1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/sse-mul16-ld64.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <smmintrin.h>
13
14 #include <xnnpack/vaddsub.h>
15
16
xnn_qu8_vaddc_minmax_ukernel__sse41_mul16_ld64_x16(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_vaddc_minmax_ukernel__sse41_mul16_ld64_x16(
18 size_t n,
19 const uint8_t* input_a,
20 const uint8_t* input_b,
21 uint8_t* output,
22 const union xnn_qu8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 const __m128i vbias = _mm_add_epi32(
25 _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
26 _mm_load_si128((const __m128i*) params->sse2.bias));
27 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
28 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
29 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
30 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
31 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
32 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
33
34 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
35 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
36 const __m128i va89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
37 input_a += 16;
38
39
40 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
41 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
42 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
43 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
44
45 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
46 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
47
48
49 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
50 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
51 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
52 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
53
54 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
55 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
56 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
57 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
58
59 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
60 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
61
62
63 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
64
65 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
66
67 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
68
69 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
70 output += 16;
71 }
72 if XNN_UNLIKELY(n != 0) {
73 do {
74 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
75 input_a += 8;
76
77
78 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
79 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
80
81 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
82
83
84 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
85 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
86
87 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
88 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
89
90 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
91
92 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
93 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
94 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
95
96 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
97 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
98 output += 8;
99 n -= 8 * sizeof(uint8_t);
100 } else {
101 if (n & (4 * sizeof(uint8_t))) {
102 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
103 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
104 output += 4;
105 }
106 if (n & (2 * sizeof(uint8_t))) {
107 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
108 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
109 output += 2;
110 }
111 if (n & (1 * sizeof(uint8_t))) {
112 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
113 }
114 n = 0;
115 }
116 } while (n != 0);
117 }
118 }
119