Debian Codegraph @

src/qu8-vcvt/gen/qu8-vcvt-ssse3-u32.c

raw
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
19 void xnn_qu8_vcvt_ukernel__ssse3_u32(
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 vinput_zero_point = _mm_set1_epi16(params->scalar.input_zero_point);
31 const __m128i vmultiplier = _mm_set1_epi16(-params->scalar.multiplier);
32 const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point);
33 XNN_FORCE_REALIZATION(vinput_zero_point);
34 XNN_FORCE_REALIZATION(vmultiplier);
35 XNN_FORCE_REALIZATION(voutput_zero_point);
36 const __m128i vzero = _mm_setzero_si128();
37 for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) {
38 const __m128i vx0 = _mm_loadu_si128((const __m128i*) input);
39 const __m128i vx1 = _mm_loadu_si128((const __m128i*) (input + 16));
40 input += 32;
41
42 __m128i vacc0 = _mm_unpacklo_epi8(vx0, vzero);
43 __m128i vacc1 = _mm_unpackhi_epi8(vx0, vzero);
44 __m128i vacc2 = _mm_unpacklo_epi8(vx1, vzero);
45 __m128i vacc3 = _mm_unpackhi_epi8(vx1, vzero);
46
47 vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
48 vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
49 vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
50 vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
51
52 vacc0 = _mm_slli_epi16(vacc0, 7);
53 vacc1 = _mm_slli_epi16(vacc1, 7);
54 vacc2 = _mm_slli_epi16(vacc2, 7);
55 vacc3 = _mm_slli_epi16(vacc3, 7);
56
57 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
58 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
59 vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier);
60 vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier);
61
62 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
63 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
64 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
65 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
66
67 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
68 const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
69
70 _mm_storeu_si128((__m128i*) output, vy0);
71 _mm_storeu_si128((__m128i*) (output + 16), vy1);
72 output += 32;
73 }
74 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
75 const __m128i vx = _mm_loadu_si128((const __m128i*) input);
76 input += 16;
77
78 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
79 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
80 vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
81 vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
82 vacc_lo = _mm_slli_epi16(vacc_lo, 7);
83 vacc_hi = _mm_slli_epi16(vacc_hi, 7);
84 vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
85 vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
86 vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
87 vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
88
89 const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
90 _mm_storeu_si128((__m128i*) output, vy);
91 output += 16;
92 }
93 if XNN_UNLIKELY(batch != 0) {
94 assert(batch >= 1 * sizeof(uint8_t));
95 assert(batch <= 15 * sizeof(uint8_t));
96
97 const __m128i vx = _mm_loadu_si128((const __m128i*) input);
98
99 __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
100 __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
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_packus_epi16(vacc_lo, vacc_hi);
111 if (batch & (8 * sizeof(uint8_t))) {
112 _mm_storel_epi64((__m128i*) output, vy);
113 vy = _mm_unpackhi_epi64(vy, vy);
114 output += 8;
115 }
116 if (batch & (4 * sizeof(uint8_t))) {
117 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vy));
118 vy = _mm_srli_epi64(vy, 32);
119 output += 4;
120 }
121 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
122 if (batch & (2 * sizeof(uint8_t))) {
123 unaligned_store_u16(output, (uint16_t) vy_lo);
124 vy_lo >>= 16;
125 output += 2;
126 }
127 if (batch & (1 * sizeof(uint8_t))) {
128 *output = (uint8_t) vy_lo;
129 }
130 }
131 }