Debian Codegraph @

src/qu8-vcvt/gen/qu8-vcvt-avx2-u32.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_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 __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 >= 32 * sizeof(uint8_t); batch -= 32 * 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 input += 32;
40
41 vacc0 = _mm256_sub_epi16(vinput_zero_point, vacc0);
42 vacc1 = _mm256_sub_epi16(vinput_zero_point, vacc1);
43
44 vacc0 = _mm256_slli_epi16(vacc0, 7);
45 vacc1 = _mm256_slli_epi16(vacc1, 7);
46
47 vacc0 = _mm256_mulhrs_epi16(vacc0, vmultiplier);
48 vacc1 = _mm256_mulhrs_epi16(vacc1, vmultiplier);
49
50 vacc0 = _mm256_adds_epi16(vacc0, voutput_zero_point);
51 vacc1 = _mm256_adds_epi16(vacc1, voutput_zero_point);
52
53 __m256i vy0 = _mm256_packus_epi16(vacc0, vacc1);
54
55 vy0 = _mm256_permute4x64_epi64(vy0, _MM_SHUFFLE(3, 1, 2, 0));
56
57 _mm256_storeu_si256((__m256i*) output, vy0);
58 output += 32;
59 }
60 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
61 __m256i vacc = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) input));
62 vacc = _mm256_sub_epi16(vinput_zero_point, vacc);
63 vacc = _mm256_slli_epi16(vacc, 7);
64 vacc = _mm256_mulhrs_epi16(vacc, vmultiplier);
65 vacc = _mm256_adds_epi16(vacc, voutput_zero_point);
66 input += 16;
67
68 const __m128i vacc_hi = _mm256_extracti128_si256(vacc, 1);
69 const __m128i vy = _mm_packus_epi16(_mm256_castsi256_si128(vacc), vacc_hi);
70 _mm_storeu_si128((__m128i*) output, vy);
71 output += 16;
72 }
73 if XNN_UNLIKELY(batch != 0) {
74 assert(batch >= 1 * sizeof(uint8_t));
75 assert(batch <= 15 * sizeof(uint8_t));
76
77 __m256i vacc = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) input));
78 vacc = _mm256_sub_epi16(vinput_zero_point, vacc);
79 vacc = _mm256_slli_epi16(vacc, 7);
80 vacc = _mm256_mulhrs_epi16(vacc, vmultiplier);
81 vacc = _mm256_adds_epi16(vacc, voutput_zero_point);
82
83 const __m128i vacc_hi = _mm256_extracti128_si256(vacc, 1);
84 __m128i vy = _mm_packus_epi16(_mm256_castsi256_si128(vacc), vacc_hi);
85 if (batch & (8 * sizeof(uint8_t))) {
86 _mm_storel_epi64((__m128i*) output, vy);
87 vy = _mm_unpackhi_epi64(vy, vy);
88 output += 8;
89 }
90 if (batch & (4 * sizeof(uint8_t))) {
91 _mm_storeu_si32(output, vy);
92 vy = _mm_srli_epi64(vy, 32);
93 output += 4;
94 }
95 if (batch & (2 * sizeof(uint8_t))) {
96 _mm_storeu_si16(output, vy);
97 vy = _mm_srli_epi32(vy, 16);
98 output += 2;
99 }
100 if (batch & (1 * sizeof(uint8_t))) {
101 *output = (uint8_t) _mm_extract_epi8(vy, 0);
102 }
103 }
104 }