xref: /aosp_15_r20/external/XNNPACK/src/f16-vclamp/gen/vclamp-f16c-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-vclamp/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/vunary.h>
17 
18 
xnn_f16_vclamp_ukernel__f16c_x16(size_t n,const void * restrict x_ptr,void * restrict y_ptr,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f16_vclamp_ukernel__f16c_x16(
20     size_t n,
21     const void* restrict x_ptr,
22     void* restrict y_ptr,
23     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25   assert(n != 0);
26   assert(n % sizeof(uint16_t) == 0);
27   assert(x_ptr != NULL);
28   assert(y_ptr != NULL);
29 
30   const uint16_t* x = (const uint16_t*) x_ptr;
31   uint16_t* y = (uint16_t*) y_ptr;
32 
33   const __m256 vy_min = _mm256_load_ps(params->avx.min);
34   const __m256 vy_max = _mm256_load_ps(params->avx.max);
35 
36   for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
37     __m256 vacc01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x));
38     __m256 vacc89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (x + 8)));
39     x += 16;
40 
41     vacc01234567 = _mm256_max_ps(vacc01234567, vy_min);
42     vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEF, vy_min);
43 
44     vacc01234567 = _mm256_min_ps(vacc01234567, vy_max);
45     vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vy_max);
46 
47     _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
48     _mm_storeu_si128((__m128i*) (y + 8), _mm256_cvtps_ph(vacc89ABCDEF, _MM_FROUND_NO_EXC));
49     y += 16;
50   }
51   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
52     __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x));
53     x += 8;
54     vacc = _mm256_max_ps(vacc, vy_min);
55     vacc = _mm256_min_ps(vacc, vy_max);
56     _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC));
57     y += 8;
58   }
59   if XNN_UNLIKELY(n != 0) {
60     __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x));
61     vacc = _mm256_max_ps(vacc, vy_min);
62     vacc = _mm256_min_ps(vacc, vy_max);
63 
64     __m128i vh = _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC);
65     if (n & (4 * sizeof(uint16_t))) {
66       _mm_storel_epi64((__m128i*) y, vh);
67       vh = _mm_unpackhi_epi64(vh, vh);
68       y += 4;
69     }
70     if (n & (2 * sizeof(uint16_t))) {
71       _mm_storeu_si32(y, vh);
72       vh = _mm_srli_epi64(vh, 32);
73       y += 2;
74     }
75     if (n & (1 * sizeof(uint16_t))) {
76       *y = _mm_extract_epi16(vh, 0);
77     }
78   }
79 }
80