1 // Auto-generated file. Do not edit!
2 // Template: src/f16-vsqrt/f16c-sqrt.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_vsqrt_ukernel__f16c_sqrt_x16(size_t n,const void * input,void * output,const union xnn_f16_sqrt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f16_vsqrt_ukernel__f16c_sqrt_x16(
20 size_t n,
21 const void* input,
22 void* output,
23 const union xnn_f16_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(uint16_t) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 const uint16_t* i = (const uint16_t*) input;
31 uint16_t* o = (uint16_t*) output;
32 for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
33 __m256 vacc0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
34 __m256 vacc1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 8)));
35 i += 16;
36
37 vacc0 = _mm256_sqrt_ps(vacc0);
38 vacc1 = _mm256_sqrt_ps(vacc1);
39
40 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc0, _MM_FROUND_NO_EXC));
41 _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vacc1, _MM_FROUND_NO_EXC));
42 o += 16;
43 }
44 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
45 __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
46 i += 8;
47 vacc = _mm256_sqrt_ps(vacc);
48 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC));
49 o += 8;
50 }
51 if XNN_UNLIKELY(n != 0) {
52 __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
53 vacc = _mm256_sqrt_ps(vacc);
54 __m128i vh = _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC);
55 if (n & (4 * sizeof(uint16_t))) {
56 _mm_storel_epi64((__m128i*) o, vh);
57 o += 4;
58 vh = _mm_unpackhi_epi64(vh, vh);
59 }
60 if (n & (2 * sizeof(uint16_t))) {
61 _mm_storeu_si32(o, vh);
62 o += 2;
63 vh = _mm_srli_epi64(vh, 32);
64 }
65 if (n & (1 * sizeof(uint16_t))) {
66 *o = (uint16_t) _mm_extract_epi16(vh, 0);
67 }
68 }
69 }
70