xref: /aosp_15_r20/external/XNNPACK/src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.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_x24(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_x24(
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 >= 24 * sizeof(int8_t); n -= 24 * 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     input_a += 24;
43     input_b += 24;
44 
45     va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
46     vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
47     va89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(va89ABCDEF, va89ABCDEF), 8);
48     vb89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vb89ABCDEF, vb89ABCDEF), 8);
49     vaGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vaGHIJKLMN, vaGHIJKLMN), 8);
50     vbGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vbGHIJKLMN, vbGHIJKLMN), 8);
51 
52     __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
53     __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
54     const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
55     const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
56     __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
57     __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
58     const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
59     const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
60     __m128i vaprodGHIJKLMNhi = _mm_mulhi_epu16(vaGHIJKLMN, va_multiplier_lo);
61     __m128i vbprodGHIJKLMNhi = _mm_mulhi_epu16(vbGHIJKLMN, vb_multiplier_lo);
62     const __m128i vaprodGHIJKLMNlo = _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_lo);
63     const __m128i vbprodGHIJKLMNlo = _mm_mullo_epi16(vbGHIJKLMN, 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 
72     vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
73     vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
74     vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
75     vbprod89ABCDEFhi = _mm_sub_epi16(vbprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vb89ABCDEF, 15), vb_multiplier_lo));
76     vaprodGHIJKLMNhi = _mm_sub_epi16(vaprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vaGHIJKLMN, 15), va_multiplier_lo));
77     vbprodGHIJKLMNhi = _mm_sub_epi16(vbprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vbGHIJKLMN, 15), vb_multiplier_lo));
78 
79     __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
80     __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
81     __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
82     __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
83     __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
84     __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
85 
86     vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
87     vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
88     vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
89     vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
90     vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
91     vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
92 
93     vacc0123 = _mm_sra_epi32(vacc0123, vshift);
94     vacc4567 = _mm_sra_epi32(vacc4567, vshift);
95     vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
96     vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
97     vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
98     vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
99 
100     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
101     __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
102     __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
103 
104     vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
105     vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
106     voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
107 
108     vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
109     vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
110     voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
111 
112     __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
113     __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
114 
115 
116     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
117     _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
118     output += 24;
119   }
120   if XNN_UNLIKELY(n != 0) {
121     do {
122       __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
123       __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
124       input_a += 8;
125       input_b += 8;
126 
127       va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
128       vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
129 
130       __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
131       __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
132       const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
133       const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
134 
135       vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
136       vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
137 
138       vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
139       vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
140 
141       __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
142       __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
143 
144       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
145       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
146 
147       vacc0123 = _mm_sra_epi32(vacc0123, vshift);
148       vacc4567 = _mm_sra_epi32(vacc4567, vshift);
149 
150       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
151       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
152       vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
153 
154       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
155 
156       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
157         _mm_storel_epi64((__m128i*) output, vout0123456701234567);
158         output += 8;
159         n -= 8 * sizeof(int8_t);
160       } else {
161         if (n & (4 * sizeof(int8_t))) {
162           unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
163           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
164           output += 4;
165         }
166         if (n & (2 * sizeof(int8_t))) {
167           unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
168           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
169           output += 2;
170         }
171         if (n & (1 * sizeof(int8_t))) {
172           *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
173         }
174         n = 0;
175       }
176     } while (n != 0);
177   }
178 }
179