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 #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_qu8_vaddc_minmax_ukernel__xop_mul32_ld32_x16(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qu8_vaddc_minmax_ukernel__xop_mul32_ld32_x16(
25 size_t n,
26 const uint8_t* input_a,
27 const uint8_t* input_b,
28 uint8_t* output,
29 const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4.a_multiplier);
32 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse4.shift);
33 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
34 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
35 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4.output_max);
36
37 __m128i vbias = _mm_cvtsi32_si128(params->sse4.b_multiplier[0] * (int32_t) *input_b);
38 vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
39 vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4.bias));
40 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
41 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
42 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
43 const __m128i va89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 8)));
44 const __m128i vaCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 12)));
45 input_a += 16;
46 input_b += 16;
47
48 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
49 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
50 __m128i vacc89AB = _mm_macc_epi32(va89AB, va_multiplier, vbias);
51 __m128i vaccCDEF = _mm_macc_epi32(vaCDEF, va_multiplier, vbias);
52
53 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
54 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
55 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
56 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
57
58 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
59 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
60
61 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
62
63 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
64
65 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
66
67 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
68 output += 16;
69 }
70 if XNN_UNLIKELY(n != 0) {
71 do {
72 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
73 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
74 input_a += 8;
75
76 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
77 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
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_packus_epi16(vout01234567, vout01234567);
85 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
86 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
87
88 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
89 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
90 output += 8;
91 n -= 8 * sizeof(uint8_t);
92 } else {
93 if (n & (4 * sizeof(uint8_t))) {
94 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
95 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
96 output += 4;
97 }
98 if (n & (2 * sizeof(uint8_t))) {
99 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
100 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
101 output += 2;
102 }
103 if (n & (1 * sizeof(uint8_t))) {
104 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
105 }
106 n = 0;
107 }
108 } while (n != 0);
109 }
110 }
111