• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-f32-vcvt/avx512skx.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 <immintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vcvt.h>
17 
18 
xnn_qs8_f32_vcvt_ukernel__avx512skx_x48(size_t n,const int8_t * x,float * y,const union xnn_qs8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_f32_vcvt_ukernel__avx512skx_x48(
20     size_t n,
21     const int8_t* x,
22     float* y,
23     const union xnn_qs8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25   assert(n != 0);
26   assert(n % sizeof(int8_t) == 0);
27   assert(x != NULL);
28   assert(y != NULL);
29 
30   const __m512i vminus_zero_point = _mm512_load_si512(params->avx512.minus_zero_point);
31   const __m512 vscale = _mm512_load_ps(params->avx512.scale);
32   for (; n >= 48 * sizeof(int8_t); n -= 48 * sizeof(int8_t)) {
33     __m512i vx0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) x));
34     __m512i vxGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (x + 16)));
35     __m512i vxWXYZ = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (x + 32)));
36     x += 48;
37 
38     vx0123456789ABCDEF = _mm512_add_epi32(vx0123456789ABCDEF, vminus_zero_point);
39     vxGHIJKLMNOPQRSTUV = _mm512_add_epi32(vxGHIJKLMNOPQRSTUV, vminus_zero_point);
40     vxWXYZ = _mm512_add_epi32(vxWXYZ, vminus_zero_point);
41 
42     __m512 vy0123456789ABCDEF = _mm512_cvtepi32_ps(vx0123456789ABCDEF);
43     __m512 vyGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vxGHIJKLMNOPQRSTUV);
44     __m512 vyWXYZ = _mm512_cvtepi32_ps(vxWXYZ);
45 
46     vy0123456789ABCDEF = _mm512_mul_ps(vy0123456789ABCDEF, vscale);
47     vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vscale);
48     vyWXYZ = _mm512_mul_ps(vyWXYZ, vscale);
49 
50     _mm512_storeu_ps(y, vy0123456789ABCDEF);
51     _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV);
52     _mm512_storeu_ps(y + 32, vyWXYZ);
53     y += 48;
54   }
55   for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
56     __m512i vx = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) x));
57     vx = _mm512_add_epi32(vx, vminus_zero_point);
58     x += 16;
59 
60     __m512 vy = _mm512_cvtepi32_ps(vx);
61     vy = _mm512_mul_ps(vy, vscale);
62 
63     _mm512_storeu_ps(y, vy);
64     y += 16;
65   }
66   if XNN_UNLIKELY(n != 0) {
67     assert(n >= 1 * sizeof(int8_t));
68     assert(n <= 15 * sizeof(int8_t));
69 
70     // Prepare mask for valid elements (depends on n).
71     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
72 
73     __m512i vx = _mm512_cvtepi8_epi32(_mm_maskz_loadu_epi8(vmask, x));
74     vx = _mm512_add_epi32(vx, vminus_zero_point);
75 
76     __m512 vy = _mm512_cvtepi32_ps(vx);
77     vy = _mm512_mul_ps(vy, vscale);
78 
79     _mm512_mask_storeu_ps(y, vmask, vy);
80   }
81 }
82