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_qu8_vaddc_minmax_ukernel__sse2_mul16_ld64_x16(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_vaddc_minmax_ukernel__sse2_mul16_ld64_x16(
19 size_t n,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const union xnn_qu8_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(uint8_t); n -= 16 * sizeof(uint8_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 const __m128i vzero = _mm_setzero_si128();
41 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
42 va89ABCDEF = _mm_unpacklo_epi8(va89ABCDEF, vzero);
43
44 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
45 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
46 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
47 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
48
49 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
50 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
51
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
67 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
68
69 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
70
71 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
72
73 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
74 output += 16;
75 }
76 if XNN_UNLIKELY(n != 0) {
77 do {
78 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
79 input_a += 8;
80
81 va01234567 = _mm_unpacklo_epi8(va01234567, _mm_setzero_si128());
82
83 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
84 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
85
86 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
87
88
89 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
90 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
91
92 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
93 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
94
95 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
96
97 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
98 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
99 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
100
101 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
102 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
103 output += 8;
104 n -= 8 * sizeof(uint8_t);
105 } else {
106 if (n & (4 * sizeof(uint8_t))) {
107 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
108 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
109 output += 4;
110 }
111 if (n & (2 * sizeof(uint8_t))) {
112 unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
113 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
114 output += 2;
115 }
116 if (n & (1 * sizeof(uint8_t))) {
117 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
118 }
119 n = 0;
120 }
121 } while (n != 0);
122 }
123 }
124