[Bf-committers] Cycles CPU Kernel on MSVC *all versions*

[Brecht Van Lommel]

Hi Jürgen ,

BVH / triangle intersection is always going to take up a major part of
rendering time, but I managed to optimize it some. For MSVC the BVH
intersection is now twice as fast for e.g. the BMW scene, which gives
a 30% render time reduction overall. I also added the /Ox compiler
flag (which contains the other /O compiler flags you posted for the
benchmark), that helps maybe 2% here.

On Fri, Jun 7, 2013 at 9:32 AM, Jürgen Herrmann <shadowrom at me.com> wrote:
> Hi there,
>
> after the Benchmarking I did I decided to try analyze the diferences between MinGW and MSVC build in depth.
> First of all I ran several Benchmarks of math functions in both GCC an MSVC and found out:
> Amazing! they are both equally fast! So math functions isn't the bottleneck here.
> So I tried to get a grip on real data and tried out "Very Sleepy" ( free download at: http://www.codersnotes.com/sleepy ) to profile cycles on MSVC.
>
> What I found here was quite interresting:
> While rendering the most time is spent in ccl::bvh_intersect_instancing with 1292.15s (53.70%) O_o
> Within this call the most time is spent in ccl::bvh_node_intersect with 872.47s (36.26%) .
>
> So this is the part we'll have to look at ... I wonder how MinGW builds perform here, but I don't have a proper optimized build with debug info to compare :/
>
> /Jürgen
> _______________________________________________
> Bf-committers mailing list
> Bf-committers at blender.org
> http://lists.blender.org/mailman/listinfo/bf-committers