Debian Codegraph @

src/qu8-vaddc/gen/qu8-vaddc-minmax-sse41-mul32-ld32-u16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/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_vaddc_minmax_ukernel__sse41_mul32_ld32_u16(
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.b_multiplier * (int32_t) *input_b + params->scalar.bias);
34 const __m128i va_multiplier = _mm_set1_epi32(params->scalar.a_multiplier);
35 const __m128i vshift = _mm_cvtsi32_si128((int) params->scalar.shift);
36 const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point);
37 const __m128i voutput_min = _mm_set1_epi8(params->scalar.output_min);
38 const __m128i voutput_max = _mm_set1_epi8(params->scalar.output_max);
39
40 XNN_FORCE_REALIZATION(vbias);
41 XNN_FORCE_REALIZATION(va_multiplier);
42 XNN_FORCE_REALIZATION(vshift);
43 XNN_FORCE_REALIZATION(voutput_zero_point);
44 XNN_FORCE_REALIZATION(voutput_min);
45 XNN_FORCE_REALIZATION(voutput_max);
46
47 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
48 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
49 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
50 const __m128i va89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 8)));
51 const __m128i vaCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 12)));
52 input_a += 16;
53 input_b += 16;
54
55 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
56 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
57 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_mullo_epi32(va89AB, va_multiplier));
58 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_mullo_epi32(vaCDEF, va_multiplier));
59
60 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
61 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
62 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
63 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
64
65 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
66 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
67
68 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
69
70 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
71
72 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
73
74 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
75 output += 16;
76 }
77 if XNN_UNLIKELY(batch != 0) {
78 do {
79 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
80 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
81 input_a += 8;
82
83 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
84 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
85
86 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
87 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
88
89 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
90
91 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
92 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
93 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
94
95 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
96 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
97 output += 8;
98 batch -= 8 * sizeof(uint8_t);
99 } else {
100 if (batch & (4 * sizeof(uint8_t))) {
101 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
102 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
103 output += 4;
104 }
105 if (batch & (2 * sizeof(uint8_t))) {
106 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
107 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
108 output += 2;
109 }
110 if (batch & (1 * sizeof(uint8_t))) {
111 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
112 }
113 batch = 0;
114 }
115 } while (batch != 0);
116 }
117 }