1 // Auto-generated file. Do not edit!
2 // Template: src/s8-ibilinear/sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/ibilinear.h>
16
17
xnn_s8_ibilinear_ukernel__sse2_c8(size_t output_pixels,size_t channels,const int8_t ** restrict input,size_t input_offset,const int16_t * restrict weights,int8_t * restrict output,size_t output_increment)18 void xnn_s8_ibilinear_ukernel__sse2_c8(
19 size_t output_pixels,
20 size_t channels,
21 const int8_t**restrict input,
22 size_t input_offset,
23 const int16_t*restrict weights,
24 int8_t*restrict output,
25 size_t output_increment) XNN_OOB_READS
26 {
27 assert(output_pixels != 0);
28 assert(channels != 0);
29
30 do {
31 const int8_t* i0 = (const int8_t*) ((uintptr_t) input[0] + input_offset);
32 const int8_t* i1 = (const int8_t*) ((uintptr_t) input[1] + input_offset);
33 const int8_t* i2 = (const int8_t*) ((uintptr_t) input[2] + input_offset);
34 const int8_t* i3 = (const int8_t*) ((uintptr_t) input[3] + input_offset);
35 input += 4;
36
37 const __m128i valpha = _mm_cvtsi32_si128(*((const int*) weights));
38 weights += 2;
39 __m128i valphah = _mm_shufflelo_epi16(valpha, _MM_SHUFFLE(0, 0, 0, 0));
40 valphah = _mm_unpacklo_epi64(valphah, valphah);
41 __m128i valphav = _mm_shufflelo_epi16(valpha, _MM_SHUFFLE(1, 1, 1, 1));
42 valphav = _mm_unpacklo_epi64(valphav, valphav);
43
44 valphah = _mm_xor_si128(valphah, _mm_set1_epi32(0xFFFF0000));
45 valphah = _mm_add_epi16(valphah, _mm_set1_epi32(0x08010000));
46
47 const __m128i vrounding = _mm_set1_epi32(0x00200000);
48
49 size_t c = channels;
50 for (; c >= 8 * sizeof(int8_t); c -= 8 * sizeof(int8_t)) {
51 __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0);
52 i0 += 8;
53 __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1);
54 i1 += 8;
55 __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2);
56 i2 += 8;
57 __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3);
58 i3 += 8;
59
60 vtl01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vtl01234567, vtl01234567), 8);
61 vtr01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vtr01234567, vtr01234567), 8);
62 vbl01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vbl01234567, vbl01234567), 8);
63 vbr01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vbr01234567, vbr01234567), 8);
64
65 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
66 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
67 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
68 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
69
70 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
71 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
72
73 __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16);
74 __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16);
75
76 vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123);
77 vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567);
78
79 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
80 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
81
82 vacc0123 = _mm_srai_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
83 vacc4567 = _mm_srai_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
84
85 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
86
87 const __m128i vo01234567 = _mm_packs_epi16(vacc01234567, vacc01234567);
88
89 _mm_storel_epi64((__m128i*) output, vo01234567);
90 output += 8;
91 }
92 if XNN_UNLIKELY(c != 0) {
93 __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0);
94 __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1);
95 __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2);
96 __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3);
97
98 vtl01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vtl01234567, vtl01234567), 8);
99 vtr01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vtr01234567, vtr01234567), 8);
100 vbl01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vbl01234567, vbl01234567), 8);
101 vbr01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vbr01234567, vbr01234567), 8);
102
103 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
104 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
105 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
106 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
107
108 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
109 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
110
111 __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16);
112 __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16);
113
114 vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123);
115 vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567);
116
117 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
118 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
119
120 vacc0123 = _mm_srai_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
121 vacc4567 = _mm_srai_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
122
123 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
124
125 __m128i vo01234567 = _mm_packs_epi16(vacc01234567, vacc01234567);
126
127 if (c & (4 * sizeof(int8_t))) {
128 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vo01234567);
129 output += 4;
130 vo01234567 = _mm_srli_epi64(vo01234567, 32);
131 }
132 uint32_t vo0123 = (uint32_t) _mm_cvtsi128_si32(vo01234567);
133 if (c & (2 * sizeof(int8_t))) {
134 *((uint16_t*) output) = (uint16_t) vo0123;
135 output += 2;
136 vo0123 >>= 16;
137 }
138 if (c & (1 * sizeof(int8_t))) {
139 *output++ = (uint8_t) vo0123;
140 }
141 }
142
143 output = (int8_t*) ((uintptr_t) output + output_increment);
144 } while (--output_pixels != 0);
145 }
146