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/unaligned.h>
15 #include <xnnpack/vadd.h>
16
17
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_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x8(
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_load_si128((const __m128i*) params->sse4_mul16.bias);
26 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
27 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
28 const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.b_multiplier_lo);
29 const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.b_multiplier_hi);
30 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
33 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
34
35 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
36 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
37 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
38 input_a += 8;
39 input_b += 8;
40
41
42 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
43 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
44 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
45 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
46
47 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
48 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
49
50 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
51 vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_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
56 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
57 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
58
59 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
60 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
61
62 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
63
64
65 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
66
67 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
68
69 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
70
71 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
72 output += 8;
73 }
74 if XNN_UNLIKELY(n != 0) {
75 {
76 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
77 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
78
79
80 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
81 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
82 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
83 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
84
85 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
86 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
87
88 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
89 vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
90
91 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
92 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
93
94 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
95 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
96
97 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
98 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
99
100 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
101
102 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
103 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
104 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
105
106 if (n & (4 * sizeof(int8_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(int8_t))) {
112 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
113 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
114 output += 2;
115 }
116 if (n & (1 * sizeof(int8_t))) {
117 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
118 }
119 }
120 }
121 }
122