Fix T84167: Saving half-float EXR might result in NaN pixels

Clamp value to the -HALF_MAX .. HALF_MAX.

The non-clamped values were causing NaN and inf values saved to
the file, which was the root cause of glare node giving unexpected
result.

The nan/inf on overflow is something mentioned in the half data
type in OpenEXR header.

Differential Revision: https://developer.blender.org/D10105
This commit is contained in:
Sergey Sharybin 2021-01-13 14:20:26 +01:00
parent 74f6b81c12
commit 53bd58993e
Notes: blender-bot 2023-02-14 01:57:12 +01:00
Referenced by issue #84167, Compositor: Glare Node artifact in OpenEXR file with half float color depth
1 changed files with 10 additions and 5 deletions

View File

@ -324,6 +324,11 @@ struct _RGBAZ {
using RGBAZ = _RGBAZ;
static half float_to_half_safe(const float value)
{
return half(clamp_f(value, -HALF_MAX, HALF_MAX));
}
extern "C" {
/**
@ -472,10 +477,10 @@ static bool imb_save_openexr_half(ImBuf *ibuf, const char *name, const int flags
from = ibuf->rect_float + channels * i * width;
for (int j = ibuf->x; j > 0; j--) {
to->r = from[0];
to->g = (channels >= 2) ? from[1] : from[0];
to->b = (channels >= 3) ? from[2] : from[0];
to->a = (channels >= 4) ? from[3] : 1.0f;
to->r = float_to_half_safe(from[0]);
to->g = float_to_half_safe((channels >= 2) ? from[1] : from[0]);
to->b = float_to_half_safe((channels >= 3) ? from[2] : from[0]);
to->a = float_to_half_safe((channels >= 4) ? from[3] : 1.0f);
to++;
from += channels;
}
@ -1116,7 +1121,7 @@ void IMB_exr_write_channels(void *handle)
float *rect = echan->rect;
half *cur = current_rect_half;
for (size_t i = 0; i < num_pixels; i++, cur++) {
*cur = rect[i * echan->xstride];
*cur = float_to_half_safe(rect[i * echan->xstride]);
}
half *rect_to_write = current_rect_half + (data->height - 1L) * data->width;
frameBuffer.insert(