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_x8(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_x8(
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 >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
37 __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x));
38 x += 8;
39 vacc = _mm256_max_ps(vacc, vy_min);
40 vacc = _mm256_min_ps(vacc, vy_max);
41 _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC));
42 y += 8;
43 }
44 if XNN_UNLIKELY(n != 0) {
45 __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x));
46 vacc = _mm256_max_ps(vacc, vy_min);
47 vacc = _mm256_min_ps(vacc, vy_max);
48
49 __m128i vh = _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC);
50 if (n & (4 * sizeof(uint16_t))) {
51 _mm_storel_epi64((__m128i*) y, vh);
52 vh = _mm_unpackhi_epi64(vh, vh);
53 y += 4;
54 }
55 if (n & (2 * sizeof(uint16_t))) {
56 _mm_storeu_si32(y, vh);
57 vh = _mm_srli_epi64(vh, 32);
58 y += 2;
59 }
60 if (n & (1 * sizeof(uint16_t))) {
61 *y = _mm_extract_epi16(vh, 0);
62 }
63 }
64 }
65