Debian Codegraph @

src/qu8-vcvt/gen/qu8-vcvt-avx2-u64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vcvt/avx2.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 <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/vcvt.h"
17
18
19 void xnn_qu8_vcvt_ukernel__avx2_u64(
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 __m256i vinput_zero_point = _mm256_set1_epi16(params->scalar.input_zero_point);
31 const __m256i vmultiplier = _mm256_set1_epi16(-params->scalar.multiplier);
32 const __m256i voutput_zero_point = _mm256_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 for (; batch >= 64 * sizeof(uint8_t); batch -= 64 * sizeof(uint8_t)) {
37 __m256i vacc0 = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) input));
38 __m256i vacc1 = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) (input + 16)));
39 __m256i vacc2 = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) (input + 32)));
40 __m256i vacc3 = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) (input + 48)));
41 input += 64;
42
43 vacc0 = _mm256_sub_epi16(vinput_zero_point, vacc0);
44 vacc1 = _mm256_sub_epi16(vinput_zero_point, vacc1);
45 vacc2 = _mm256_sub_epi16(vinput_zero_point, vacc2);
46 vacc3 = _mm256_sub_epi16(vinput_zero_point, vacc3);
47
48 vacc0 = _mm256_slli_epi16(vacc0, 7);
49 vacc1 = _mm256_slli_epi16(vacc1, 7);
50 vacc2 = _mm256_slli_epi16(vacc2, 7);
51 vacc3 = _mm256_slli_epi16(vacc3, 7);
52
53 vacc0 = _mm256_mulhrs_epi16(vacc0, vmultiplier);
54 vacc1 = _mm256_mulhrs_epi16(vacc1, vmultiplier);
55 vacc2 = _mm256_mulhrs_epi16(vacc2, vmultiplier);
56 vacc3 = _mm256_mulhrs_epi16(vacc3, vmultiplier);
57
58 vacc0 = _mm256_adds_epi16(vacc0, voutput_zero_point);
59 vacc1 = _mm256_adds_epi16(vacc1, voutput_zero_point);
60 vacc2 = _mm256_adds_epi16(vacc2, voutput_zero_point);
61 vacc3 = _mm256_adds_epi16(vacc3, voutput_zero_point);
62
63 __m256i vy0 = _mm256_packus_epi16(vacc0, vacc1);
64 __m256i vy1 = _mm256_packus_epi16(vacc2, vacc3);
65
66 vy0 = _mm256_permute4x64_epi64(vy0, _MM_SHUFFLE(3, 1, 2, 0));
67 vy1 = _mm256_permute4x64_epi64(vy1, _MM_SHUFFLE(3, 1, 2, 0));
68
69 _mm256_storeu_si256((__m256i*) output, vy0);
70 _mm256_storeu_si256((__m256i*) (output + 32), vy1);
71 output += 64;
72 }
73 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
74 __m256i vacc = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) input));
75 vacc = _mm256_sub_epi16(vinput_zero_point, vacc);
76 vacc = _mm256_slli_epi16(vacc, 7);
77 vacc = _mm256_mulhrs_epi16(vacc, vmultiplier);
78 vacc = _mm256_adds_epi16(vacc, voutput_zero_point);
79 input += 16;
80
81 const __m128i vacc_hi = _mm256_extracti128_si256(vacc, 1);
82 const __m128i vy = _mm_packus_epi16(_mm256_castsi256_si128(vacc), vacc_hi);
83 _mm_storeu_si128((__m128i*) output, vy);
84 output += 16;
85 }
86 if XNN_UNLIKELY(batch != 0) {
87 assert(batch >= 1 * sizeof(uint8_t));
88 assert(batch <= 15 * sizeof(uint8_t));
89
90 __m256i vacc = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) input));
91 vacc = _mm256_sub_epi16(vinput_zero_point, vacc);
92 vacc = _mm256_slli_epi16(vacc, 7);
93 vacc = _mm256_mulhrs_epi16(vacc, vmultiplier);
94 vacc = _mm256_adds_epi16(vacc, voutput_zero_point);
95
96 const __m128i vacc_hi = _mm256_extracti128_si256(vacc, 1);
97 __m128i vy = _mm_packus_epi16(_mm256_castsi256_si128(vacc), vacc_hi);
98 if (batch & (8 * sizeof(uint8_t))) {
99 _mm_storel_epi64((__m128i*) output, vy);
100 vy = _mm_unpackhi_epi64(vy, vy);
101 output += 8;
102 }
103 if (batch & (4 * sizeof(uint8_t))) {
104 _mm_storeu_si32(output, vy);
105 vy = _mm_srli_epi64(vy, 32);
106 output += 4;
107 }
108 if (batch & (2 * sizeof(uint8_t))) {
109 _mm_storeu_si16(output, vy);
110 vy = _mm_srli_epi32(vy, 16);
111 output += 2;
112 }
113 if (batch & (1 * sizeof(uint8_t))) {
114 *output = (uint8_t) _mm_extract_epi8(vy, 0);
115 }
116 }
117 }