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/unaligned.h>
15 #include <xnnpack/vadd.h>
16
17
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_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x16(
19 size_t n,
20 const int8_t* input_a,
21 const int8_t* input_b,
22 int8_t* output,
23 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 const __m128i vbias = _mm_add_epi32(
26 _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
27 _mm_load_si128((const __m128i*) params->sse2.bias));
28 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
29 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
30 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
33 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
34
35 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
36 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
37 __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
38 input_a += 16;
39
40 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
41 va89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(va89ABCDEF, va89ABCDEF), 8);
42
43 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
44 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
45 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
46 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
47
48 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
49 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
50
51 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
52 vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
53
54 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
55 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
56 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
57 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
58
59 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
60 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
61 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
62 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
63
64 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
65 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
66
67 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
68 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
69
70 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
71 vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
72
73 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
74
75
76 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
77 output += 16;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 do {
81 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
82 input_a += 8;
83
84 va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
85
86 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
87 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
88
89 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
90
91 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
92
93 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
94 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
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 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
101 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
102
103 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
104
105 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
106 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
107 output += 8;
108 n -= 8 * sizeof(int8_t);
109 } else {
110 if (n & (4 * sizeof(int8_t))) {
111 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
112 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
113 output += 4;
114 }
115 if (n & (2 * sizeof(int8_t))) {
116 unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
117 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
118 output += 2;
119 }
120 if (n & (1 * sizeof(int8_t))) {
121 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
122 }
123 n = 0;
124 }
125 } while (n != 0);
126 }
127 }
128