• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <emmintrin.h>
13 
14 #include <xnnpack/vaddsub.h>
15 
16 
xnn_qs8_vadd_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_vadd_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_load_si128((const __m128i*) params->sse2.bias);
25   const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
26   const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
27   const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_lo);
28   const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.b_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 vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
37     __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
38     __m128i vb89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_b + 8));
39     input_a += 16;
40     input_b += 16;
41 
42     va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
43     vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
44     va89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(va89ABCDEF, va89ABCDEF), 8);
45     vb89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vb89ABCDEF, vb89ABCDEF), 8);
46 
47     __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
48     __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
49     const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
50     const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
51     __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
52     __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
53     const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
54     const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
55 
56     vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
57     vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
58     vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
59     vbprod89ABCDEFhi = _mm_add_epi16(vbprod89ABCDEFhi, _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_hi));
60 
61     vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
62     vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
63     vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
64     vbprod89ABCDEFhi = _mm_sub_epi16(vbprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vb89ABCDEF, 15), vb_multiplier_lo));
65 
66     __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
67     __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
68     __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
69     __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
70 
71     vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
72     vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
73     vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
74     vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
75 
76     vacc0123 = _mm_sra_epi32(vacc0123, vshift);
77     vacc4567 = _mm_sra_epi32(vacc4567, vshift);
78     vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
79     vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
80 
81     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
82     __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
83 
84     vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
85     vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
86 
87     vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
88     vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
89 
90     __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
91 
92 
93     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
94     output += 16;
95   }
96   if XNN_UNLIKELY(n != 0) {
97     do {
98       __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
99       __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
100       input_a += 8;
101       input_b += 8;
102 
103       va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
104       vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
105 
106       __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
107       __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
108       const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
109       const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
110 
111       vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
112       vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
113 
114       vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
115       vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
116 
117       __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
118       __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
119 
120       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
121       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
122 
123       vacc0123 = _mm_sra_epi32(vacc0123, vshift);
124       vacc4567 = _mm_sra_epi32(vacc4567, vshift);
125 
126       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
127       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
128       vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
129 
130       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
131 
132       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
133         _mm_storel_epi64((__m128i*) output, vout0123456701234567);
134         output += 8;
135         n -= 8 * sizeof(int8_t);
136       } else {
137         if (n & (4 * sizeof(int8_t))) {
138           *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
139           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
140           output += 4;
141         }
142         if (n & (2 * sizeof(int8_t))) {
143           *((uint16_t*) output) = (uint16_t) _mm_cvtsi128_si32(vout0123456701234567);
144           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
145           output += 2;
146         }
147         if (n & (1 * sizeof(int8_t))) {
148           *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
149         }
150         n = 0;
151       }
152     } while (n != 0);
153   }
154 }
155