Debian Codegraph @

src/qu8-f32-vcvt/gen/qu8-f32-vcvt-avx512skx-u48.c

raw
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
19 void xnn_qu8_f32_vcvt_ukernel__avx512skx_u48(
20 size_t batch,
21 const uint8_t* input,
22 float* output,
23 const struct xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 const __m512i vzero_point = _mm512_set1_epi32(params->scalar.zero_point);
31 const __m512 vscale = _mm512_set1_ps(params->scalar.scale);
32 XNN_FORCE_REALIZATION(vzero_point);
33 XNN_FORCE_REALIZATION(vscale);
34 for (; batch >= 48 * sizeof(uint8_t); batch -= 48 * sizeof(uint8_t)) {
35 __m512i vx0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input));
36 __m512i vxGHIJKLMNOPQRSTUV = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) (input + 16)));
37 __m512i vxWXYZ = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) (input + 32)));
38 input += 48;
39
40 vx0123456789ABCDEF = _mm512_sub_epi32(vx0123456789ABCDEF, vzero_point);
41 vxGHIJKLMNOPQRSTUV = _mm512_sub_epi32(vxGHIJKLMNOPQRSTUV, vzero_point);
42 vxWXYZ = _mm512_sub_epi32(vxWXYZ, vzero_point);
43
44 __m512 vy0123456789ABCDEF = _mm512_cvtepi32_ps(vx0123456789ABCDEF);
45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vxGHIJKLMNOPQRSTUV);
46 __m512 vyWXYZ = _mm512_cvtepi32_ps(vxWXYZ);
47
48 vy0123456789ABCDEF = _mm512_mul_ps(vy0123456789ABCDEF, vscale);
49 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vscale);
50 vyWXYZ = _mm512_mul_ps(vyWXYZ, vscale);
51
52 _mm512_storeu_ps(output, vy0123456789ABCDEF);
53 _mm512_storeu_ps(output + 16, vyGHIJKLMNOPQRSTUV);
54 _mm512_storeu_ps(output + 32, vyWXYZ);
55 output += 48;
56 }
57 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
58 __m512i vx = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) input));
59 vx = _mm512_sub_epi32(vx, vzero_point);
60 input += 16;
61
62 __m512 vy = _mm512_cvtepi32_ps(vx);
63 vy = _mm512_mul_ps(vy, vscale);
64
65 _mm512_storeu_ps(output, vy);
66 output += 16;
67 }
68 if XNN_UNLIKELY(batch != 0) {
69 assert(batch >= 1 * sizeof(uint8_t));
70 assert(batch <= 15 * sizeof(uint8_t));
71
72 // Prepare mask for valid elements (depends on batch).
73 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
74
75 __m512i vx = _mm512_cvtepu8_epi32(_mm_maskz_loadu_epi8(vmask, input));
76 vx = _mm512_sub_epi32(vx, vzero_point);
77
78 __m512 vy = _mm512_cvtepi32_ps(vx);
79 vy = _mm512_mul_ps(vy, vscale);
80
81 _mm512_mask_storeu_ps(output, vmask, vy);
82 }
83 }