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 <emmintrin.h>
13
14 #include <xnnpack/vaddsub.h>
15
16
xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x16(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__sse2_mul16_ld64_x16(
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->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(int8_t); n -= 16 * sizeof(int8_t)) {
35 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
36 __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
37 input_a += 16;
38
39 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
40 va89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(va89ABCDEF, va89ABCDEF), 8);
41
42 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
43 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
44 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
45 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
46
47 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
48 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
49
50 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
51 vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
52
53 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
54 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
55 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
56 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
57
58 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
59 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
60 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
61 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
62
63 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
64 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
65
66 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
67 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
68
69 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
70 vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
71
72 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
73
74
75 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
76 output += 16;
77 }
78 if XNN_UNLIKELY(n != 0) {
79 do {
80 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
81 input_a += 8;
82
83 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
84
85 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
86 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
87
88 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
89
90 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
91
92 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
93 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
94
95 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
96 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
97
98 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
99 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
100 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
101
102 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
103
104 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
105 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
106 output += 8;
107 n -= 8 * sizeof(int8_t);
108 } else {
109 if (n & (4 * sizeof(int8_t))) {
110 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
111 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
112 output += 4;
113 }
114 if (n & (2 * sizeof(int8_t))) {
115 *((uint16_t*) output) = (uint16_t) _mm_cvtsi128_si32(vout0123456701234567);
116 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
117 output += 2;
118 }
119 if (n & (1 * sizeof(int8_t))) {
120 *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
121 }
122 n = 0;
123 }
124 } while (n != 0);
125 }
126 }
127