xref: /aosp_15_r20/external/XNNPACK/src/qs8-vadd/gen/minmax-sse2-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 <emmintrin.h>
13 
14 #include <xnnpack/unaligned.h>
15 #include <xnnpack/vadd.h>
16 
17 
xnn_qs8_vadd_minmax_ukernel__sse2_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__sse2_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->sse2.bias);
26   const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
27   const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
28   const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_lo);
29   const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.b_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 >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
36     __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
37     __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
38     __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
39     __m128i vb89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_b + 8));
40     __m128i vaGHIJKLMN = _mm_loadl_epi64((const __m128i*) (input_a + 16));
41     __m128i vbGHIJKLMN = _mm_loadl_epi64((const __m128i*) (input_b + 16));
42     __m128i vaOPQRSTUV = _mm_loadl_epi64((const __m128i*) (input_a + 24));
43     __m128i vbOPQRSTUV = _mm_loadl_epi64((const __m128i*) (input_b + 24));
44     input_a += 32;
45     input_b += 32;
46 
47     va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
48     vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
49     va89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(va89ABCDEF, va89ABCDEF), 8);
50     vb89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vb89ABCDEF, vb89ABCDEF), 8);
51     vaGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vaGHIJKLMN, vaGHIJKLMN), 8);
52     vbGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vbGHIJKLMN, vbGHIJKLMN), 8);
53     vaOPQRSTUV = _mm_srai_epi16(_mm_unpacklo_epi8(vaOPQRSTUV, vaOPQRSTUV), 8);
54     vbOPQRSTUV = _mm_srai_epi16(_mm_unpacklo_epi8(vbOPQRSTUV, vbOPQRSTUV), 8);
55 
56     __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
57     __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
58     const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
59     const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
60     __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
61     __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
62     const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
63     const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
64     __m128i vaprodGHIJKLMNhi = _mm_mulhi_epu16(vaGHIJKLMN, va_multiplier_lo);
65     __m128i vbprodGHIJKLMNhi = _mm_mulhi_epu16(vbGHIJKLMN, vb_multiplier_lo);
66     const __m128i vaprodGHIJKLMNlo = _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_lo);
67     const __m128i vbprodGHIJKLMNlo = _mm_mullo_epi16(vbGHIJKLMN, vb_multiplier_lo);
68     __m128i vaprodOPQRSTUVhi = _mm_mulhi_epu16(vaOPQRSTUV, va_multiplier_lo);
69     __m128i vbprodOPQRSTUVhi = _mm_mulhi_epu16(vbOPQRSTUV, vb_multiplier_lo);
70     const __m128i vaprodOPQRSTUVlo = _mm_mullo_epi16(vaOPQRSTUV, va_multiplier_lo);
71     const __m128i vbprodOPQRSTUVlo = _mm_mullo_epi16(vbOPQRSTUV, vb_multiplier_lo);
72 
73     vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
74     vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
75     vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
76     vbprod89ABCDEFhi = _mm_add_epi16(vbprod89ABCDEFhi, _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_hi));
77     vaprodGHIJKLMNhi = _mm_add_epi16(vaprodGHIJKLMNhi, _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_hi));
78     vbprodGHIJKLMNhi = _mm_add_epi16(vbprodGHIJKLMNhi, _mm_mullo_epi16(vbGHIJKLMN, vb_multiplier_hi));
79     vaprodOPQRSTUVhi = _mm_add_epi16(vaprodOPQRSTUVhi, _mm_mullo_epi16(vaOPQRSTUV, va_multiplier_hi));
80     vbprodOPQRSTUVhi = _mm_add_epi16(vbprodOPQRSTUVhi, _mm_mullo_epi16(vbOPQRSTUV, vb_multiplier_hi));
81 
82     vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
83     vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
84     vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
85     vbprod89ABCDEFhi = _mm_sub_epi16(vbprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vb89ABCDEF, 15), vb_multiplier_lo));
86     vaprodGHIJKLMNhi = _mm_sub_epi16(vaprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vaGHIJKLMN, 15), va_multiplier_lo));
87     vbprodGHIJKLMNhi = _mm_sub_epi16(vbprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vbGHIJKLMN, 15), vb_multiplier_lo));
88     vaprodOPQRSTUVhi = _mm_sub_epi16(vaprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vaOPQRSTUV, 15), va_multiplier_lo));
89     vbprodOPQRSTUVhi = _mm_sub_epi16(vbprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vbOPQRSTUV, 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     __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
94     __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
95     __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
96     __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
97     __m128i vaccOPQR = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
98     __m128i vaccSTUV = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
99 
100     vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
101     vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
102     vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
103     vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
104     vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
105     vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
106     vaccOPQR = _mm_add_epi32(vaccOPQR, _mm_unpacklo_epi16(vbprodOPQRSTUVlo, vbprodOPQRSTUVhi));
107     vaccSTUV = _mm_add_epi32(vaccSTUV, _mm_unpackhi_epi16(vbprodOPQRSTUVlo, vbprodOPQRSTUVhi));
108 
109     vacc0123 = _mm_sra_epi32(vacc0123, vshift);
110     vacc4567 = _mm_sra_epi32(vacc4567, vshift);
111     vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
112     vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
113     vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
114     vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
115     vaccOPQR = _mm_sra_epi32(vaccOPQR, vshift);
116     vaccSTUV = _mm_sra_epi32(vaccSTUV, vshift);
117 
118     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
119     __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
120     __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
121     __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
122 
123     vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
124     vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
125     voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
126     voutOPQRSTUV = _mm_max_epi16(voutOPQRSTUV, voutput_min);
127 
128     vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
129     vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
130     voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
131     voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
132 
133     __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
134     __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
135 
136 
137     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
138     _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
139     output += 32;
140   }
141   if XNN_UNLIKELY(n != 0) {
142     do {
143       __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
144       __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
145       input_a += 8;
146       input_b += 8;
147 
148       va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
149       vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
150 
151       __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
152       __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
153       const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
154       const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
155 
156       vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
157       vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
158 
159       vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
160       vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
161 
162       __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
163       __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
164 
165       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
166       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
167 
168       vacc0123 = _mm_sra_epi32(vacc0123, vshift);
169       vacc4567 = _mm_sra_epi32(vacc4567, vshift);
170 
171       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
172       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
173       vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
174 
175       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
176 
177       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
178         _mm_storel_epi64((__m128i*) output, vout0123456701234567);
179         output += 8;
180         n -= 8 * sizeof(int8_t);
181       } else {
182         if (n & (4 * sizeof(int8_t))) {
183           unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
184           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
185           output += 4;
186         }
187         if (n & (2 * sizeof(int8_t))) {
188           unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
189           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
190           output += 2;
191         }
192         if (n & (1 * sizeof(int8_t))) {
193           *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
194         }
195         n = 0;
196       }
197     } while (n != 0);
198   }
199 }
200