diff options
| author | Timo Rothenpieler <timo@rothenpieler.org> | 2022-08-09 22:16:50 +0200 |
|---|---|---|
| committer | Timo Rothenpieler <timo@rothenpieler.org> | 2022-08-19 22:09:36 +0200 |
| commit | cb8ad005bb73b1adf0d36eeb794c4c375fd3ee12 (patch) | |
| tree | 2c8e552c20075b2da7dc536d139c3307026f163b /libavutil/half2float.h | |
| parent | b42925264a910e6807e9e7134feaa44ae47bf911 (diff) | |
| download | ffmpeg-cb8ad005bb73b1adf0d36eeb794c4c375fd3ee12.tar.gz | |
avutil/half2float: adjust conversion of NaN
IEEE-754 differentiates two different kind of NaNs.
Quiet and Signaling ones. They are differentiated by the MSB of the
mantissa.
For whatever reason, actual hardware conversion of half to single always
sets the signaling bit to 1 if the mantissa is != 0, and to 0 if it's 0.
So our code has to follow suite or fate-testing hardware float16 will be
impossible.
Diffstat (limited to 'libavutil/half2float.h')
| -rw-r--r-- | libavutil/half2float.h | 5 |
1 files changed, 5 insertions, 0 deletions
diff --git a/libavutil/half2float.h b/libavutil/half2float.h index 1f6deade07..5af4690cfe 100644 --- a/libavutil/half2float.h +++ b/libavutil/half2float.h @@ -45,6 +45,9 @@ static void half2float_table(uint32_t *mantissatable, uint32_t *exponenttable, mantissatable[i] = convertmantissa(i); for (int i = 1024; i < 2048; i++) mantissatable[i] = 0x38000000UL + ((i - 1024) << 13UL); + for (int i = 2048; i < 3072; i++) + mantissatable[i] = mantissatable[i - 1024] | 0x400000UL; + mantissatable[2048] = mantissatable[1024]; exponenttable[0] = 0; for (int i = 1; i < 31; i++) @@ -58,7 +61,9 @@ static void half2float_table(uint32_t *mantissatable, uint32_t *exponenttable, offsettable[0] = 0; for (int i = 1; i < 64; i++) offsettable[i] = 1024; + offsettable[31] = 2048; offsettable[32] = 0; + offsettable[63] = 2048; } static uint32_t half2float(uint16_t h, const uint32_t *mantissatable, const uint32_t *exponenttable, |