blob: 4a7511ca1f74de54a7e4fd58785ceac4958b2f3d [file] [log] [blame]
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
$assert BATCH_TILE % 4 == 0
$assert BATCH_TILE >= 4
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>
#include <math.h>
#include <arm_neon.h>
#include <xnnpack/common.h>
#include <xnnpack/vunary.h>
void xnn_f32_vsqrt_ukernel__neon_sqrt_x${BATCH_TILE}(
size_t n,
const float* x,
float* y,
const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
assert(n != 0);
assert(n % sizeof(float) == 0);
$if BATCH_TILE > 4:
for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
$for N in range(0, BATCH_TILE, 4):
const float32x4_t vx${ABC[N:N+4]} = vld1q_f32(x); x += 4;
$for N in range(0, BATCH_TILE, 4):
const float32x4_t vy${ABC[N:N+4]} = vsqrtq_f32(vx${ABC[N:N+4]});
$for N in range(0, BATCH_TILE, 4):
vst1q_f32(y, vy${ABC[N:N+4]}); y += 4;
}
for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
const float32x4_t vx = vld1q_f32(x); x += 4;
const float32x4_t vy = vsqrtq_f32(vx);
vst1q_f32(y, vy); y += 4;
}
if XNN_UNLIKELY(n != 0) {
do {
const float vx = *x++;
const float vy = sqrtf(vx);
*y++ = vy;
n -= sizeof(float);
} while (n != 0);
}
}