1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-f32-vcvt/sse2.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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_qu8_f32_vcvt_ukernel__sse2_x32(size_t n,const uint8_t * x,float * y,const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_f32_vcvt_ukernel__sse2_x32(
20 size_t n,
21 const uint8_t* x,
22 float* y,
23 const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(uint8_t) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m128i vmagic_exp = _mm_load_si128((const __m128i*) params->sse2.magic_exp);
31 const __m128 vmagic_bias = _mm_load_ps(params->sse2.magic_bias);
32 const __m128 vscale = _mm_load_ps(params->sse2.scale);
33 const __m128i vzero = _mm_setzero_si128();
34 for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
35 __m128i vx01234567 = _mm_loadl_epi64((const __m128i*) x);
36 __m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (x + 8));
37 __m128i vxGHIJKLMN = _mm_loadl_epi64((const __m128i*) (x + 16));
38 __m128i vxOPQRSTUV = _mm_loadl_epi64((const __m128i*) (x + 24));
39 x += 32;
40
41
42 vx01234567 = _mm_unpacklo_epi8(vx01234567, vzero);
43 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero);
44 vxGHIJKLMN = _mm_unpacklo_epi8(vxGHIJKLMN, vzero);
45 vxOPQRSTUV = _mm_unpacklo_epi8(vxOPQRSTUV, vzero);
46
47 __m128 vy0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(vx01234567, vmagic_exp));
48 __m128 vy4567 = _mm_castsi128_ps(_mm_unpackhi_epi16(vx01234567, vmagic_exp));
49 __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp));
50 __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp));
51 __m128 vyGHIJ = _mm_castsi128_ps(_mm_unpacklo_epi16(vxGHIJKLMN, vmagic_exp));
52 __m128 vyKLMN = _mm_castsi128_ps(_mm_unpackhi_epi16(vxGHIJKLMN, vmagic_exp));
53 __m128 vyOPQR = _mm_castsi128_ps(_mm_unpacklo_epi16(vxOPQRSTUV, vmagic_exp));
54 __m128 vySTUV = _mm_castsi128_ps(_mm_unpackhi_epi16(vxOPQRSTUV, vmagic_exp));
55
56 vy0123 = _mm_sub_ps(vy0123, vmagic_bias);
57 vy4567 = _mm_sub_ps(vy4567, vmagic_bias);
58 vy89AB = _mm_sub_ps(vy89AB, vmagic_bias);
59 vyCDEF = _mm_sub_ps(vyCDEF, vmagic_bias);
60 vyGHIJ = _mm_sub_ps(vyGHIJ, vmagic_bias);
61 vyKLMN = _mm_sub_ps(vyKLMN, vmagic_bias);
62 vyOPQR = _mm_sub_ps(vyOPQR, vmagic_bias);
63 vySTUV = _mm_sub_ps(vySTUV, vmagic_bias);
64
65 vy0123 = _mm_mul_ps(vy0123, vscale);
66 vy4567 = _mm_mul_ps(vy4567, vscale);
67 vy89AB = _mm_mul_ps(vy89AB, vscale);
68 vyCDEF = _mm_mul_ps(vyCDEF, vscale);
69 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale);
70 vyKLMN = _mm_mul_ps(vyKLMN, vscale);
71 vyOPQR = _mm_mul_ps(vyOPQR, vscale);
72 vySTUV = _mm_mul_ps(vySTUV, vscale);
73
74 _mm_storeu_ps(y, vy0123);
75 _mm_storeu_ps(y + 4, vy4567);
76 _mm_storeu_ps(y + 8, vy89AB);
77 _mm_storeu_ps(y + 12, vyCDEF);
78 _mm_storeu_ps(y + 16, vyGHIJ);
79 _mm_storeu_ps(y + 20, vyKLMN);
80 _mm_storeu_ps(y + 24, vyOPQR);
81 _mm_storeu_ps(y + 28, vySTUV);
82 y += 32;
83 }
84 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
85 __m128i vx = _mm_loadl_epi64((const __m128i*) x);
86 vx = _mm_unpacklo_epi8(vx, vzero);
87 x += 8;
88
89 __m128 vy_lo = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
90 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));
91
92 vy_lo = _mm_sub_ps(vy_lo, vmagic_bias);
93 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias);
94
95 vy_lo = _mm_mul_ps(vy_lo, vscale);
96 vy_hi = _mm_mul_ps(vy_hi, vscale);
97
98 _mm_storeu_ps(y, vy_lo);
99 _mm_storeu_ps(y + 4, vy_hi);
100 y += 8;
101 }
102 if XNN_UNLIKELY(n != 0) {
103 assert(n >= 1 * sizeof(uint8_t));
104 assert(n <= 7 * sizeof(uint8_t));
105
106 __m128i vx = _mm_loadl_epi64((const __m128i*) x);
107 vx = _mm_unpacklo_epi8(vx, vzero);
108
109 __m128 vy = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
110 vy = _mm_sub_ps(vy, vmagic_bias);
111 vy = _mm_mul_ps(vy, vscale);
112
113 if (n & (4 * sizeof(uint8_t))) {
114 _mm_storeu_ps(y, vy);
115 vy = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));
116 vy = _mm_sub_ps(vy, vmagic_bias);
117 vy = _mm_mul_ps(vy, vscale);
118 y += 4;
119 }
120 if (n & (2 * sizeof(uint8_t))) {
121 _mm_storel_pi((__m64*) y, vy);
122 vy = _mm_movehl_ps(vy, vy);
123 y += 2;
124 }
125 if (n & (1 * sizeof(uint8_t))) {
126 _mm_store_ss(y, vy);
127 }
128 }
129 }
130