1 // Auto-generated file. Do not edit! 2 // Template: src/f32-f16-vcvt/scalar-fabsf.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 #include <math.h> 12 13 #include <xnnpack/common.h> 14 #include <xnnpack/math.h> 15 #include <xnnpack/vcvt.h> 16 17 #include <fp16.h> 18 19 xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x2(size_t n,const float * input,void * output,const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])20void xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x2( 21 size_t n, 22 const float* input, 23 void* output, 24 const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) 25 { 26 assert(n != 0); 27 assert(n % sizeof(float) == 0); 28 assert(input != NULL); 29 assert(output != NULL); 30 31 const float vscale_to_inf = params->scalar_fabsf.scale_to_inf; 32 const uint32_t vexp_bias = params->scalar_fabsf.exp_bias; 33 const float vscale_to_zero = params->scalar_fabsf.scale_to_zero; 34 const uint32_t vexpw_max = params->scalar_fabsf.expw_max; 35 const uint32_t vbias_min = params->scalar_fabsf.bias_min; 36 const uint16_t vexph_mask = params->scalar_fabsf.exph_mask; 37 const uint16_t vmanth_mask = params->scalar_fabsf.manth_mask; 38 const uint16_t vnanh = params->scalar_fabsf.nanh; 39 40 uint16_t* o = (uint16_t*) output; 41 for (; n >= 2 * sizeof(float); n -= 2 * sizeof(float)) { 42 const float vx0 = input[0]; 43 const float vx1 = input[1]; 44 input += 2; 45 46 const float vabsx0 = fabsf(vx0); 47 const float vabsx1 = fabsf(vx1); 48 uint32_t vsignw0 = fp32_to_bits(vx0); 49 uint32_t vsignw1 = fp32_to_bits(vx1); 50 51 const uint32_t vnonsignw0 = fp32_to_bits(vabsx0); 52 const uint32_t vnonsignw1 = fp32_to_bits(vabsx1); 53 float vf0 = vabsx0 * vscale_to_inf; 54 float vf1 = vabsx1 * vscale_to_inf; 55 56 uint32_t vbias0 = vnonsignw0 + vexp_bias; 57 uint32_t vbias1 = vnonsignw1 + vexp_bias; 58 vsignw0 ^= vnonsignw0; 59 vsignw1 ^= vnonsignw1; 60 61 vf0 *= vscale_to_zero; 62 vf1 *= vscale_to_zero; 63 vbias0 &= vexpw_max; 64 vbias1 &= vexpw_max; 65 66 vbias0 = math_max_u32(vbias0, vbias_min); 67 vbias1 = math_max_u32(vbias1, vbias_min); 68 69 vf0 += fp32_from_bits(vbias0); 70 vf1 += fp32_from_bits(vbias1); 71 72 const uint32_t vbits0 = fp32_to_bits(vf0); 73 const uint32_t vbits1 = fp32_to_bits(vf1); 74 75 const uint16_t vexph0 = (uint16_t) (vbits0 >> 13) & vexph_mask; 76 const uint16_t vexph1 = (uint16_t) (vbits1 >> 13) & vexph_mask; 77 const uint16_t vmanth0 = (uint16_t) vbits0 & vmanth_mask; 78 const uint16_t vmanth1 = (uint16_t) vbits1 & vmanth_mask; 79 const uint16_t vsignh0 = (uint16_t) (vsignw0 >> 16); 80 const uint16_t vsignh1 = (uint16_t) (vsignw1 >> 16); 81 82 uint16_t vh0 = vexph0 + vmanth0; 83 uint16_t vh1 = vexph1 + vmanth1; 84 if XNN_UNPREDICTABLE(vnonsignw0 > vexpw_max) { 85 vh0 = vnanh; 86 } 87 if XNN_UNPREDICTABLE(vnonsignw1 > vexpw_max) { 88 vh1 = vnanh; 89 } 90 vh0 |= vsignh0; 91 vh1 |= vsignh1; 92 93 o[0] = vh0; 94 o[1] = vh1; 95 o += 2; 96 } 97 if XNN_UNLIKELY(n != 0) { 98 const float vx = *input; 99 100 const float vabsx = fabsf(vx); 101 uint32_t vsignw = fp32_to_bits(vx); 102 103 const uint32_t vnonsignw = fp32_to_bits(vabsx); 104 float vf = vabsx * vscale_to_inf; 105 106 uint32_t vbias = vnonsignw + vexp_bias; 107 vsignw ^= vnonsignw; 108 109 vf *= vscale_to_zero; 110 vbias &= vexpw_max; 111 112 vbias = math_max_u32(vbias, vbias_min); 113 114 vf += fp32_from_bits(vbias); 115 116 const uint32_t vbits = fp32_to_bits(vf); 117 118 const uint16_t vexph = (uint16_t) (vbits >> 13) & vexph_mask; 119 const uint16_t vmanth = (uint16_t) vbits & vmanth_mask; 120 const uint16_t vsignh = (uint16_t) (vsignw >> 16); 121 122 uint16_t vh = vexph + vmanth; 123 if XNN_UNPREDICTABLE(vnonsignw > vexpw_max) { 124 vh = vnanh; 125 } 126 vh |= vsignh; 127 128 *o = vh; 129 } 130 } 131