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_x24(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_x24(
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 >= 24 * sizeof(uint8_t); n -= 24 * 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 x += 24;
39
40
41 vx01234567 = _mm_unpacklo_epi8(vx01234567, vzero);
42 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero);
43 vxGHIJKLMN = _mm_unpacklo_epi8(vxGHIJKLMN, vzero);
44
45 __m128 vy0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(vx01234567, vmagic_exp));
46 __m128 vy4567 = _mm_castsi128_ps(_mm_unpackhi_epi16(vx01234567, vmagic_exp));
47 __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp));
48 __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp));
49 __m128 vyGHIJ = _mm_castsi128_ps(_mm_unpacklo_epi16(vxGHIJKLMN, vmagic_exp));
50 __m128 vyKLMN = _mm_castsi128_ps(_mm_unpackhi_epi16(vxGHIJKLMN, vmagic_exp));
51
52 vy0123 = _mm_sub_ps(vy0123, vmagic_bias);
53 vy4567 = _mm_sub_ps(vy4567, vmagic_bias);
54 vy89AB = _mm_sub_ps(vy89AB, vmagic_bias);
55 vyCDEF = _mm_sub_ps(vyCDEF, vmagic_bias);
56 vyGHIJ = _mm_sub_ps(vyGHIJ, vmagic_bias);
57 vyKLMN = _mm_sub_ps(vyKLMN, vmagic_bias);
58
59 vy0123 = _mm_mul_ps(vy0123, vscale);
60 vy4567 = _mm_mul_ps(vy4567, vscale);
61 vy89AB = _mm_mul_ps(vy89AB, vscale);
62 vyCDEF = _mm_mul_ps(vyCDEF, vscale);
63 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale);
64 vyKLMN = _mm_mul_ps(vyKLMN, vscale);
65
66 _mm_storeu_ps(y, vy0123);
67 _mm_storeu_ps(y + 4, vy4567);
68 _mm_storeu_ps(y + 8, vy89AB);
69 _mm_storeu_ps(y + 12, vyCDEF);
70 _mm_storeu_ps(y + 16, vyGHIJ);
71 _mm_storeu_ps(y + 20, vyKLMN);
72 y += 24;
73 }
74 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
75 __m128i vx = _mm_loadl_epi64((const __m128i*) x);
76 vx = _mm_unpacklo_epi8(vx, vzero);
77 x += 8;
78
79 __m128 vy_lo = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
80 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));
81
82 vy_lo = _mm_sub_ps(vy_lo, vmagic_bias);
83 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias);
84
85 vy_lo = _mm_mul_ps(vy_lo, vscale);
86 vy_hi = _mm_mul_ps(vy_hi, vscale);
87
88 _mm_storeu_ps(y, vy_lo);
89 _mm_storeu_ps(y + 4, vy_hi);
90 y += 8;
91 }
92 if XNN_UNLIKELY(n != 0) {
93 assert(n >= 1 * sizeof(uint8_t));
94 assert(n <= 7 * sizeof(uint8_t));
95
96 __m128i vx = _mm_loadl_epi64((const __m128i*) x);
97 vx = _mm_unpacklo_epi8(vx, vzero);
98
99 __m128 vy = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
100 vy = _mm_sub_ps(vy, vmagic_bias);
101 vy = _mm_mul_ps(vy, vscale);
102
103 if (n & (4 * sizeof(uint8_t))) {
104 _mm_storeu_ps(y, vy);
105 vy = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));
106 vy = _mm_sub_ps(vy, vmagic_bias);
107 vy = _mm_mul_ps(vy, vscale);
108 y += 4;
109 }
110 if (n & (2 * sizeof(uint8_t))) {
111 _mm_storel_pi((__m64*) y, vy);
112 vy = _mm_movehl_ps(vy, vy);
113 y += 2;
114 }
115 if (n & (1 * sizeof(uint8_t))) {
116 _mm_store_ss(y, vy);
117 }
118 }
119 }
120