xref: /aosp_15_r20/external/XNNPACK/src/f16-vbinary/gen/vadd-minmax-f16c-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-vbinary/vop-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/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vbinary.h>
17 
18 
xnn_f16_vadd_minmax_ukernel__f16c_x8(size_t n,const void * restrict a_ptr,const void * restrict b_ptr,void * restrict y_ptr,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f16_vadd_minmax_ukernel__f16c_x8(
20     size_t n,
21     const void* restrict a_ptr,
22     const void* restrict b_ptr,
23     void* restrict y_ptr,
24     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26   assert(n != 0);
27   assert(n % sizeof(uint16_t) == 0);
28   assert(a_ptr != NULL);
29   assert(b_ptr != NULL);
30   assert(y_ptr != NULL);
31 
32   const uint16_t* a = (const uint16_t*) a_ptr;
33   const uint16_t* b = (const uint16_t*) b_ptr;
34   uint16_t* y = (uint16_t*) y_ptr;
35 
36   const __m256 vy_min = _mm256_load_ps(params->avx.min);
37   const __m256 vy_max = _mm256_load_ps(params->avx.max);
38 
39   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
40     const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
41     const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
42     a += 8;
43     b += 8;
44 
45     __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(va, vb), _MM_FROUND_NO_EXC));
46 
47     vy = _mm256_max_ps(vy, vy_min);
48     vy = _mm256_min_ps(vy, vy_max);
49 
50     _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC));
51     y += 8;
52   }
53   if XNN_UNLIKELY(n != 0) {
54     const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
55     const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
56 
57     __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(va, vb), _MM_FROUND_NO_EXC));
58 
59     vy = _mm256_max_ps(vy, vy_min);
60     vy = _mm256_min_ps(vy, vy_max);
61 
62     __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC);
63     if (n & (4 * sizeof(uint16_t))) {
64       _mm_storel_epi64((__m128i*) y, vh);
65       vh = _mm_unpackhi_epi64(vh, vh);
66       y += 4;
67     }
68     if (n & (2 * sizeof(uint16_t))) {
69       _mm_storeu_si32(y, vh);
70       vh = _mm_srli_epi64(vh, 32);
71       y += 2;
72     }
73     if (n & (1 * sizeof(uint16_t))) {
74       *y = (uint16_t) _mm_extract_epi16(vh, 0);
75     }
76   }
77 }
78