1 // Auto-generated file. Do not edit!
2 // Template: src/f16-vunary/sse2.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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17 #include <xnnpack/vunary.h>
18
19
xnn_f16_vneg_ukernel__sse2_x16(size_t n,const void * input,void * output,const union xnn_f16_neg_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_vneg_ukernel__sse2_x16(
21 size_t n,
22 const void* input,
23 void* output,
24 const union xnn_f16_neg_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(n != 0);
27 assert(n % sizeof(uint16_t) == 0);
28 assert(input != NULL);
29 assert(output != NULL);
30
31 const uint16_t* i = (const uint16_t*) input;
32 uint16_t* o = (uint16_t*) output;
33 const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse.sign_mask);
34 for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
35 __m128i vacc0 = _mm_loadu_si128((const __m128i*) i);
36 __m128i vacc1 = _mm_loadu_si128((const __m128i*) (i + 8));
37 i += 16;
38
39 vacc0 = _mm_xor_si128(vacc0, vsign_mask);
40 vacc1 = _mm_xor_si128(vacc1, vsign_mask);
41
42 _mm_storeu_si128((__m128i*) o, vacc0);
43 _mm_storeu_si128((__m128i*) (o + 8), vacc1);
44 o += 16;
45 }
46 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
47 __m128i vacc = _mm_loadu_si128((const __m128i*) i);
48 i += 8;
49 vacc = _mm_xor_si128(vacc, vsign_mask);
50 _mm_storeu_si128((__m128i*) o, vacc);
51 o += 8;
52 }
53 if XNN_UNLIKELY(n != 0) {
54 __m128i vacc = _mm_loadu_si128((const __m128i*) i);
55 vacc = _mm_xor_si128(vacc, vsign_mask);
56 if (n & (4 * sizeof(uint16_t))) {
57 _mm_storel_epi64((__m128i*) o, vacc);
58 o += 4;
59 vacc = _mm_unpackhi_epi64(vacc, vacc);
60 }
61 if (n & (2 * sizeof(uint16_t))) {
62 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vacc));
63 o += 2;
64 vacc = _mm_srli_epi64(vacc, 32);
65 }
66 if (n & (1 * sizeof(uint16_t))) {
67 *o = (uint16_t) _mm_extract_epi16(vacc, 0);
68 }
69 }
70 }
71