xref: /aosp_15_r20/external/XNNPACK/src/qs8-f32-vcvt/gen/vcvt-avx512skx-x64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_x64(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_x64(
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 >= 64 * sizeof(int8_t); n -= 64 * 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     __m512i vx = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (x + 48)));
37     x += 64;
38 
39     vx0123456789ABCDEF = _mm512_add_epi32(vx0123456789ABCDEF, vminus_zero_point);
40     vxGHIJKLMNOPQRSTUV = _mm512_add_epi32(vxGHIJKLMNOPQRSTUV, vminus_zero_point);
41     vxWXYZ = _mm512_add_epi32(vxWXYZ, vminus_zero_point);
42     vx = _mm512_add_epi32(vx, vminus_zero_point);
43 
44     __m512 vy0123456789ABCDEF = _mm512_cvtepi32_ps(vx0123456789ABCDEF);
45     __m512 vyGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vxGHIJKLMNOPQRSTUV);
46     __m512 vyWXYZ = _mm512_cvtepi32_ps(vxWXYZ);
47     __m512 vy = _mm512_cvtepi32_ps(vx);
48 
49     vy0123456789ABCDEF = _mm512_mul_ps(vy0123456789ABCDEF, vscale);
50     vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vscale);
51     vyWXYZ = _mm512_mul_ps(vyWXYZ, vscale);
52     vy = _mm512_mul_ps(vy, vscale);
53 
54     _mm512_storeu_ps(y, vy0123456789ABCDEF);
55     _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV);
56     _mm512_storeu_ps(y + 32, vyWXYZ);
57     _mm512_storeu_ps(y + 48, vy);
58     y += 64;
59   }
60   for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
61     __m512i vx = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) x));
62     vx = _mm512_add_epi32(vx, vminus_zero_point);
63     x += 16;
64 
65     __m512 vy = _mm512_cvtepi32_ps(vx);
66     vy = _mm512_mul_ps(vy, vscale);
67 
68     _mm512_storeu_ps(y, vy);
69     y += 16;
70   }
71   if XNN_UNLIKELY(n != 0) {
72     assert(n >= 1 * sizeof(int8_t));
73     assert(n <= 15 * sizeof(int8_t));
74 
75     // Prepare mask for valid elements (depends on n).
76     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
77 
78     __m512i vx = _mm512_cvtepi8_epi32(_mm_maskz_loadu_epi8(vmask, x));
79     vx = _mm512_add_epi32(vx, vminus_zero_point);
80 
81     __m512 vy = _mm512_cvtepi32_ps(vx);
82     vy = _mm512_mul_ps(vy, vscale);
83 
84     _mm512_mask_storeu_ps(y, vmask, vy);
85   }
86 }
87