Debian Codegraph @

src/qu8-vcvt/gen/qu8-vcvt-sse41-u32.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vcvt/sse4.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__sse41_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 for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) {
37 __m128i vacc0 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input));
38 __m128i vacc1 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input + 8)));
39 __m128i vacc2 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input + 16)));
40 __m128i vacc3 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input + 24)));
41 input += 32;
42
43 vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
44 vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
45 vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
46 vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
47
48 vacc0 = _mm_slli_epi16(vacc0, 7);
49 vacc1 = _mm_slli_epi16(vacc1, 7);
50 vacc2 = _mm_slli_epi16(vacc2, 7);
51 vacc3 = _mm_slli_epi16(vacc3, 7);
52
53 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
54 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
55 vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier);
56 vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier);
57
58 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
59 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
60 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
61 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
62
63 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
64 const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
65
66 _mm_storeu_si128((__m128i*) output, vy0);
67 _mm_storeu_si128((__m128i*) (output + 16), vy1);
68 output += 32;
69 }
70 for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) {
71 __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input));
72 vacc = _mm_sub_epi16(vinput_zero_point, vacc);
73 vacc = _mm_slli_epi16(vacc, 7);
74 vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
75 vacc = _mm_adds_epi16(vacc, voutput_zero_point);
76 input += 8;
77
78 const __m128i vy = _mm_packus_epi16(vacc, vacc);
79 _mm_storel_epi64((__m128i*) output, vy);
80 output += 8;
81 }
82 if XNN_UNLIKELY(batch != 0) {
83 assert(batch >= 1 * sizeof(uint8_t));
84 assert(batch <= 7 * sizeof(uint8_t));
85
86 __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input));
87 vacc = _mm_sub_epi16(vinput_zero_point, vacc);
88 vacc = _mm_slli_epi16(vacc, 7);
89 vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
90 vacc = _mm_adds_epi16(vacc, voutput_zero_point);
91
92 __m128i vy = _mm_packus_epi16(vacc, vacc);
93 if (batch & (4 * sizeof(uint8_t))) {
94 _mm_storeu_si32(output, vy);
95 vy = _mm_srli_epi64(vy, 32);
96 output += 4;
97 }
98 if (batch & (2 * sizeof(uint8_t))) {
99 _mm_storeu_si16(output, vy);
100 vy = _mm_srli_epi32(vy, 16);
101 output += 2;
102 }
103 if (batch & (1 * sizeof(uint8_t))) {
104 *output = (uint8_t) _mm_extract_epi8(vy, 0);
105 }
106 }
107 }