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_x32(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_x32(
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 >= 32 * sizeof(uint8_t); n -= 32 * 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 __m128i vacc2 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (x + 16)));
37 __m128i vacc3 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (x + 24)));
38 x += 32;
39
40 vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
41 vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
42 vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
43 vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
44
45 vacc0 = _mm_slli_epi16(vacc0, 7);
46 vacc1 = _mm_slli_epi16(vacc1, 7);
47 vacc2 = _mm_slli_epi16(vacc2, 7);
48 vacc3 = _mm_slli_epi16(vacc3, 7);
49
50 vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
51 vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
52 vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier);
53 vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier);
54
55 vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
56 vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
57 vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
58 vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
59
60 const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
61 const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
62
63 _mm_storeu_si128((__m128i*) y, vy0);
64 _mm_storeu_si128((__m128i*) (y + 16), vy1);
65 y += 32;
66 }
67 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
68 __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
69 vacc = _mm_sub_epi16(vinput_zero_point, vacc);
70 vacc = _mm_slli_epi16(vacc, 7);
71 vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
72 vacc = _mm_adds_epi16(vacc, voutput_zero_point);
73 x += 8;
74
75 const __m128i vy = _mm_packus_epi16(vacc, vacc);
76 _mm_storel_epi64((__m128i*) y, vy);
77 y += 8;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 assert(n >= 1 * sizeof(uint8_t));
81 assert(n <= 7 * sizeof(uint8_t));
82
83 __m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) x));
84 vacc = _mm_sub_epi16(vinput_zero_point, vacc);
85 vacc = _mm_slli_epi16(vacc, 7);
86 vacc = _mm_mulhrs_epi16(vacc, vmultiplier);
87 vacc = _mm_adds_epi16(vacc, voutput_zero_point);
88
89 __m128i vy = _mm_packus_epi16(vacc, vacc);
90 if (n & (4 * sizeof(uint8_t))) {
91 _mm_storeu_si32(y, vy);
92 vy = _mm_srli_epi64(vy, 32);
93 y += 4;
94 }
95 if (n & (2 * sizeof(uint8_t))) {
96 _mm_storeu_si16(y, vy);
97 vy = _mm_srli_epi32(vy, 16);
98 y += 2;
99 }
100 if (n & (1 * sizeof(uint8_t))) {
101 *y = (uint8_t) _mm_extract_epi8(vy, 0);
102 }
103 }
104 }
105