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 <smmintrin.h>
13
14 #include <xnnpack/unaligned.h>
15 #include <xnnpack/vadd.h>
16
17
xnn_qs8_vaddc_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_vaddc_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_add_epi32(
26 _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse4_mul16.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)),
27 _mm_load_si128((const __m128i*) params->sse4_mul16.bias));
28 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
29 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_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 input_a += 8;
38
39
40 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
41 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
42
43 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
44
45 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
46
47 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
48 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
49
50 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
51 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
52
53 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
54
55
56 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
57
58 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
59
60 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
61
62 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
63 output += 8;
64 }
65 if XNN_UNLIKELY(n != 0) {
66 {
67 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
68
69
70 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
71 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
72
73 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
74
75 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
76
77 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
78 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
79
80 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
81 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
82
83 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
84
85 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
86 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
87 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
88
89 if (n & (4 * sizeof(int8_t))) {
90 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
91 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
92 output += 4;
93 }
94 if (n & (2 * sizeof(int8_t))) {
95 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
96 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
97 output += 2;
98 }
99 if (n & (1 * sizeof(int8_t))) {
100 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
101 }
102 }
103 }
104 }
105