On 12 September 2013 12:23, Stefan Dösinger stefan@codeweavers.com wrote:
* On the r500 driver on Windows D3DCMP_GREATER and D3DCMP_GREATEREQUAL are broken for depth* values > 0.5. The range appears to be distored, apparently an incoming value of ~0.875 is
"distorted"?
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Am 2013-09-12 12:44, schrieb Henri Verbeet:
On 12 September 2013 12:23, Stefan Dösinger stefan@codeweavers.com wrote:
* On the r500 driver on Windows D3DCMP_GREATER andD3DCMP_GREATEREQUAL are broken for depth + * values > 0.5. The range appears to be distored, apparently an incoming value of ~0.875 is
"distorted"?
Sorry, I wasn't able to find out what the driver is doing exactly. Everything seems ok when the value in the Z buffer and the incoming fragment's Z are 0.5 or smaller. At first I thought D3DCMP_GREATER is broken in general and behaves like D3DCMP_NEVER - that was when I investigated the failure in the d3d8 version of this test. On closer inspection it turned out that 0.5000001 is roughly equal to 0.0877, and 0.95 roughly matches the 0.75 we write into the buffer with the clear call. (The numbers are from my memory, they may be a bit off.)
D3DCMP_LESS works correctly. I think D3DCMP_GREATER also produces expected results when just comparing geometry Z vs geometry Z, so a game that uses it and doesn't depend on a match with the clear value should still work, maybe with reduced precision.
D3DCMP_GREATEREQUAL has the same bug as D3DCMP_GREATER. D3DCMP_LESSEQUAL works ok as well.
After two hours of trying to find out what the driver does I decided it's not worth the time and marked the tests broken. It is after all a no longer maintained Windows driver.
On 12 September 2013 13:33, Stefan Dösinger stefan@codeweavers.com wrote:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Am 2013-09-12 12:44, schrieb Henri Verbeet:
On 12 September 2013 12:23, Stefan Dösinger stefan@codeweavers.com wrote:
* On the r500 driver on Windows D3DCMP_GREATER andD3DCMP_GREATEREQUAL are broken for depth + * values > 0.5. The range appears to be distored, apparently an incoming value of ~0.875 is
"distorted"?
Sorry, I wasn't able to find out what the driver is doing exactly. Everything seems ok when the value in the Z buffer and the incoming fragment's Z are 0.5 or smaller. At first I thought D3DCMP_GREATER is broken in general and behaves like D3DCMP_NEVER - that was when I investigated the failure in the d3d8 version of this test. On closer inspection it turned out that 0.5000001 is roughly equal to 0.0877, and 0.95 roughly matches the 0.75 we write into the buffer with the clear call. (The numbers are from my memory, they may be a bit off.)
D3DCMP_LESS works correctly. I think D3DCMP_GREATER also produces expected results when just comparing geometry Z vs geometry Z, so a game that uses it and doesn't depend on a match with the clear value should still work, maybe with reduced precision.
D3DCMP_GREATEREQUAL has the same bug as D3DCMP_GREATER. D3DCMP_LESSEQUAL works ok as well.
After two hours of trying to find out what the driver does I decided it's not worth the time and marked the tests broken. It is after all a no longer maintained Windows driver.
I really just meant the typo in the comment.
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Am 2013-09-12 13:36, schrieb Henri Verbeet:
I really just meant the typo in the comment.
Ah yes, that should be "distorted".
-
Henri Verbeet -
Stefan Dösinger