On Wed, Jul 08, 2015 at 02:00:02PM -0700, Matt Turner wrote: > By keeping a pointer to the next available location, we reduce the > number of memory accesses needed to write to the batchbuffer. > > A net ~7k reduction of .text size, 7.5k of which is from the change to > intel_batchbuffer_emit_dword(). > > text data bss dec hex filename > 4943740 195152 26192 5165084 4ed01c i965_dri.so before > 4936804 195152 26192 5158148 4eb504 i965_dri.so after > > Combined with the previous patch, improves performance of Synmark > OglBatch7 by 4.05914% +/- 1.49373% (n=270) on Haswell. > --- > Full disclosure: when testing on an IVB desktop, I measured a > regression in the same benchmark of -4.19005% +/- 1.15188% (n=30). > I don't have any explanation.
The problem is that it seems to generate worse code with multiple adjacent emit_dwords. I have seen similar regressions when doing the same batch[index] to *batch++ elsewhere. -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
