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/intrinsics-polyfill.h>
15 #include <xnnpack/unaligned.h>
16 #include <xnnpack/vadd.h>
17
18
xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16(
20 size_t n,
21 const int8_t* input_a,
22 const int8_t* input_b,
23 int8_t* output,
24 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 const __m128i vbias = _mm_load_si128((const __m128i*) params->sse4_mul32.bias);
27 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
28 const __m128i vb_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.b_multiplier);
29 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse4_mul32.shift);
30 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
31 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
32 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
33
34 for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
35 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
36 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
37 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
38 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
39 const __m128i va89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 8)));
40 const __m128i vb89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 8)));
41 const __m128i vaCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 12)));
42 const __m128i vbCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 12)));
43 input_a += 16;
44 input_b += 16;
45
46 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
47 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
48 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_mullo_epi32(va89AB, va_multiplier));
49 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_mullo_epi32(vaCDEF, va_multiplier));
50
51 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
52 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
53 vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vb89AB, vb_multiplier));
54 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vbCDEF, vb_multiplier));
55
56 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
57 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
58 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
59 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
60
61 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
62 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
63
64 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
65
66 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
67
68 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
69
70 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
71 output += 16;
72 }
73 if XNN_UNLIKELY(n != 0) {
74 do {
75 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a)));
76 const __m128i vb0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b)));
77 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4)));
78 const __m128i vb4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4)));
79 input_a += 8;
80 input_b += 8;
81
82 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier));
83 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier));
84
85 vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier));
86 vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier));
87
88 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
89 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
90
91 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
92
93 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
94 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
95 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
96
97 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
98 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
99 output += 8;
100 n -= 8 * sizeof(int8_t);
101 } else {
102 if (n & (4 * sizeof(int8_t))) {
103 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
104 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
105 output += 4;
106 }
107 if (n & (2 * sizeof(int8_t))) {
108 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
109 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
110 output += 2;
111 }
112 if (n & (1 * sizeof(int8_t))) {
113 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
114 }
115 n = 0;
116 }
117 } while (n != 0);
118 }
119 }
120