc1, 8);
62 vacc2 = _mm_srai_epi32(vacc2, 8);
63 vacc3 = _mm_srai_epi32(vacc3, 8);
64
65 vacc0 = _mm_packs_epi32(vacc0, vacc1);
66 vacc1 = _mm_packs_epi32(vacc2, vacc3);
67
68 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
69
70 _mm_storeu_si128((__m128i*) output, vy0);
71 output += 16;
72 }
73 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
74 const __m128i vx = _mm_loadu_si128((const __m128i*) input);
75 input += 16;
76
77 const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vzero);
78 const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vzero);
79
80 const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier);
81 const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier);
82 const __m128i vprodhi_lo = _mm_mulhi_epu16(vextx_lo, vmultiplier);
83 const __m128i vprodhi_hi = _mm_mulhi_epu16(vextx_hi, vmultiplier);
84
85 __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo);
86 __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo);
87 __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi);
88 __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi);
89
90 vacc_ll = _mm_add_epi32(vacc_ll, vbias);
91 vacc_lh = _mm_add_epi32(vacc_lh, vbias);
92 vacc_hl = _mm_add_epi32(vacc_hl, vbias);
93 vacc_hh = _mm_add_epi32(vacc_hh, vbias);
94
95 vacc_ll = _mm_srai_epi32(vacc_ll, 8);
96 vacc_lh = _mm_srai_epi32(vacc_lh, 8);
97 vacc_hl = _mm_srai_epi32(vacc_hl, 8);
98 vacc_hh = _mm_srai_epi32(vacc_hh, 8);
99
100 const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh);
101 const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh);
102
103 const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
104 _mm_storeu_si128((__m128i*) output, vy);
105 output += 16;
106 }
107 if XNN_UNLIKELY(batch != 0) {
108 assert(batch >= 1 * sizeof(uint8_t));
109 assert(batch <= 15 * sizeof(uint8_t));
110
111 const __m128i vx = _mm_loadu_si128((const __m128i*) input);
112
113 const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vzero);
114 const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vzero);
115
116 const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier);
117 const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier);
118 const __m128i vprodhi_lo = _mm_mulhi_epu16(vextx_lo, vmultiplier);
119 const __m128i vprodhi_hi = _mm_mulhi_epu16(vextx_hi, vmultiplier);
120
121 __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo);
122 __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo);
123 __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi);
124 __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi);
125
126 vacc_ll = _mm_add_epi32(vacc_ll, vbias);
127 vacc_lh = _mm_add_epi32(vacc_lh, vbias);
128 vacc_hl = _mm_add_epi32(vacc_hl, vbias);
129 vacc_hh = _mm_add_epi32(vacc_hh, vbias);
130
131 vacc_ll = _mm_srai_epi32(vacc_ll, 8);
132 vacc_lh = _mm_srai_epi32(vacc_lh, 8);
133 vacc_hl = _mm_srai_epi32(vacc_hl, 8);
134 vacc_hh = _mm_srai_epi32(vacc_hh, 8);
135
136 const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh);
137 const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh);
138
139 __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
140 if (batch & (8 * sizeof(uint8_t))) {
141 _mm_storel_epi64((__m128i*) output, vy);
142 vy = _mm_unpackhi_epi64(vy, vy);
143 output += 8;
144 }
145 if (batch & (4 * sizeof(uint8_t))) {
146 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vy));
147 vy = _mm_srli_epi64(vy, 32);
148 output += 4;
149 }
150 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
151 if (batch & (2 * sizeof(uint8_t))) {
152 unaligned_store_u16(output, (uint16_t) vy_lo);
153 vy_lo >>= 16;
154 output += 2;
155 }
156 if (batch & (1 * sizeof(uint8_t))) {
157 *output = (uint8_t) vy_lo;
158 }
159 }
160 }