1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/sse-mul32-ld32.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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/intrinsics-polyfill.h>
20 #include <xnnpack/unaligned.h>
21 #include <xnnpack/vadd.h>
22
23
xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_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)])24 void xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24(
25 size_t n,
26 const int8_t* input_a,
27 const int8_t* input_b,
28 int8_t* output,
29 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse4_mul32.bias);
32 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
33 const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.b_multiplier);
34 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse4_mul32.shift);
35 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
36 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
37 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
38
39 for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
40 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
41 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
42 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
43 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
44 const __m128i va89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 8)));
45 const __m128i vb89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 8)));
46 const __m128i vaCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 12)));
47 const __m128i vbCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 12)));
48 const __m128i vaGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 16)));
49 const __m128i vbGHIJ = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 16)));
50 const __m128i vaKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 20)));
51 const __m128i vbKLMN = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 20)));
52 input_a += 24;
53 input_b += 24;
54
55 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
56 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
57 __m128i vacc89AB = _mm_macc_epi32(va89AB, va_multiplier, vbias);
58 __m128i vaccCDEF = _mm_macc_epi32(vaCDEF, va_multiplier, vbias);
59 __m128i vaccGHIJ = _mm_macc_epi32(vaGHIJ, va_multiplier, vbias);
60 __m128i vaccKLMN = _mm_macc_epi32(vaKLMN, va_multiplier, vbias);
61
62 vacc0123 = _mm_macc_epi32(vb0123, vb_multiplier, vacc0123);
63 vacc4567 = _mm_macc_epi32(vb4567, vb_multiplier, vacc4567);
64 vacc89AB = _mm_macc_epi32(vb89AB, vb_multiplier, vacc89AB);
65 vaccCDEF = _mm_macc_epi32(vbCDEF, vb_multiplier, vaccCDEF);
66 vaccGHIJ = _mm_macc_epi32(vbGHIJ, vb_multiplier, vaccGHIJ);
67 vaccKLMN = _mm_macc_epi32(vbKLMN, vb_multiplier, vaccKLMN);
68
69 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
70 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
71 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
72 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
73 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
74 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
75
76 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
77 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
78 const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
79
80 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
81 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
82
83 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
84 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
85
86 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
87 voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
88
89 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
90 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
91 output += 24;
92 }
93 if XNN_UNLIKELY(n != 0) {
94 do {
95 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
96 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
97 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
98 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
99 input_a += 8;
100 input_b += 8;
101
102 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
103 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
104
105 vacc0123 = _mm_macc_epi32(vb0123, vb_multiplier, vacc0123);
106 vacc4567 = _mm_macc_epi32(vb4567, vb_multiplier, vacc4567);
107
108 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
109 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
110
111 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
112
113 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
114 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
115 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
116
117 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
118 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
119 output += 8;
120 n -= 8 * sizeof(int8_t);
121 } else {
122 if (n & (4 * sizeof(int8_t))) {
123 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
124 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
125 output += 4;
126 }
127 if (n & (2 * sizeof(int8_t))) {
128 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
129 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
130 output += 2;
131 }
132 if (n & (1 * sizeof(int8_t))) {
133 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
134 }
135 n = 0;
136 }
137 } while (n != 0);
138 }
139 }
140