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
xnn_qu8_vcvt_ukernel__sse41_x16(size_t n,const uint8_t * x,uint8_t * y,const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_vcvt_ukernel__sse41_x16(
20 size_t n,
21 const uint8_t* x,
22 uint8_t* y,
23 const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(uint8_t) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.input_zero_point);
31 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->ssse3.multiplier);
32 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.output_zero_point);
33 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
34 __m128i vacc0 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
35 __m128i vacc1 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (x + 8)));
36 x += 16;
37
38 vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
39 vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
40
41 vacc0 = _mm_slli_epi16(vacc0, 7);
42 vacc1 = _mm_slli_epi16(vacc1, 7);
43
44 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
45 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
46
47 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
48 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
49
50 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
51
52 _mm_storeu_si128((__m128i*) y, vy0);
53 y += 16;
54 }
55 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
56 __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
57 vacc = _mm_sub_epi16(vinput_zero_point, vacc);
58 vacc = _mm_slli_epi16(vacc, 7);
59 vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
60 vacc = _mm_adds_epi16(vacc, voutput_zero_point);
61 x += 8;
62
63 const __m128i vy = _mm_packus_epi16(vacc, vacc);
64 _mm_storel_epi64((__m128i*) y, vy);
65 y += 8;
66 }
67 if XNN_UNLIKELY(n != 0) {
68 assert(n >= 1 * sizeof(uint8_t));
69 assert(n <= 7 * sizeof(uint8_t));
70
71 __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
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
77 __m128i vy = _mm_packus_epi16(vacc, vacc);
78 if (n & (4 * sizeof(uint8_t))) {
79 _mm_storeu_si32(y, vy);
80 vy = _mm_srli_epi64(vy, 32);
81 y += 4;
82 }
83 if (n & (2 * sizeof(uint8_t))) {
84 _mm_storeu_si16(y, vy);
85 vy = _mm_srli_epi32(vy, 16);
86 y += 2;
87 }
88 if (n & (1 * sizeof(uint8_t))) {
89 *y = (uint8_t) _mm_extract_epi8(vy, 0);
90 }
91 }
92 }
93