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_qu8_vadd_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_vadd_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 vbias = _mm_load_si128((const __m128i*) params->sse4.bias);
32 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4.a_multiplier);
33 const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse4.b_multiplier);
34 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse4.shift);
35 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
36 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
37 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4.output_max);
38
39 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
40 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
41 const __m128i vb0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
42 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
43 const __m128i vb4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
44 const __m128i va89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 8)));
45 const __m128i vb89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 8)));
46 const __m128i vaCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 12)));
47 const __m128i vbCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 12)));
48 input_a += 16;
49 input_b += 16;
50
51 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
52 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
53 __m128i vacc89AB = _mm_macc_epi32(va89AB, va_multiplier, vbias);
54 __m128i vaccCDEF = _mm_macc_epi32(vaCDEF, va_multiplier, vbias);
55
56 vacc0123 = _mm_macc_epi32(vb0123, vb_multiplier, vacc0123);
57 vacc4567 = _mm_macc_epi32(vb4567, vb_multiplier, vacc4567);
58 vacc89AB = _mm_macc_epi32(vb89AB, vb_multiplier, vacc89AB);
59 vaccCDEF = _mm_macc_epi32(vbCDEF, vb_multiplier, vaccCDEF);
60
61 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
62 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
63 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
64 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
65
66 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
67 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
68
69 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
70
71 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
72
73 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
74
75 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
76 output += 16;
77 }
78 if XNN_UNLIKELY(n != 0) {
79 do {
80 const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
81 const __m128i vb0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
82 const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
83 const __m128i vb4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
84 input_a += 8;
85 input_b += 8;
86
87 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
88 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
89
90 vacc0123 = _mm_macc_epi32(vb0123, vb_multiplier, vacc0123);
91 vacc4567 = _mm_macc_epi32(vb4567, vb_multiplier, vacc4567);
92
93 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
94 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
95
96 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
97
98 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
99 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
100 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
101
102 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
103 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
104 output += 8;
105 n -= 8 * sizeof(uint8_t);
106 } else {
107 if (n & (4 * sizeof(uint8_t))) {
108 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
109 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
110 output += 4;
111 }
112 if (n & (2 * sizeof(uint8_t))) {
113 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
114 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
115 output += 2;
116 }
117 if (n & (1 * sizeof(uint8_t))) {
118 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
119 }
120 n = 0;
121 }
122 } while (n != 0);
123 }
124 }
125