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/vaddsub.h>
21
22
xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8(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)])23 void xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8(
24 size_t n,
25 const int8_t* input_a,
26 const int8_t* input_b,
27 int8_t* output,
28 const union xnn_qs8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse4_mul32.bias);
31 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
32 const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.b_multiplier);
33 const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
34 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
35 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
36 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
37
38 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
39 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
40 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b));
41 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
42 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 4));
43 input_a += 8;
44 input_b += 8;
45
46 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
47 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
48
49 vacc0123 = _mm_macc_epi32(vb0123, vb_multiplier, vacc0123);
50 vacc4567 = _mm_macc_epi32(vb4567, vb_multiplier, vacc4567);
51
52 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
53 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
54
55 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
56
57 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
58
59 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
60
61 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
62
63 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
64 output += 8;
65 }
66 if XNN_UNLIKELY(n != 0) {
67 {
68 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
69 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b));
70 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
71 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_b + 4));
72
73 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
74 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
75
76 vacc0123 = _mm_macc_epi32(vb0123, vb_multiplier, vacc0123);
77 vacc4567 = _mm_macc_epi32(vb4567, vb_multiplier, vacc4567);
78
79 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
80 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
81
82 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
83
84 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
85 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
86 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
87
88 if (n & (4 * sizeof(int8_t))) {
89 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
90 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
91 output += 4;
92 }
93 if (n & (2 * sizeof(int8_t))) {
94 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
95 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
96 output += 2;
97 }
98 if (n & (1 * sizeof(int8_t))) {
99 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
100 }
101 }
102 }
103 }
104