1 // Auto-generated file. Do not edit! 2 // Template: src/f16-f32-vcvt/scalar.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 12 #include <xnnpack/common.h> 13 #include <xnnpack/math.h> 14 #include <xnnpack/vcvt.h> 15 16 xnn_f16_f32_vcvt_ukernel__scalar_x4(size_t n,const void * input,float * output,const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])17void xnn_f16_f32_vcvt_ukernel__scalar_x4( 18 size_t n, 19 const void* input, 20 float* output, 21 const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) 22 { 23 assert(n != 0); 24 assert(n % sizeof(uint16_t) == 0); 25 assert(input != NULL); 26 assert(output != NULL); 27 28 const uint32_t vsign_mask = params->scalar.sign_mask; 29 const uint32_t vexp_offset = params->scalar.exp_offset; 30 const float vexp_scale = params->scalar.exp_scale; 31 const uint32_t vmagic_mask = params->scalar.magic_mask; 32 const float vmagic_bias = params->scalar.magic_bias; 33 const uint32_t vdenorm_cutoff = params->scalar.denorm_cutoff; 34 35 const uint16_t* i = (const uint16_t*) input; 36 uint32_t* o = (uint32_t*) output; 37 for (; n >= 4 * sizeof(uint16_t); n -= 4 * sizeof(uint16_t)) { 38 const uint16_t vh0 = i[0]; 39 const uint16_t vh1 = i[1]; 40 const uint16_t vh2 = i[2]; 41 const uint16_t vh3 = i[3]; 42 i += 4; 43 44 const uint32_t vw0 = (uint32_t) vh0 << 16; 45 const uint32_t vw1 = (uint32_t) vh1 << 16; 46 const uint32_t vw2 = (uint32_t) vh2 << 16; 47 const uint32_t vw3 = (uint32_t) vh3 << 16; 48 49 const uint32_t vsign0 = vw0 & vsign_mask; 50 const uint32_t vsign1 = vw1 & vsign_mask; 51 const uint32_t vsign2 = vw2 & vsign_mask; 52 const uint32_t vsign3 = vw3 & vsign_mask; 53 54 const uint32_t v2w0 = vw0 + vw0; 55 const uint32_t v2w1 = vw1 + vw1; 56 const uint32_t v2w2 = vw2 + vw2; 57 const uint32_t v2w3 = vw3 + vw3; 58 59 const uint32_t vnorm0 = float_as_uint32(uint32_as_float((v2w0 >> 4) + vexp_offset) * vexp_scale); 60 const uint32_t vnorm1 = float_as_uint32(uint32_as_float((v2w1 >> 4) + vexp_offset) * vexp_scale); 61 const uint32_t vnorm2 = float_as_uint32(uint32_as_float((v2w2 >> 4) + vexp_offset) * vexp_scale); 62 const uint32_t vnorm3 = float_as_uint32(uint32_as_float((v2w3 >> 4) + vexp_offset) * vexp_scale); 63 64 const uint32_t vdenorm0 = float_as_uint32(uint32_as_float((v2w0 >> 17) | vmagic_mask) - vmagic_bias); 65 const uint32_t vdenorm1 = float_as_uint32(uint32_as_float((v2w1 >> 17) | vmagic_mask) - vmagic_bias); 66 const uint32_t vdenorm2 = float_as_uint32(uint32_as_float((v2w2 >> 17) | vmagic_mask) - vmagic_bias); 67 const uint32_t vdenorm3 = float_as_uint32(uint32_as_float((v2w3 >> 17) | vmagic_mask) - vmagic_bias); 68 69 const uint32_t vf0 = vsign0 | (XNN_UNPREDICTABLE(v2w0 < vdenorm_cutoff) ? vdenorm0 : vnorm0); 70 const uint32_t vf1 = vsign1 | (XNN_UNPREDICTABLE(v2w1 < vdenorm_cutoff) ? vdenorm1 : vnorm1); 71 const uint32_t vf2 = vsign2 | (XNN_UNPREDICTABLE(v2w2 < vdenorm_cutoff) ? vdenorm2 : vnorm2); 72 const uint32_t vf3 = vsign3 | (XNN_UNPREDICTABLE(v2w3 < vdenorm_cutoff) ? vdenorm3 : vnorm3); 73 74 o[0] = vf0; 75 o[1] = vf1; 76 o[2] = vf2; 77 o[3] = vf3; 78 o += 4; 79 } 80 if XNN_UNLIKELY(n != 0) { 81 do { 82 const uint16_t vh = *i++; 83 84 const uint32_t vw = (uint32_t) vh << 16; 85 const uint32_t vsign = vw & vsign_mask; 86 const uint32_t v2w = vw + vw; 87 const uint32_t vnorm = float_as_uint32(uint32_as_float((v2w >> 4) + vexp_offset) * vexp_scale); 88 const uint32_t vdenorm = float_as_uint32(uint32_as_float((v2w >> 17) | vmagic_mask) - vmagic_bias); 89 const uint32_t vf = vsign | (XNN_UNPREDICTABLE(v2w < vdenorm_cutoff) ? vdenorm : vnorm); 90 91 *o++ = vf; 92 93 n -= sizeof(uint16_t); 94 } while (n != 0); 95 } 96 } 97