xref: /aosp_15_r20/external/XNNPACK/src/f32-vrelu/avx512f.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker// Copyright 2020 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker//
3*4bdc9457SAndroid Build Coastguard Worker// This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker// LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker
6*4bdc9457SAndroid Build Coastguard Worker$assert BATCH_TILE % 16 == 0
7*4bdc9457SAndroid Build Coastguard Worker$assert BATCH_TILE >= 16
8*4bdc9457SAndroid Build Coastguard Worker$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9*4bdc9457SAndroid Build Coastguard Worker#include <assert.h>
10*4bdc9457SAndroid Build Coastguard Worker
11*4bdc9457SAndroid Build Coastguard Worker#include <immintrin.h>
12*4bdc9457SAndroid Build Coastguard Worker
13*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/vunary.h>
14*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/common.h>
15*4bdc9457SAndroid Build Coastguard Worker#include <xnnpack/intrinsics-polyfill.h>
16*4bdc9457SAndroid Build Coastguard Worker
17*4bdc9457SAndroid Build Coastguard Worker
18*4bdc9457SAndroid Build Coastguard Workervoid xnn_f32_vrelu_ukernel__avx512f_x${BATCH_TILE}(
19*4bdc9457SAndroid Build Coastguard Worker    size_t n,
20*4bdc9457SAndroid Build Coastguard Worker    const float* x,
21*4bdc9457SAndroid Build Coastguard Worker    float* y,
22*4bdc9457SAndroid Build Coastguard Worker    const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
23*4bdc9457SAndroid Build Coastguard Worker{
24*4bdc9457SAndroid Build Coastguard Worker  assert(n != 0);
25*4bdc9457SAndroid Build Coastguard Worker  assert(n % sizeof(float) == 0);
26*4bdc9457SAndroid Build Coastguard Worker  assert(x != NULL);
27*4bdc9457SAndroid Build Coastguard Worker  assert(y != NULL);
28*4bdc9457SAndroid Build Coastguard Worker
29*4bdc9457SAndroid Build Coastguard Worker  const __m512 vzero = _mm512_setzero_ps();
30*4bdc9457SAndroid Build Coastguard Worker
31*4bdc9457SAndroid Build Coastguard Worker  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
32*4bdc9457SAndroid Build Coastguard Worker    __m512 vacc${ABC[0:16]} = _mm512_loadu_ps(x);
33*4bdc9457SAndroid Build Coastguard Worker    $for N in range(16, BATCH_TILE, 16):
34*4bdc9457SAndroid Build Coastguard Worker      __m512 vacc${ABC[N:N+16]} = _mm512_loadu_ps(x + ${N});
35*4bdc9457SAndroid Build Coastguard Worker    x += ${BATCH_TILE};
36*4bdc9457SAndroid Build Coastguard Worker
37*4bdc9457SAndroid Build Coastguard Worker    $for N in range(0, BATCH_TILE, 16):
38*4bdc9457SAndroid Build Coastguard Worker      vacc${ABC[N:N+16]} = _mm512_max_ps(vacc${ABC[N:N+16]}, vzero);
39*4bdc9457SAndroid Build Coastguard Worker
40*4bdc9457SAndroid Build Coastguard Worker    _mm512_storeu_ps(y, vacc${ABC[0:16]});
41*4bdc9457SAndroid Build Coastguard Worker    $for N in range(16, BATCH_TILE, 16):
42*4bdc9457SAndroid Build Coastguard Worker      _mm512_storeu_ps(y + ${N}, vacc${ABC[N:N+16]});
43*4bdc9457SAndroid Build Coastguard Worker    y += ${BATCH_TILE};
44*4bdc9457SAndroid Build Coastguard Worker  }
45*4bdc9457SAndroid Build Coastguard Worker  $if BATCH_TILE > 16:
46*4bdc9457SAndroid Build Coastguard Worker    for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
47*4bdc9457SAndroid Build Coastguard Worker      __m512 vacc = _mm512_loadu_ps(x);
48*4bdc9457SAndroid Build Coastguard Worker      x += 16;
49*4bdc9457SAndroid Build Coastguard Worker
50*4bdc9457SAndroid Build Coastguard Worker      vacc = _mm512_max_ps(vacc, vzero);
51*4bdc9457SAndroid Build Coastguard Worker
52*4bdc9457SAndroid Build Coastguard Worker      _mm512_storeu_ps(y, vacc);
53*4bdc9457SAndroid Build Coastguard Worker      y += 16;
54*4bdc9457SAndroid Build Coastguard Worker    }
55*4bdc9457SAndroid Build Coastguard Worker  if XNN_UNLIKELY(n != 0) {
56*4bdc9457SAndroid Build Coastguard Worker    assert(n >= 1 * sizeof(float));
57*4bdc9457SAndroid Build Coastguard Worker    assert(n <= 15 * sizeof(float));
58*4bdc9457SAndroid Build Coastguard Worker    // Prepare mask for valid 32-bit elements (depends on n).
59*4bdc9457SAndroid Build Coastguard Worker    n >>= 2 /* log2(sizeof(float)) */;
60*4bdc9457SAndroid Build Coastguard Worker    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
61*4bdc9457SAndroid Build Coastguard Worker
62*4bdc9457SAndroid Build Coastguard Worker    __m512 vacc = _mm512_maskz_loadu_ps(vmask, x);
63*4bdc9457SAndroid Build Coastguard Worker    vacc = _mm512_max_ps(vacc, vzero);
64*4bdc9457SAndroid Build Coastguard Worker    _mm512_mask_storeu_ps(y, vmask, vacc);
65*4bdc9457SAndroid Build Coastguard Worker  }
66*4bdc9457SAndroid Build Coastguard Worker}
67