Debian Codegraph @

src/qu8-vadd/gen/qu8-vadd-minmax-sse41-mul32-ld32-u8.c

raw
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 #include <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/unaligned.h"
17 #include "xnnpack/vbinary.h"
18
19
20 void xnn_qu8_vadd_minmax_ukernel__sse41_mul32_ld32_u8(
21 size_t batch,
22 const uint8_t* input_a,
23 const uint8_t* input_b,
24 uint8_t* output,
25 const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
26 {
27 assert(batch != 0);
28 assert(batch % sizeof(uint8_t) == 0);
29 assert(input_a != NULL);
30 assert(input_b != NULL);
31 assert(output != NULL);
32
33 const __m128i vbias = _mm_set1_epi32(params->scalar.bias);
34 const __m128i va_multiplier = _mm_set1_epi32(params->scalar.a_multiplier);
35 const __m128i vb_multiplier = _mm_set1_epi32(params->scalar.b_multiplier);
36 const __m128i vshift = _mm_cvtsi32_si128((int) params->scalar.shift);
37 const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point);
38 const __m128i voutput_min = _mm_set1_epi8(params->scalar.output_min);
39 const __m128i voutput_max = _mm_set1_epi8(params->scalar.output_max);
40
41 XNN_FORCE_REALIZATION(vbias);
42 XNN_FORCE_REALIZATION(va_multiplier);
43 XNN_FORCE_REALIZATION(vb_multiplier);
44 XNN_FORCE_REALIZATION(vshift);
45 XNN_FORCE_REALIZATION(voutput_zero_point);
46 XNN_FORCE_REALIZATION(voutput_min);
47 XNN_FORCE_REALIZATION(voutput_max);
48
49 for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) {
50 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
51 const __m128i vb0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
52 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
53 const __m128i vb4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
54 input_a += 8;
55 input_b += 8;
56
57 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
58 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
59
60 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
61 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
62
63 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
64 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
65
66 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
67
68 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
69
70 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
71
72 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
73
74 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
75 output += 8;
76 }
77 if XNN_UNLIKELY(batch != 0) {
78 {
79 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
80 const __m128i vb0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
81 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
82 const __m128i vb4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
83
84 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
85 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
86
87 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
88 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
89
90 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
91 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
92
93 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
94
95 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
96 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
97 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
98
99 if (batch & (4 * sizeof(uint8_t))) {
100 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
101 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
102 output += 4;
103 }
104 if (batch & (2 * sizeof(uint8_t))) {
105 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
106 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
107 output += 2;
108 }
109 if (batch & (1 * sizeof(uint8_t))) {
110 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
111 }
112 }
113 }
114 }