xref: /aosp_15_r20/external/XNNPACK/src/qs8-vadd/gen/minmax-avx-mul16-ld64-x32.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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__avx_mul16_ld64_x32(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__avx_mul16_ld64_x32(
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 >= 32 * sizeof(int8_t); n -= 32 * 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     const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
39     const __m128i vb89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_b + 8)));
40     const __m128i vaGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 16)));
41     const __m128i vbGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_b + 16)));
42     const __m128i vaOPQRSTUV = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 24)));
43     const __m128i vbOPQRSTUV = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_b + 24)));
44     input_a += 32;
45     input_b += 32;
46 
47 
48     __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
49     __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
50     const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
51     const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
52     __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
53     __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
54     const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
55     const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
56     __m128i vaprodGHIJKLMNhi = _mm_mulhi_epu16(vaGHIJKLMN, va_multiplier_lo);
57     __m128i vbprodGHIJKLMNhi = _mm_mulhi_epu16(vbGHIJKLMN, vb_multiplier_lo);
58     const __m128i vaprodGHIJKLMNlo = _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_lo);
59     const __m128i vbprodGHIJKLMNlo = _mm_mullo_epi16(vbGHIJKLMN, vb_multiplier_lo);
60     __m128i vaprodOPQRSTUVhi = _mm_mulhi_epu16(vaOPQRSTUV, va_multiplier_lo);
61     __m128i vbprodOPQRSTUVhi = _mm_mulhi_epu16(vbOPQRSTUV, vb_multiplier_lo);
62     const __m128i vaprodOPQRSTUVlo = _mm_mullo_epi16(vaOPQRSTUV, va_multiplier_lo);
63     const __m128i vbprodOPQRSTUVlo = _mm_mullo_epi16(vbOPQRSTUV, vb_multiplier_lo);
64 
65     vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
66     vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
67     vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
68     vbprod89ABCDEFhi = _mm_add_epi16(vbprod89ABCDEFhi, _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_hi));
69     vaprodGHIJKLMNhi = _mm_add_epi16(vaprodGHIJKLMNhi, _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_hi));
70     vbprodGHIJKLMNhi = _mm_add_epi16(vbprodGHIJKLMNhi, _mm_mullo_epi16(vbGHIJKLMN, vb_multiplier_hi));
71     vaprodOPQRSTUVhi = _mm_add_epi16(vaprodOPQRSTUVhi, _mm_mullo_epi16(vaOPQRSTUV, va_multiplier_hi));
72     vbprodOPQRSTUVhi = _mm_add_epi16(vbprodOPQRSTUVhi, _mm_mullo_epi16(vbOPQRSTUV, vb_multiplier_hi));
73 
74     vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
75     vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
76     vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
77     vbprod89ABCDEFhi = _mm_sub_epi16(vbprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vb89ABCDEF, 15), vb_multiplier_lo));
78     vaprodGHIJKLMNhi = _mm_sub_epi16(vaprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vaGHIJKLMN, 15), va_multiplier_lo));
79     vbprodGHIJKLMNhi = _mm_sub_epi16(vbprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vbGHIJKLMN, 15), vb_multiplier_lo));
80     vaprodOPQRSTUVhi = _mm_sub_epi16(vaprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vaOPQRSTUV, 15), va_multiplier_lo));
81     vbprodOPQRSTUVhi = _mm_sub_epi16(vbprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vbOPQRSTUV, 15), vb_multiplier_lo));
82 
83     __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
84     __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
85     __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
86     __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
87     __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
88     __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
89     __m128i vaccOPQR = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
90     __m128i vaccSTUV = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
91 
92     vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
93     vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
94     vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
95     vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
96     vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
97     vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
98     vaccOPQR = _mm_add_epi32(vaccOPQR, _mm_unpacklo_epi16(vbprodOPQRSTUVlo, vbprodOPQRSTUVhi));
99     vaccSTUV = _mm_add_epi32(vaccSTUV, _mm_unpackhi_epi16(vbprodOPQRSTUVlo, vbprodOPQRSTUVhi));
100 
101     vacc0123 = _mm_sra_epi32(vacc0123, vshift);
102     vacc4567 = _mm_sra_epi32(vacc4567, vshift);
103     vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
104     vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
105     vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
106     vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
107     vaccOPQR = _mm_sra_epi32(vaccOPQR, vshift);
108     vaccSTUV = _mm_sra_epi32(vaccSTUV, vshift);
109 
110     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
111     __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
112     __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
113     __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
114 
115 
116     __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
117     __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
118 
119     vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
120     voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
121 
122     vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
123     voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
124 
125     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
126     _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
127     output += 32;
128   }
129   if XNN_UNLIKELY(n != 0) {
130     do {
131       const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
132       const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
133       input_a += 8;
134       input_b += 8;
135 
136 
137       __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
138       __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
139       const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
140       const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
141 
142       vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
143       vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
144 
145       vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
146       vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
147 
148       __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
149       __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
150 
151       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
152       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
153 
154       vacc0123 = _mm_sra_epi32(vacc0123, vshift);
155       vacc4567 = _mm_sra_epi32(vacc4567, vshift);
156 
157       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
158 
159       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
160       vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
161       vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
162 
163       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
164         _mm_storel_epi64((__m128i*) output, vout0123456701234567);
165         output += 8;
166         n -= 8 * sizeof(int8_t);
167       } else {
168         if (n & (4 * sizeof(int8_t))) {
169           unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
170           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
171           output += 4;
172         }
173         if (n & (2 * sizeof(int8_t))) {
174           unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
175           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
176           output += 2;
177         }
178         if (n & (1 * sizeof(int8_t))) {
179           *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
180         }
181         n = 0;
182       }
183     } while (n != 0);
184   }
185 }
186