[Mesa-dev] [PATCH 3/3] mesa: Only flag _NEW_TEXTURE on texture updates when a texture changes.
Brian Paul
brianp at vmware.com
Thu Feb 16 16:27:25 PST 2012
On 02/16/2012 04:40 PM, Eric Anholt wrote:
> There are three reasons left for this function flagging it:
> Re-computing completeness (and therefore the number of levels),
> binding a new current texture object, or you're doing some sort of
> fixed function (and nobody cares). For other cases, like rebinding a
> shader, we no longer trigger the driver recomputing every piece of
> texture state.
>
> Improves a VS state change microbenchmark by 14.8284% +/- 1.02% (n=10)
> on gen7. No statistically significant performance difference on
> 640x480 nexuiz (n=78).
>
> v2: Remove obsoleted comment, fix state flagging on transition to
> !_ReallyEnabled by unreffing the object (which seems like a good
> idea anyway).
> ---
>
> The unreffing appears to have cleared up my hang. I'm thinking about
> how to write a testcase.
>
> src/mesa/main/texstate.c | 36 ++++++++++++++++++++++++++----------
> 1 files changed, 26 insertions(+), 10 deletions(-)
>
> diff --git a/src/mesa/main/texstate.c b/src/mesa/main/texstate.c
> index 641e1f8..8c811d7 100644
> --- a/src/mesa/main/texstate.c
> +++ b/src/mesa/main/texstate.c
> @@ -592,9 +592,6 @@ update_texture_state( struct gl_context *ctx )
> * FINISHME: here.
> */
>
> - /* TODO: only set this if there are actual changes */
> - ctx->NewState |= _NEW_TEXTURE;
> -
> /*
> * Update texture unit state.
> */
> @@ -604,6 +601,7 @@ update_texture_state( struct gl_context *ctx )
> GLbitfield enabledFragTargets = 0x0;
> GLbitfield enabledTargets = 0x0;
> GLuint texIndex;
> + struct gl_texture_object *oldObj = texUnit->_Current;
>
> /* Get the bitmask of texture target enables.
> * enableBits will be a mask of the TEXTURE_*_BIT flags indicating
> @@ -625,10 +623,18 @@ update_texture_state( struct gl_context *ctx )
>
> enabledTargets = enabledVertTargets | enabledFragTargets;
>
> + /* This is not flagged with _NEW_TEXTURE, because the test for changing of
> + * the _Current texture object will cover every case where _ReallyEnabled
> + * could have changed.
> + */
> texUnit->_ReallyEnabled = 0x0;
>
> if (enabledTargets == 0x0) {
> /* neither vertex nor fragment processing uses this unit */
> + if (texUnit->_Current) {
> + _mesa_reference_texobj(&texUnit->_Current, NULL);
> + ctx->NewState |= _NEW_TEXTURE;
> + }
> continue;
> }
>
> @@ -642,6 +648,10 @@ update_texture_state( struct gl_context *ctx )
> if (enabledTargets& (1<< texIndex)) {
> struct gl_texture_object *texObj = texUnit->CurrentTex[texIndex];
> if (!texObj->_Complete) {
> + /* While _Complete is only consumed by us, this call also updates
> + _MaxLevel, which drivers do need to see.
> + */
> + ctx->NewState |= _NEW_TEXTURE;
> _mesa_test_texobj_completeness(ctx, texObj);
> }
> if (texObj->_Complete) {
> @@ -668,22 +678,26 @@ update_texture_state( struct gl_context *ctx )
> _mesa_reference_texobj(&texUnit->_Current, texObj);
> texUnit->_ReallyEnabled = 1<< texTarget;
> }
> - else {
> - /* fixed-function: texture unit is really disabled */
> - continue;
> - }
> }
> +
> + if (oldObj != texUnit->_Current)
> + ctx->NewState |= _NEW_TEXTURE;
> }
>
>
> /* Determine which texture coordinate sets are actually needed */
> if (fprog) {
> const GLuint coordMask = (1<< MAX_TEXTURE_COORD_UNITS) - 1;
> + GLuint new_coord_units = ((fprog->InputsRead>> FRAG_ATTRIB_TEX0)&
> + coordMask);
I have a feeling that MSVC is going to generate a warning here since
InpusRead is a 64-bit field but new_coord_units is 32 bits. If so, I
can follow-up with a fix.
> /* Note that this gets consumed by update_ffvs_texture_state(). */
> - ctx->Texture._EnabledCoordUnits
> - = (fprog->InputsRead>> FRAG_ATTRIB_TEX0)& coordMask;
> + if (ctx->Texture._EnabledCoordUnits != new_coord_units) {
> + ctx->NewState |= _NEW_TEXTURE;
> + ctx->Texture._EnabledCoordUnits = new_coord_units;
> + }
> }
> else {
> + ctx->NewState |= _NEW_TEXTURE;
> update_fffs_texture_state(ctx);
> }
>
> @@ -691,8 +705,10 @@ update_texture_state( struct gl_context *ctx )
> * program. Those state flags are only used in the case of fixed function
> * vertex shading, in the tnl pipeline or ff_vertex_shader.
> */
> - if (!vprog)
> + if (!vprog) {
> + ctx->NewState |= _NEW_TEXTURE;
> update_ffvs_texture_state(ctx);
> + }
> }
>
>
Reviewed-by: Brian Paul <brianp at vmware.com>
More information about the mesa-dev
mailing list