1 // Auto-generated file. Do not edit!
2 // Template: src/f16-prelu/f16c.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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/math.h>
15 #include <xnnpack/prelu.h>
16
17
xnn_f16_prelu_ukernel__f16c_2x8(size_t rows,size_t channels,const void * restrict input,size_t input_stride,const void * restrict weights,void * restrict output,size_t output_stride)18 void xnn_f16_prelu_ukernel__f16c_2x8(
19 size_t rows,
20 size_t channels,
21 const void* restrict input,
22 size_t input_stride,
23 const void* restrict weights,
24 void* restrict output,
25 size_t output_stride) XNN_OOB_READS
26 {
27 assert(rows != 0);
28 assert(channels != 0);
29 assert(channels % sizeof(uint16_t) == 0);
30
31 const uint16_t* i0 = (const uint16_t*) input;
32 uint16_t* o0 = (uint16_t*) output;
33 const uint16_t* i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
34 uint16_t* o1 = (uint16_t*) ((uintptr_t) o0 + output_stride);
35
36 const size_t input_increment = input_stride * 2 - channels;
37 const size_t output_increment = output_stride * 2 - channels;
38
39 do {
40 if XNN_UNPREDICTABLE(rows < 2) {
41 i1 = i0;
42 o1 = o0;
43 }
44
45 const uint16_t* w = (const uint16_t*) weights;
46 size_t c = channels;
47 for (; c >= 8 * sizeof(uint16_t); c -= 8 * sizeof(uint16_t)) {
48 const __m256 vw01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) w));
49 w += 8;
50
51 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
52 i0 += 8;
53 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
54 i1 += 8;
55
56 __m256 vacc0x01234567 = _mm256_mul_ps(vi0x01234567, vw01234567);
57 __m256 vacc1x01234567 = _mm256_mul_ps(vi1x01234567, vw01234567);
58
59 vacc0x01234567 = _mm256_blendv_ps(vi0x01234567, vacc0x01234567, vi0x01234567);
60 vacc1x01234567 = _mm256_blendv_ps(vi1x01234567, vacc1x01234567, vi1x01234567);
61
62 _mm_storeu_si128((__m128i*) o0, _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC));
63 o0 += 8;
64 _mm_storeu_si128((__m128i*) o1, _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC));
65 o1 += 8;
66 }
67 if XNN_UNLIKELY(c != 0) {
68 const __m256 vw01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) w));
69
70 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
71 i0 = (const uint16_t*) ((uintptr_t) i0 + c);
72 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
73 i1 = (const uint16_t*) ((uintptr_t) i1 + c);
74
75 __m256 vacc0x01234567 = _mm256_mul_ps(vi0x01234567, vw01234567);
76 __m256 vacc1x01234567 = _mm256_mul_ps(vi1x01234567, vw01234567);
77
78 vacc0x01234567 = _mm256_blendv_ps(vi0x01234567, vacc0x01234567, vi0x01234567);
79 vacc1x01234567 = _mm256_blendv_ps(vi1x01234567, vacc1x01234567, vi1x01234567);
80
81 __m128i vh0x01234567 = _mm256_cvtps_ph(vacc0x01234567, _MM_FROUND_NO_EXC);
82 __m128i vh1x01234567 = _mm256_cvtps_ph(vacc1x01234567, _MM_FROUND_NO_EXC);
83 if (c & (4 * sizeof(uint16_t))) {
84 _mm_storel_epi64((__m128i*) o0, vh0x01234567);
85 _mm_storel_epi64((__m128i*) o1, vh1x01234567);
86
87 vh0x01234567 = _mm_unpackhi_epi64(vh0x01234567, vh0x01234567);
88 vh1x01234567 = _mm_unpackhi_epi64(vh1x01234567, vh1x01234567);
89
90 o0 += 4;
91 o1 += 4;
92 }
93 if (c & (2 * sizeof(uint16_t))) {
94 *((uint32_t*) o0) = (uint32_t) _mm_cvtsi128_si32(vh0x01234567);
95 *((uint32_t*) o1) = (uint32_t) _mm_cvtsi128_si32(vh1x01234567);
96
97 vh0x01234567 = _mm_srli_epi64(vh0x01234567, 32);
98 vh1x01234567 = _mm_srli_epi64(vh1x01234567, 32);
99
100 o0 += 2;
101 o1 += 2;
102 }
103 if (c & (1 * sizeof(uint16_t))) {
104 *o0 = (uint16_t) _mm_extract_epi16(vh0x01234567, 0);
105 *o1 = (uint16_t) _mm_extract_epi16(vh1x01234567, 0);
106
107 o0 += 1;
108 o1 += 1;
109 }
110 }
111 i0 = (const uint16_t*) ((uintptr_t) i0 + input_increment);
112 o0 = (uint16_t*) ((uintptr_t) o0 + output_increment);
113 i1 = (const uint16_t*) ((uintptr_t) i1 + input_increment);
114 o1 = (uint16_t*) ((uintptr_t) o1 + output_increment);
115 rows = doz(rows, 2);
116 } while (rows != 0);
117 }
118