1 // Auto-generated file. Do not edit!
2 // Template: src/f32-f16-vcvt/sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <smmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16
17
xnn_f32_f16_vcvt_ukernel__avx_x8(size_t n,const float * input,void * output,const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_f16_vcvt_ukernel__avx_x8(
19 size_t n,
20 const float* input,
21 void* output,
22 const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(input != NULL);
27 assert(output != NULL);
28
29 const __m128 vnonsign_mask = _mm_load_ps((const float*) params->sse2.nonsign_mask);
30 const __m128i vexp_bias = _mm_load_si128((const __m128i*) params->sse2.exp_bias);
31 const __m128 vscale_to_inf = _mm_load_ps(params->sse2.scale_to_inf);
32 const __m128i vexpw_max = _mm_load_si128((const __m128i*) params->sse2.expw_max);
33 const __m128 vscale_to_zero = _mm_load_ps(params->sse2.scale_to_zero);
34 const __m128i vbias_min = _mm_load_si128((const __m128i*) params->sse2.bias_min);
35 const __m128i vmanth_mask = _mm_load_si128((const __m128i*) params->sse2.manth_mask);
36 const __m128i vexph_mask = _mm_load_si128((const __m128i*) params->sse2.exph_mask);
37 const __m128i vnanh = _mm_load_si128((const __m128i*) params->sse2.nanh);
38
39 uint16_t* o = (uint16_t*) output;
40 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
41 const __m128 vx_lo = _mm_loadu_ps(input);
42 const __m128 vx_hi = _mm_loadu_ps(input + 4);
43 input += 8;
44
45 const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask);
46 const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask);
47
48 const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo);
49 const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi);
50 __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias);
51 __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias);
52 __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf);
53 __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf);
54 const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max);
55 const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max);
56
57 vbias_lo = _mm_and_si128(vbias_lo, vexpw_max);
58 vbias_hi = _mm_and_si128(vbias_hi, vexpw_max);
59 vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero);
60 vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero);
61 const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi);
62 const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi));
63
64 vbias_lo = _mm_max_epi16(vbias_lo, vbias_min);
65 vbias_hi = _mm_max_epi16(vbias_hi, vbias_min);
66
67 vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo));
68 vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi));
69
70 __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13);
71 __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13);
72 const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask);
73 const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask);
74
75 vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask);
76 vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask);
77
78 const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo);
79 const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi);
80
81 const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi);
82
83 const __m128i vabsh = _mm_blendv_epi8(vnonsignh, vnanh, vnanmaskh);
84
85 const __m128i vh = _mm_or_si128(vabsh, vsignh);
86
87 _mm_storeu_si128((__m128i*) o, vh);
88 o += 8;
89 }
90 if XNN_UNPREDICTABLE(n != 0) {
91 const __m128 vx_lo = _mm_loadu_ps(input);
92 const float* input_hi = (const float*) ((uintptr_t) input + (n & (4 * sizeof(float))));
93 const __m128 vx_hi = _mm_loadu_ps(input_hi);
94
95 const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask);
96 const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask);
97
98 const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo);
99 const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi);
100 __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias);
101 __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias);
102 __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf);
103 __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf);
104 const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max);
105 const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max);
106
107 vbias_lo = _mm_and_si128(vbias_lo, vexpw_max);
108 vbias_hi = _mm_and_si128(vbias_hi, vexpw_max);
109 vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero);
110 vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero);
111 const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi);
112 const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi));
113
114 vbias_lo = _mm_max_epi16(vbias_lo, vbias_min);
115 vbias_hi = _mm_max_epi16(vbias_hi, vbias_min);
116
117 vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo));
118 vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi));
119
120 __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13);
121 __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13);
122 const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask);
123 const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask);
124
125 vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask);
126 vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask);
127
128 const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo);
129 const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi);
130
131 const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi);
132
133 const __m128i vabsh = _mm_blendv_epi8(vnonsignh, vnanh, vnanmaskh);
134
135 __m128i vh = _mm_or_si128(vabsh, vsignh);
136
137 if (n & (4 * sizeof(float))) {
138 _mm_storel_epi64((__m128i*) o, vh);
139 vh = _mm_unpackhi_epi64(vh, vh);
140 o += 4;
141 }
142 if (n & (2 * sizeof(float))) {
143 *((uint32_t*) o) = (uint32_t) _mm_cvtsi128_si32(vh);
144 vh = _mm_srli_epi64(vh, 32);
145 o += 2;
146 }
147 if (n & (1 * sizeof(float))) {
148 *o = (uint16_t) _mm_extract_epi16(vh, 0);
149 }
150 }
151 }
152