Debian Codegraph @

src/qu8-vcvt/gen/qu8-vcvt-avx2-u16.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_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 __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 >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
37 __m256i vacc = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) input));
38 vacc = _mm256_sub_epi16(vinput_zero_point, vacc);
39 vacc = _mm256_slli_epi16(vacc, 7);
40 vacc = _mm256_mulhrs_epi16(vacc, vmultiplier);
41 vacc = _mm256_adds_epi16(vacc, voutput_zero_point);
42 input += 16;
43
44 const __m128i vacc_hi = _mm256_extracti128_si256(vacc, 1);
45 const __m128i vy = _mm_packus_epi16(_mm256_castsi256_si128(vacc), vacc_hi);
46 _mm_storeu_si128((__m128i*) output, vy);
47 output += 16;
48 }
49 if XNN_UNLIKELY(batch != 0) {
50 assert(batch >= 1 * sizeof(uint8_t));
51 assert(batch <= 15 * sizeof(uint8_t));
52
53 __m256i vacc = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i*) input));
54 vacc = _mm256_sub_epi16(vinput_zero_point, vacc);
55 vacc = _mm256_slli_epi16(vacc, 7);
56 vacc = _mm256_mulhrs_epi16(vacc, vmultiplier);
57 vacc = _mm256_adds_epi16(vacc, voutput_zero_point);
58
59 const __m128i vacc_hi = _mm256_extracti128_si256(vacc, 1);
60 __m128i vy = _mm_packus_epi16(_mm256_castsi256_si128(vacc), vacc_hi);
61 if (batch & (8 * sizeof(uint8_t))) {
62 _mm_storel_epi64((__m128i*) output, vy);
63 vy = _mm_unpackhi_epi64(vy, vy);
64 output += 8;
65 }
66 if (batch & (4 * sizeof(uint8_t))) {
67 _mm_storeu_si32(output, vy);
68 vy = _mm_srli_epi64(vy, 32);
69 output += 4;
70 }
71 if (batch & (2 * sizeof(uint8_t))) {
72 _mm_storeu_si16(output, vy);
73 vy = _mm_srli_epi32(vy, 16);
74 output += 2;
75 }
76 if (batch & (1 * sizeof(uint8_t))) {
77 *output = (uint8_t) _mm_extract_epi8(vy, 0);
78 }
79 }
80 }