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_qs8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8(
18 size_t n,
19 const int8_t* input_a,
20 const int8_t* input_b,
21 int8_t* output,
22 const union xnn_qs8_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->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
26 _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
27 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
28 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
29 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
30 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
31 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
32 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
33
34 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
35 const __m128i va01234567 = _mm_cvtepi8_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 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
45
46 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
47 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
48
49 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
50 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
51
52 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
53
54
55 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
56
57 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
58
59 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
60
61 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
62 output += 8;
63 }
64 if XNN_UNLIKELY(n != 0) {
65 {
66 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
67
68
69 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
70 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
71
72 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
73
74 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
75
76 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
77 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
78
79 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
80 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
81
82 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
83
84 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
85 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
86 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
87
88 if (n & (4 * sizeof(int8_t))) {
89 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
90 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
91 output += 4;
92 }
93 if (n & (2 * sizeof(int8_t))) {
94 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
95 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
96 output += 2;
97 }
98 if (n & (1 * sizeof(int8_t))) {
99 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
100 }
101 }
102 }
103 }
104