portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-04 03:05:55 +01:00
Code Releases Activity

Move sfw IEEE fp16 into central location

Browse Source
This commit is contained in:
paboyle 2018-06-13 20:22:01 +01:00
parent 8287ed8383
commit 847c761ccc
1 changed files with 0 additions and 67 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

67
lib/simd/Grid_sse4.h
View File

@ -333,73 +333,6 @@ struct Permute{
#ifdef SFW_FP16
struct Grid_half {
Grid_half(){}
Grid_half(uint16_t raw) : x(raw) {}
uint16_t x;
};
union FP32 {
unsigned int u;
float f;
};
// PAB - Lifted and adapted from Eigen, which is GPL V2
inline float sfw_half_to_float(Grid_half h) {
const FP32 magic = { 113 << 23 };
const unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift
FP32 o;
o.u = (h.x & 0x7fff) << 13; // exponent/mantissa bits
unsigned int exp = shifted_exp & o.u; // just the exponent
o.u += (127 - 15) << 23; // exponent adjust
// handle exponent special cases
if (exp == shifted_exp) { // Inf/NaN?
o.u += (128 - 16) << 23; // extra exp adjust
} else if (exp == 0) { // Zero/Denormal?
o.u += 1 << 23; // extra exp adjust
o.f -= magic.f; // renormalize
}
o.u |= (h.x & 0x8000) << 16; // sign bit
return o.f;
}
inline Grid_half sfw_float_to_half(float ff) {
FP32 f; f.f = ff;
const FP32 f32infty = { 255 << 23 };
const FP32 f16max = { (127 + 16) << 23 };
const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
unsigned int sign_mask = 0x80000000u;
Grid_half o;
o.x = static_cast<unsigned short>(0x0u);
unsigned int sign = f.u & sign_mask;
f.u ^= sign;
// NOTE all the integer compares in this function can be safely
// compiled into signed compares since all operands are below
// 0x80000000. Important if you want fast straight SSE2 code
// (since there's no unsigned PCMPGTD).
if (f.u >= f16max.u) { // result is Inf or NaN (all exponent bits set)
o.x = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf
} else { // (De)normalized number or zero
if (f.u < (113 << 23)) { // resulting FP16 is subnormal or zero
// use a magic value to align our 10 mantissa bits at the bottom of
// the float. as long as FP addition is round-to-nearest-even this
// just works.
f.f += denorm_magic.f;
// and one integer subtract of the bias later, we have our final float!
o.x = static_cast<unsigned short>(f.u - denorm_magic.u);
} else {
unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
// update exponent, rounding bias part 1
f.u += ((unsigned int)(15 - 127) << 23) + 0xfff;
// rounding bias part 2
f.u += mant_odd;
// take the bits!
o.x = static_cast<unsigned short>(f.u >> 13);
}
}
o.x |= static_cast<unsigned short>(sign >> 16);
return o;
}
static inline __m128i Grid_mm_cvtps_ph(__m128 f,int discard) {
__m128i ret=(__m128i)_mm_setzero_ps();
float *fp = (float *)&f;