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_x8(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_x8(
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 >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
35 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
36 input_a += 8;
37
38
39 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
40 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
41
42 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
43
44
45 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
46 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
47
48 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
49 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
50
51 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
52
53
54 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
55
56 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
57
58 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
59
60 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
61 output += 8;
62 }
63 if XNN_UNLIKELY(n != 0) {
64 {
65 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
66
67
68 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
69 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
70
71 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
72
73
74 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
75 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
76
77 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
78 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
79
80 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
81
82 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
83 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
84 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
85
86 if (n & (4 * sizeof(uint8_t))) {
87 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
88 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
89 output += 4;
90 }
91 if (n & (2 * sizeof(uint8_t))) {
92 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
93 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
94 output += 2;
95 }
96 if (n & (1 * sizeof(uint8_t))) {
97 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
98 }
99 }
100 }
101 }
102