Skip to content

Commit

Permalink
Port the Float16 runtime functions to C.
Browse files Browse the repository at this point in the history 
The Julia runtime wasn't safe wrt. GC operations.
I left the commit in for archival purposes, in case
we want to revisit this again later.
  • Loading branch information
@maleadt
maleadt committed Oct 14, 2020
1 parent 3851ac5 commit 1e69b0a
Show file tree
Hide file tree
Showing 6 changed files with 204 additions and 309 deletions.
4 changes: 0 additions & 4 deletions base/Base.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -129,10 +129,6 @@ end
include(strcat((length(Core.ARGS)>=2 ? Core.ARGS[2] : ""), "build_h.jl")) # include($BUILDROOT/base/build_h.jl)
include(strcat((length(Core.ARGS)>=2 ? Core.ARGS[2] : ""), "version_git.jl")) # include($BUILDROOT/base/version_git.jl)

# run-time library
include("runtime/runtime.jl")
using .Runtime

# numeric operations
include("hashing.jl")
include("rounding.jl")
Expand Down
161 changes: 0 additions & 161 deletions base/runtime/runtime.jl
View file

This file was deleted.

197 changes: 197 additions & 0 deletions src/intrinsics.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1299,3 +1299,200 @@ static Value *emit_untyped_intrinsic(jl_codectx_t &ctx, intrinsic f, Value **arg
}
assert(0 && "unreachable");
}


// float16 intrinsics
// TODO: use LLVM's compiler-rt

static inline float half_to_float(uint16_t ival)
{
uint32_t sign = (ival & 0x8000) >> 15;
uint32_t exp = (ival & 0x7c00) >> 10;
uint32_t sig = (ival & 0x3ff) >> 0;
uint32_t ret;

if (exp == 0) {
if (sig == 0) {
sign = sign << 31;
ret = sign | exp | sig;
}
else {
int n_bit = 1;
uint16_t bit = 0x0200;
while ((bit & sig) == 0) {
n_bit = n_bit + 1;
bit = bit >> 1;
}
sign = sign << 31;
exp = ((-14 - n_bit + 127) << 23);
sig = ((sig & (~bit)) << n_bit) << (23 - 10);
ret = sign | exp | sig;
}
}
else if (exp == 0x1f) {
if (sig == 0) { // Inf
if (sign == 0)
ret = 0x7f800000;
else
ret = 0xff800000;
}
else // NaN
ret = 0x7fc00000 | (sign << 31) | (sig << (23 - 10));
}
else {
sign = sign << 31;
exp = ((exp - 15 + 127) << 23);
sig = sig << (23 - 10);
ret = sign | exp | sig;
}

float fret;
memcpy(&fret, &ret, sizeof(float));
return fret;
}

// float to half algorithm from:
// "Fast Half Float Conversion" by Jeroen van der Zijp
// ftp:https://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf
//
// With adjustments for round-to-nearest, ties to even.

static uint16_t basetable[512] = {
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0400, 0x0800, 0x0c00, 0x1000, 0x1400, 0x1800, 0x1c00, 0x2000,
0x2400, 0x2800, 0x2c00, 0x3000, 0x3400, 0x3800, 0x3c00, 0x4000, 0x4400, 0x4800, 0x4c00,
0x5000, 0x5400, 0x5800, 0x5c00, 0x6000, 0x6400, 0x6800, 0x6c00, 0x7000, 0x7400, 0x7800,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00, 0x7c00,
0x7c00, 0x7c00, 0x7c00, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8400, 0x8800, 0x8c00, 0x9000, 0x9400,
0x9800, 0x9c00, 0xa000, 0xa400, 0xa800, 0xac00, 0xb000, 0xb400, 0xb800, 0xbc00, 0xc000,
0xc400, 0xc800, 0xcc00, 0xd000, 0xd400, 0xd800, 0xdc00, 0xe000, 0xe400, 0xe800, 0xec00,
0xf000, 0xf400, 0xf800, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00,
0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00, 0xfc00};

static uint8_t shifttable[512] = {
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x18, 0x17, 0x16, 0x15, 0x14, 0x13, 0x12, 0x11, 0x10, 0x0f,
0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x0d, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19,
0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x19, 0x18, 0x17, 0x16, 0x15, 0x14, 0x13,
0x12, 0x11, 0x10, 0x0f, 0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x0d};

static inline uint16_t float_to_half(float param)
{
uint32_t f;
memcpy(&f, &param, sizeof(float));
if (isnan(param)) {
uint32_t t = 0x8000 ^ (0x8000 & ((uint16_t)(f >> 0x10)));
return t ^ ((uint16_t)(f >> 0xd));
}
int i = ((f & ~0x007fffff) >> 23);
uint8_t sh = shifttable[i];
f &= 0x007fffff;
// If `val` is subnormal, the tables are set up to force the
// result to 0, so the significand has an implicit `1` in the
// cases we care about.
f |= 0x007fffff + 0x1;
uint16_t h = (uint16_t)(basetable[i] + ((f >> sh) & 0x03ff));
// round
// NOTE: we maybe should ignore NaNs here, but the payload is
// getting truncated anyway so "rounding" it might not matter
int nextbit = (f >> (sh - 1)) & 1;
if (nextbit != 0 && (h & 0x7C00) != 0x7C00) {
// Round halfway to even or check lower bits
if ((h & 1) == 1 || (f & ((1 << (sh - 1)) - 1)) != 0)
h += UINT16_C(1);
}
return h;
}

#if !defined(_OS_DARWIN_) // xcode already links compiler-rt

extern "C" JL_DLLEXPORT float __gnu_h2f_ieee(uint16_t param)
{
return half_to_float(param);
}

extern "C" JL_DLLEXPORT float __extendhfsf2(uint16_t param)
{
return half_to_float(param);
}

extern "C" JL_DLLEXPORT uint16_t __gnu_f2h_ieee(float param)
{
return float_to_half(param);
}

extern "C" JL_DLLEXPORT uint16_t __truncdfhf2(double param)
{
return float_to_half((float)param);
}

#endif
6 changes: 6 additions & 0 deletions src/julia.expmap
View file
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,12 @@
environ;
__progname;

/* compiler run-time intrinsics */
__gnu_h2f_ieee;
__extendhfsf2;
__gnu_f2h_ieee;
__truncdfhf2;

local:
*;
};
2 changes: 1 addition & 1 deletion test/choosetests.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ in the `choices` argument:
"""
function choosetests(choices = [])
testnames = [
"subarray", "core", "compiler", "worlds", "runtime",
"subarray", "core", "compiler", "worlds",
"keywordargs", "numbers", "subtype",
"char", "strings", "triplequote", "unicode", "intrinsics",
"dict", "hashing", "iobuffer", "staged", "offsetarray",
Expand Down
Loading

0 comments on commit 1e69b0a

Please sign in to comment.