xref: /aosp_15_r20/external/XNNPACK/src/f16-f32-vcvt/gen/vcvt-avx512skx-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-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_f16_f32_vcvt_ukernel__avx512skx_x16(size_t n,const void * input,float * output,const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f16_f32_vcvt_ukernel__avx512skx_x16(
20     size_t n,
21     const void* input,
22     float* output,
23     const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25   assert(n != 0);
26   assert(n % sizeof(uint16_t) == 0);
27   assert(input != NULL);
28   assert(output != NULL);
29 
30   const uint16_t* i = (const uint16_t*) input;
31   for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
32     const __m512 vacc = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i));
33     i += 16;
34 
35     _mm512_storeu_ps(output, vacc);
36     output += 16;
37   }
38   if XNN_UNLIKELY(n != 0) {
39     assert(n >= 1 * sizeof(uint16_t));
40     assert(n <= 15 * sizeof(uint16_t));
41 
42     // Prepare mask for valid 32-bit elements (depends on n).
43     n >>= 1 /* log2(sizeof(uint16_t)) */;
44     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
45 
46     const __m512 vacc = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(vmask, i));
47 
48     _mm512_mask_storeu_ps(output, vmask, vacc);
49   }
50 }
51