1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/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_vadd_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_vadd_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_load_si128((const __m128i*) params->sse4_mul16.bias);
25 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
26 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
27 const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.b_multiplier_lo);
28 const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.b_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 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
37 input_a += 8;
38 input_b += 8;
39
40
41 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
42 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
43 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
44 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
45
46 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
47 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
48
49 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
50 vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
51
52 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
53 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
54
55 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
56 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
57
58 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
59 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
60
61 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
62
63
64 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
65
66 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
67
68 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
69
70 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
71 output += 8;
72 }
73 if XNN_UNLIKELY(n != 0) {
74 {
75 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
76 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
77
78
79 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
80 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
81 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
82 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
83
84 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
85 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
86
87 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
88 vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
89
90 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
91 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
92
93 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
94 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
95
96 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
97 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
98
99 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
100
101 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
102 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
103 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
104
105 if (n & (4 * sizeof(int8_t))) {
106 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
107 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
108 output += 4;
109 }
110 if (n & (2 * sizeof(int8_t))) {
111 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
112 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
113 output += 2;
114 }
115 if (n & (1 * sizeof(int8_t))) {
116 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
117 }
118 }
119 }
120 }
121