Sign in to follow this  
RPTD

row-major driver bug confirmation

Recommended Posts

RPTD    340

I ran around in circles for a couple of weeks now trying to figure out why a certain change in my shader caused everything to go haywire. I've nailed now the problem and it looks to me like it's a driver bug not me understanding some specs wrong. I would like to confirm with others here if this is the case or if something slipped me.

 

I found this bug on ATI 12.6_beta_pre897:legacy as well as 13.1_beta_pre897:legacy . The graphic card is a Radeon 4870 HD.

 

The bug happens with uniform buffer objects in connection with row-major oriented matrices. The test shader looks like this (fragment shader only, vertex shader is basic in-out position):

 

 

precision highp float;
precision highp int;

layout (std140, row_major) uniform TestParameters{
    mat4x3 pMatrix1;
    mat3 pMatrix2;
};

out vec3 outColor;

void main( void ){
    mat4x3 indirectAccess = pMatrix1;
    vec3 normal = indirectAccess * vec4( vec3( 0.0, 0.0, 1.0 ), 1.0 );
    outColor = normal * vec3( 0.5 ) + vec3( 0.5 );
}

 

This test is run on a 1x1 screen so the result can be easily verfied by testing the result of a single pixel. The UBO in question is laid out like this with each [X] being one float value aligned in rows to 4 floats for readability:

 

 

[ a11 ][ a12 ][ a13 ][ a14 ]
[ a21 ][ a22 ][ a23 ][ a24 ]
[ a31 ][ a32 ][ a33 ][ a34 ]
[ b11 ][ b12 ][ b13 ][ 0 ]
[ b21 ][ b22 ][ b23 ][ 0 ]
[ b31 ][ b32 ][ b33 ][ 0 ]

 

aXY refers to matrix1[X][Y] and bXY to matrix2[X][Y] both in row-major order. For testing the following values are written in the UBO:

 

 

[ 0 ][ 0 ][ 0 ][ 0 ]
[ 0 ][ 0 ][ 0 ][ 0 ]
[ 0 ][ 0 ][ 0 ][ 1 ]
[ 1 ][ 0 ][ 0 ][ 0 ]
[ 0 ][ 0 ][ 0 ][ 0 ]
[ 0 ][ 0 ][ 0 ][ 0 ]

 

You can verify that if the data is read as row-major you end up with a matrix producing the output vector (0,0,1) while in colon-major you end up with (1,0,0).

 

If I run this shader I should get (0,0,1) but I get actually (1,0,0) hence the matrix is incorrectly read as colon-major if copied first to a matrix of the same type ( indirectAccess == pMatrix1 ). If the shader is changed to this (using pMatrix1 directly):

 

 

precision highp float;
precision highp int;

layout (std140, row_major) uniform TestParameters{
    mat4x3 pMatrix1;
    mat3 pMatrix2;
};

out vec3 outColor;

void main( void ){
    vec3 normal = pMatrix1 * vec4( vec3( 0.0, 0.0, 1.0 ), 1.0 );   // ******************
    outColor = normal * vec3( 0.5 ) + vec3( 0.5 );
}

 

the output is (0,0,1) as it should be. I ended up with this bug as I had to multiply two matrices together in the shader to do a bunch of calculations. The result never had been right because the driver reads the matrix incorrectly from the UBO data storage.

 

So my questions are not these:

 

1) Can somebody else confirm this bug using this simple test-case?

 

2) Or do I misunderstand something about the specs stating that accessing a matrix directly uses the row-major qualifier while using the matrix indirectly ignores it?

 

EDIT:
I should note that UBO is working properly so far. The problem started when I tried doing a more complex shader where I had to access matrices indirectly. So far I wrote all my shaders to access matrices directly so the bug never showed before.

Edited by RPTD

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

Sign in to follow this