Response to JavaScript/Emscripten Skinning Benchmarks by Chad Austin

gistfile1.sh

# A response to
# http://chadaustin.me/2011/11/digging-into-javascript-performance-part-2/

# C++

$ g++ -Wall -O2 -o skinning_test_no_simd skinning_test_no_simd.cpp

$ ./skinning_test_no_simd
Skinned vertices per second: 19730000, blah=0.000000

# Unoptimized Emscripten

$ python tools/emmaken.py skinning_test_no_simd.cpp -o skinning.bc

$ python emscripten.py skinning_test_no_simd.bc > skinning.js

$ ~/Dev/mozilla-central/js/src/js -m -n skinning.js
Skinned vertices per second: 116279, blah=0.000000

# Optimized Emscripten

$ python tools/emmaken.py skinning_test_no_simd.cpp -o skinning.bc

$ python emscripten.py -s CORRECT_SIGNS=0 -s CORRECT_OVERFLOWS=0 -s CORRECT_ROUNDINGS=0 -s OPTIMIZE=1 -s RELOOP=1 -s QUANTUM_SIZE=1 -s USE_TYPED_ARRAYS=1 -s ASSERTIONS=0 -s INIT_STACK=0 --optimize skinning.bc > skinning.js

$ java -Xmx1024m -jar ~/Dev/closure-compiler-read-only/build/compiler.jar --compilation_level ADVANCED_OPTIMIZATIONS --js skinning.js --js_output_file skinning.opt.js --formatting PRETTY_PRINT

$ ~/Dev/mozilla-central/js/src/js -m -n skinning.opt.js
Skinned vertices per second: 2365269, blah=0.000000

# Ratios:
# C++                        1.00
# Optimized Emscripten       8.34
# Unoptimized Emscripten   169.68 (used in the link at the top of this doc)
#
# This is slower than other Emscripten benchmarks due to this test being
# heavy on memory operations, which tends to be an issue even with typed
# arrays. (A similar thing happens in the 'raytrace' benchmark in the
# Emscripten benchmark suite, which is the slowest benchmark there.)
#
# Note that there are plenty of obvious optimizations not done in the
# generated code. We are working on fixing that, but meanwhile some
# projects using Emscripten are hand-optimizing their inner loops,
# for example Broadway (of course hand-optimizing inner loops will
# probably help even with the best automatic optimizer).
#
# We realize that generating fully optimized code with Emscripten is
# not trivial (as can be seen by the options used above), and we will
# improve that. Meanwhile, see the makefiles in projects using
# Emscripten (ammo.js, j2k.js, Broadway, etc.) for concrete examples
# and also the docs in
# https://github.com/kripken/emscripten/wiki/Optimizing-Code
# (slightly outdated, the makefiles are current).