File size: 3,760 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 |
// Auto-generated file. Do not edit!
// Template: src/f32-f16-vcvt/scalar-fabsf.c.in
// Generator: tools/xngen
//
// Copyright 2021 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.
#include <assert.h>
#include <math.h>
#include <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/vcvt.h>
void xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x2(
size_t batch,
const float* input,
void* output,
const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
{
assert(batch != 0);
assert(batch % sizeof(float) == 0);
assert(input != NULL);
assert(output != NULL);
const float vscale_to_inf = params->scalar_fabsf.scale_to_inf;
const uint32_t vexp_bias = params->scalar_fabsf.exp_bias;
const float vscale_to_zero = params->scalar_fabsf.scale_to_zero;
const uint32_t vexpw_max = params->scalar_fabsf.expw_max;
const uint32_t vbias_min = params->scalar_fabsf.bias_min;
const uint16_t vexph_mask = params->scalar_fabsf.exph_mask;
const uint16_t vmanth_mask = params->scalar_fabsf.manth_mask;
const uint16_t vnanh = params->scalar_fabsf.nanh;
uint16_t* o = (uint16_t*) output;
for (; batch >= 2 * sizeof(float); batch -= 2 * sizeof(float)) {
const float vx0 = input[0];
const float vx1 = input[1];
input += 2;
const float vabsx0 = fabsf(vx0);
const float vabsx1 = fabsf(vx1);
uint32_t vsignw0 = float_as_uint32(vx0);
uint32_t vsignw1 = float_as_uint32(vx1);
const uint32_t vnonsignw0 = float_as_uint32(vabsx0);
const uint32_t vnonsignw1 = float_as_uint32(vabsx1);
float vf0 = vabsx0 * vscale_to_inf;
float vf1 = vabsx1 * vscale_to_inf;
uint32_t vbias0 = vnonsignw0 + vexp_bias;
uint32_t vbias1 = vnonsignw1 + vexp_bias;
vsignw0 ^= vnonsignw0;
vsignw1 ^= vnonsignw1;
vf0 *= vscale_to_zero;
vf1 *= vscale_to_zero;
vbias0 &= vexpw_max;
vbias1 &= vexpw_max;
vbias0 = math_max_u32(vbias0, vbias_min);
vbias1 = math_max_u32(vbias1, vbias_min);
vf0 += uint32_as_float(vbias0);
vf1 += uint32_as_float(vbias1);
const uint32_t vbits0 = float_as_uint32(vf0);
const uint32_t vbits1 = float_as_uint32(vf1);
const uint16_t vexph0 = (uint16_t) (vbits0 >> 13) & vexph_mask;
const uint16_t vexph1 = (uint16_t) (vbits1 >> 13) & vexph_mask;
const uint16_t vmanth0 = (uint16_t) vbits0 & vmanth_mask;
const uint16_t vmanth1 = (uint16_t) vbits1 & vmanth_mask;
const uint16_t vsignh0 = (uint16_t) (vsignw0 >> 16);
const uint16_t vsignh1 = (uint16_t) (vsignw1 >> 16);
uint16_t vh0 = vexph0 + vmanth0;
uint16_t vh1 = vexph1 + vmanth1;
if XNN_UNPREDICTABLE(vnonsignw0 > vexpw_max) {
vh0 = vnanh;
}
if XNN_UNPREDICTABLE(vnonsignw1 > vexpw_max) {
vh1 = vnanh;
}
vh0 |= vsignh0;
vh1 |= vsignh1;
o[0] = vh0;
o[1] = vh1;
o += 2;
}
if XNN_UNLIKELY(batch != 0) {
const float vx = *input;
const float vabsx = fabsf(vx);
uint32_t vsignw = float_as_uint32(vx);
const uint32_t vnonsignw = float_as_uint32(vabsx);
float vf = vabsx * vscale_to_inf;
uint32_t vbias = vnonsignw + vexp_bias;
vsignw ^= vnonsignw;
vf *= vscale_to_zero;
vbias &= vexpw_max;
vbias = math_max_u32(vbias, vbias_min);
vf += uint32_as_float(vbias);
const uint32_t vbits = float_as_uint32(vf);
const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask;
const uint16_t vmanth = (uint16_t) vbits & vmanth_mask;
const uint16_t vsignh = (uint16_t) (vsignw >> 16);
uint16_t vh = vexph + vmanth;
if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) {
vh = vnanh;
}
vh |= vsignh;
*o = vh;
}
}
|