xref: /aosp_15_r20/external/XNNPACK/src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-vbinary/vop-avx512f.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <immintrin.h>
13*4bdc9457SAndroid Build Coastguard Worker 
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/intrinsics-polyfill.h>
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vbinary.h>
17*4bdc9457SAndroid Build Coastguard Worker 
18*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_vadd_minmax_ukernel__avx512f_x16(size_t n,const float * a,const float * b,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_vadd_minmax_ukernel__avx512f_x16(
20*4bdc9457SAndroid Build Coastguard Worker     size_t n,
21*4bdc9457SAndroid Build Coastguard Worker     const float* a,
22*4bdc9457SAndroid Build Coastguard Worker     const float* b,
23*4bdc9457SAndroid Build Coastguard Worker     float* y,
24*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
25*4bdc9457SAndroid Build Coastguard Worker {
26*4bdc9457SAndroid Build Coastguard Worker   assert(n != 0);
27*4bdc9457SAndroid Build Coastguard Worker   assert(n % sizeof(float) == 0);
28*4bdc9457SAndroid Build Coastguard Worker   assert(a != NULL);
29*4bdc9457SAndroid Build Coastguard Worker   assert(b != NULL);
30*4bdc9457SAndroid Build Coastguard Worker   assert(y != NULL);
31*4bdc9457SAndroid Build Coastguard Worker 
32*4bdc9457SAndroid Build Coastguard Worker   const __m512 vy_min = _mm512_set1_ps(params->scalar.min);
33*4bdc9457SAndroid Build Coastguard Worker   const __m512 vy_max = _mm512_set1_ps(params->scalar.max);
34*4bdc9457SAndroid Build Coastguard Worker 
35*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
36*4bdc9457SAndroid Build Coastguard Worker     const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a);
37*4bdc9457SAndroid Build Coastguard Worker     a += 16;
38*4bdc9457SAndroid Build Coastguard Worker 
39*4bdc9457SAndroid Build Coastguard Worker     const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b);
40*4bdc9457SAndroid Build Coastguard Worker     b += 16;
41*4bdc9457SAndroid Build Coastguard Worker 
42*4bdc9457SAndroid Build Coastguard Worker     __m512 vy0123456789ABCDEF = _mm512_add_ps(va0123456789ABCDEF, vb0123456789ABCDEF);
43*4bdc9457SAndroid Build Coastguard Worker 
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker     vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min);
46*4bdc9457SAndroid Build Coastguard Worker 
47*4bdc9457SAndroid Build Coastguard Worker     vy0123456789ABCDEF = _mm512_min_ps(vy0123456789ABCDEF, vy_max);
48*4bdc9457SAndroid Build Coastguard Worker 
49*4bdc9457SAndroid Build Coastguard Worker     _mm512_storeu_ps(y, vy0123456789ABCDEF);
50*4bdc9457SAndroid Build Coastguard Worker     y += 16;
51*4bdc9457SAndroid Build Coastguard Worker   }
52*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
53*4bdc9457SAndroid Build Coastguard Worker     assert(n >= 1 * sizeof(float));
54*4bdc9457SAndroid Build Coastguard Worker     assert(n <= 15 * sizeof(float));
55*4bdc9457SAndroid Build Coastguard Worker     // Prepare mask for valid 32-bit elements (depends on n).
56*4bdc9457SAndroid Build Coastguard Worker     n >>= 2 /* log2(sizeof(float)) */;
57*4bdc9457SAndroid Build Coastguard Worker     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
58*4bdc9457SAndroid Build Coastguard Worker 
59*4bdc9457SAndroid Build Coastguard Worker     const __m512 va = _mm512_maskz_loadu_ps(vmask, a);
60*4bdc9457SAndroid Build Coastguard Worker     const __m512 vb = _mm512_maskz_loadu_ps(vmask, b);
61*4bdc9457SAndroid Build Coastguard Worker 
62*4bdc9457SAndroid Build Coastguard Worker     __m512 vy = _mm512_add_ps(va, vb);
63*4bdc9457SAndroid Build Coastguard Worker     vy = _mm512_max_ps(vy, vy_min);
64*4bdc9457SAndroid Build Coastguard Worker     vy = _mm512_min_ps(vy, vy_max);
65*4bdc9457SAndroid Build Coastguard Worker     _mm512_mask_storeu_ps(y, vmask, vy);
66*4bdc9457SAndroid Build Coastguard Worker   }
67*4bdc9457SAndroid Build Coastguard Worker }
68