Re: wined3d: fix a memory leak and a performance regression in gen_arbfp_ffp_shader

[Henri Verbeet]

2008/9/3 Jeff Zaroyko :
> Please ignore the previous patch sent two days ago, this one fixes a
> performance regression as well as removing the memory leak.
>
> -Jeff

Do you have some kind of benchmark data for this change? I don't think
allocating relatively large buffers like that on the stack is a good
idea. If it does make a lot of difference in terms of performance, it
probably makes sense to allocate the buffer in arbfp_alloc() instead
and just reuse it. In that case it probably also makes sense to do
that for the ARB and GLSL shader backends as well.

wined3d: fix a memory leak and a performance regression in gen_arbfp_ffp_shader

[Jeff Zaroyko]

On Wed, Sep 3, 2008 at 10:25 PM, Henri Verbeet  wrote:
> 2008/9/3 Jeff Zaroyko :
> > Please ignore the previous patch sent two days ago, this one fixes a
> > performance regression as well as removing the memory leak.
> >
> > -Jeff
>
> Do you have some kind of benchmark data for this change? I don't think
> allocating relatively large buffers like that on the stack is a good
> idea. If it does make a lot of difference in terms of performance, it
> probably makes sense to allocate the buffer in arbfp_alloc() instead
> and just reuse it. In that case it probably also makes sense to do
> that for the ARB and GLSL shader backends as well.

None as such since difference in performance is so large.  I can say
that without a doubt using HeapAlloc and HeapFree in this function
causes noticeable delays such that in Battlefield 1942 the framerate
is falls so much that it is unplayable, changing the allocation leaves
no trace of this.

wined3d: fix a memory leak and a performance regression in gen_arbfp_ffp_shader

[Alexandre Julliard]

"Jeff Zaroyko"  writes:
> None as such since difference in performance is so large.  I can say
> that without a doubt using HeapAlloc and HeapFree in this function
> causes noticeable delays such that in Battlefield 1942 the framerate
> is falls so much that it is unplayable, changing the allocation leaves
> no trace of this.

The first thing to try would be to get rid of the useless
HEAP_ZERO_MEMORY. If that's not enough you should profile it to find
where the bottleneck is.

wined3d: fix a memory leak and a performance regression in gen_arbfp_ffp_shader

[Stefan Dösinger]

The real problems are that the ffp description structure still contains 3
uninitialized padding bytes, and that a HeapFree is missing. I'm going to
send patches for those