Issue 639042: Review: runtime optimizations, part 2 (Closed)
DescriptionSignificant additional work on "runtime code optimization" introduced recently. Quick run-down:
* New constant folding for additional ops: neg, aref (was previously commented out due to bugs), compref, strlen, endswith, concat, clamp, sqrt, matrix, transform[nv].
* All previous optimizations had been ones where no *new* constant values were needed (we could do new constant symbols, but only using VALUES that were stored statically or were the defaults of parameters stored in the master). Now we can make up brand new constants with new values, which are stored in the ShadingSystem via the helper class ConstantPool. This allows add, sub, mul, and div where both operands are constant will just turn into a single constant, as well as several of the other new constant folding routines to work.
* Refactor to eliminate the if-else-else-else of constant folding routine selection, in favor of a hash map acting as a dispatch table.
* Allow "local aliasing" within basic blocks -- previously, "A=const" would remember that A had the constant value (so subsequent C=A*B, say, would treat A as a constant) but only if the statement was in the global scope and A was never subsequently reassigned. Now we can remember that A has that value, but just for the basic block it's in and/or until it's reassigned. This allows lots of additional constant folding. When each basic block is done, its local aliases are forgotten.
* Some preliminary groundwork for tracking which params will be used by downstream layers, so I can take advantage of that layer. (Apologies for the incomplete nature of that.)
* Add -O0, -O1, -O2 optional flags to "testshade" to specify level of runtime optimization.
* New statistics track how many actual instructions we execute over a render, and how much total time we spend doing the runtime optimization.
* A couple minor renames and cleanups I promised Chris I'd do in the last review, but somehow botched the checkin and didn't get them all committed.
* IMPORTANT bug fix to the last checkin, in constfold_add -- lines 340 and 347 got the argument numbers mixed up, resulting in turning "A = B + 0" into "A=0" rather than "A=B". Oops!
So, the net result of all this is summarized by the following stats from a production asset we like to use as a benchmark:
optimize=1 5:35 2046 MB peak 1,806,370,431 instructions run
optimize=2 4:32 1403 MB peak 1,046,991,997 instructions run
As for how we can reduce the number of instructions executed by almost 45% but only reduce runtime by 20%, let's just say there are other bottlenecks in our renderer we're working on and interpreter overhead is obviously not the major (or at least not the sole) bottleneck.
Patch Set 1 #
Total comments: 6
MessagesTotal messages: 5
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
