On 2009-09-08 19:47, Reimar D?ffinger wrote:
> On Tue, Sep 08, 2009 at 03:30:26PM +0200, Jimmy Christensen wrote:
>> + switch (s->bits_per_color_id) {
>> + // 32-bit
>> + case 2:
>> + avctx->pix_fmt = PIX_FMT_RGB48LE;
>> + break;
>> + // 16-bit
>> + case 1:
>> + avctx->pix_fmt = PIX_FMT_RGB48LE;
>> + break;
>
> I missed that, this obviously can't be right, you write the uint16_t
> values directly, thus in native format - not little-endian.
>
You're right. Fixed in rev17.
>> + for (x = 0; x< xdelta; x++) {
>> + *ptr_x++ = exr_flt2uint(bytestream_get_le32(&red_channel_buffer));
>> + *ptr_x++ = exr_flt2uint(bytestream_get_le32(&green_channel_buffer));
>> + *ptr_x++ = exr_flt2uint(bytestream_get_le32(&blue_channel_buffer));
>> + }
>
> Not that this must used code suitable for unaligned access, thus e.g. on
> Sparc
> *ptr_x++ = exr_flt2uint(le2me_32(*(uint32_t *)red_channel_buffer));
> will be faster.