Avoid resending lots of D3D state

Avoid resending lots of D3D state

This change uses trivial caching to determines whether to reset shaders, the viewport, and the currently set vertex declaration. It also caches the render target desc to avoid rereading that. Serial numbers are added to vertex and index buffers, so resending those can be avoided.

These changes can give a big speedup (30% has been measured) on simple content, particularly when used directly or through pepper/native client.

BUG=
TEST=bunch of pages using webgl

Committed: http://code.google.com/p/angleproject/source/detail?r=743

[dgkoch, 2011-09-01 21:02:59]

One question about the vertex declarations, but otherwise I think it looks good.

http://codereview.appspot.com/4964057/diff/1001/src/libGLESv2/Context.cpp
File src/libGLESv2/Context.cpp (right):

http://codereview.appspot.com/4964057/diff/1001/src/libGLESv2/Context.cpp#new...
src/libGLESv2/Context.cpp:3755: device->SetStreamSource(i,
attributes[i].vertexBuffer, attributes[i].offset, attributes[i].stride);
tab character?

http://codereview.appspot.com/4964057/diff/1001/src/libGLESv2/Context.cpp#new...
src/libGLESv2/Context.cpp:3783: mLastSetVDecl = entry->vertexDeclaration;
shouldn't we be using serials here (in case we get the same pointer for a new
declaration after freeing a previous one)?

[John Bauman, 2011-09-01 21:08:09]

http://codereview.appspot.com/4964057/diff/1001/src/libGLESv2/Context.cpp
File src/libGLESv2/Context.cpp (right):

http://codereview.appspot.com/4964057/diff/1001/src/libGLESv2/Context.cpp#new...
src/libGLESv2/Context.cpp:3755: device->SetStreamSource(i,
attributes[i].vertexBuffer, attributes[i].offset, attributes[i].stride);
On 2011/09/01 21:03:00, dgkoch wrote:
> tab character?
oops.

http://codereview.appspot.com/4964057/diff/1001/src/libGLESv2/Context.cpp#new...
src/libGLESv2/Context.cpp:3783: mLastSetVDecl = entry->vertexDeclaration;
On 2011/09/01 21:03:00, dgkoch wrote:
> shouldn't we be using serials here (in case we get the same pointer for a new
> declaration after freeing a previous one)?

We only free an entry in a cache to add a new decl, so in that case the new decl
will be mLastSetVDecl and we don't have to worry about it.

[dgkoch, 2011-09-01 23:19:35]

On 2011/09/01 21:08:09, John Bauman wrote:
> http://codereview.appspot.com/4964057/diff/1001/src/libGLESv2/Context.cpp
> File src/libGLESv2/Context.cpp (right):
> 
>
http://codereview.appspot.com/4964057/diff/1001/src/libGLESv2/Context.cpp#new...
> src/libGLESv2/Context.cpp:3755: device->SetStreamSource(i,
> attributes[i].vertexBuffer, attributes[i].offset, attributes[i].stride);
> On 2011/09/01 21:03:00, dgkoch wrote:
> > tab character?
> oops.
> 
>
http://codereview.appspot.com/4964057/diff/1001/src/libGLESv2/Context.cpp#new...
> src/libGLESv2/Context.cpp:3783: mLastSetVDecl = entry->vertexDeclaration;
> On 2011/09/01 21:03:00, dgkoch wrote:
> > shouldn't we be using serials here (in case we get the same pointer for a
new
> > declaration after freeing a previous one)?
> 
> We only free an entry in a cache to add a new decl, so in that case the new
decl
> will be mLastSetVDecl and we don't have to worry about it.

ok. LGTM then, provided you've tested it thoroughly..