• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_x8(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_x8(
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 >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
35     __m128i vx = _mm_loadl_epi64((const __m128i*) x);
36     vx = _mm_unpacklo_epi8(vx, vzero);
37     x += 8;
38 
39     __m128 vy_lo = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
40     __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));
41 
42     vy_lo = _mm_sub_ps(vy_lo, vmagic_bias);
43     vy_hi = _mm_sub_ps(vy_hi, vmagic_bias);
44 
45     vy_lo = _mm_mul_ps(vy_lo, vscale);
46     vy_hi = _mm_mul_ps(vy_hi, vscale);
47 
48     _mm_storeu_ps(y, vy_lo);
49     _mm_storeu_ps(y + 4, vy_hi);
50     y += 8;
51   }
52   if XNN_UNLIKELY(n != 0) {
53     assert(n >= 1 * sizeof(uint8_t));
54     assert(n <= 7 * sizeof(uint8_t));
55 
56     __m128i vx = _mm_loadl_epi64((const __m128i*) x);
57     vx = _mm_unpacklo_epi8(vx, vzero);
58 
59     __m128 vy = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
60     vy = _mm_sub_ps(vy, vmagic_bias);
61     vy = _mm_mul_ps(vy, vscale);
62 
63     if (n & (4 * sizeof(uint8_t))) {
64       _mm_storeu_ps(y, vy);
65       vy = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));
66       vy = _mm_sub_ps(vy, vmagic_bias);
67       vy = _mm_mul_ps(vy, vscale);
68       y += 4;
69     }
70     if (n & (2 * sizeof(uint8_t))) {
71       _mm_storel_pi((__m64*) y, vy);
72       vy = _mm_movehl_ps(vy, vy);
73       y += 2;
74     }
75     if (n & (1 * sizeof(uint8_t))) {
76       _mm_store_ss(y, vy);
77     }
78   }
79 }
80