Fwd: Patchwatcher: failed regression tests: d3dx8: replace loops with explicit computations to make them faster [PATCH] d3dx8: replace loops by explicit computations
Patchwatcher notes that the conformance test fails here:
../../../tools/runtest -q -P wine -M d3dx8.dll -T ../../.. -p d3dx8_test.exe.so math.c && touch math.ok math.c:442: Test failed: Expected matrix= (-0.214800,1.311600,0.475200,0.000000 0.950400,-0.883600,0.924400,0.000000 1.021200,0.193600,-1.358800,0.000000 18.298500,-29.624001,15.683499,1.000000 )
Got matrix= (-0.214800,1.311600,0.475200,0.000000 0.950401,-0.883601,0.924400,0.000000 1.021204,0.193593,-1.358803,0.000000 18.298538,-29.624159,15.683425,1.000000)
Looks like rounding errors, but one way or another, you have to take care of that.
Also: did you benchmark this change to verify it actually made things faster? Unrolling loops is no longer as sure-fire a win as it used to be. I'm inclined to prefer the more compact, clear loop unless there's hard evidence it's > 1% slower. - Dan
---------- Forwarded message ---------- From: patchwatcher@kegel.com Date: Fri, Aug 22, 2008 at 8:35 PM Subject: Patchwatcher: failed regression tests: d3dx8: replace loops with explicit computations to make them faster [PATCH] d3dx8: replace loops by explicit computations To: dank@kegel.com
Hi! This is the experimental automated wine patchwatcher thingy. The latest git sources were built and tested with your patch "d3dx8: replace loops with explicit computations to make them faster [PATCH] d3dx8: replace loops by explicit computations" Result: the patch failed regression tests.
You can retrieve the full build results at http://kegel.com/wine/patchwatcher/results/264.log and see the patch as parsed at http://kegel.com/wine/patchwatcher/results/264.txt See http://kegel.com/wine/patchwatcher/results for more info.
On Sat, Aug 23, 2008 at 8:05 AM, Dan Kegel dank@kegel.com wrote:
Patchwatcher notes that the conformance test fails here:
Also: did you benchmark this change to verify it actually made things faster? Unrolling loops is no longer as sure-fire a win as it used to be. I'm inclined to prefer the more compact, clear loop unless there's hard evidence it's > 1% slower.
- Dan
I'de bet good money that GCC already unrolled those loops with -O2 and greater. I do agree with Dan though that doing a test on a standard intel core machine and a k8 machine to see what reality looks like is best.
You can't just declare a patch series for the last patch, as you seem to have done here:
d3dx8, test [patch 3/3] test the relative error instead of the absolute one to avoid rounded errors
That confuses patchwatcher (and maybe even humans).
That patch should have come first, to avoid having the tree in broken state. - Dan
On Sat, Aug 23, 2008 at 5:05 AM, Dan Kegel dank@kegel.com wrote:
Patchwatcher notes that the conformance test fails here:
../../../tools/runtest -q -P wine -M d3dx8.dll -T ../../.. -p d3dx8_test.exe.so math.c && touch math.ok math.c:442: Test failed: Expected matrix= (-0.214800,1.311600,0.475200,0.000000 0.950400,-0.883600,0.924400,0.000000 1.021200,0.193600,-1.358800,0.000000 18.298500,-29.624001,15.683499,1.000000 )
Got matrix= (-0.214800,1.311600,0.475200,0.000000 0.950401,-0.883601,0.924400,0.000000 1.021204,0.193593,-1.358803,0.000000 18.298538,-29.624159,15.683425,1.000000)
Looks like rounding errors, but one way or another, you have to take care of that.
Also: did you benchmark this change to verify it actually made things faster? Unrolling loops is no longer as sure-fire a win as it used to be. I'm inclined to prefer the more compact, clear loop unless there's hard evidence it's > 1% slower.
- Dan
---------- Forwarded message ---------- From: patchwatcher@kegel.com Date: Fri, Aug 22, 2008 at 8:35 PM Subject: Patchwatcher: failed regression tests: d3dx8: replace loops with explicit computations to make them faster [PATCH] d3dx8: replace loops by explicit computations To: dank@kegel.com
Hi! This is the experimental automated wine patchwatcher thingy. The latest git sources were built and tested with your patch "d3dx8: replace loops with explicit computations to make them faster [PATCH] d3dx8: replace loops by explicit computations" Result: the patch failed regression tests.
You can retrieve the full build results at http://kegel.com/wine/patchwatcher/results/264.log and see the patch as parsed at http://kegel.com/wine/patchwatcher/results/264.txt See http://kegel.com/wine/patchwatcher/results for more info.
-
Dan Kegel -
Zachary Goldberg