This repo benchmarks various half-sampling implementations for uint8_t data. The target is currently shipping WebAssembly runtimes in browsers that have support for uint64_t types but no explicit SIMD instructions yet.
There are various implementations that handle multiple data elements packed into regular uint32_t or uint64_t elements. Some bit twiddling is required to stop the elements from interfering. Here are some handy references to the general ideas and some specific implementations:
- http://aggregate.org/SWAR/over.html (describes general ideas of SWAR - "SIMD Within a Register")
- http://aggregate.org/MAGIC (general twiddling algorithms)
- http://graphics.stanford.edu/~seander/bithacks.html (more bit twiddling stuff)
To compile/test wasm implementation in d8 (with baseline "liftoff compiler" disabled):
$ emcc -O3 test.c halfsample.c -o test.js
$ path/to/v8/out/x64.release/d8 --no-wasm-tier-up -no-liftoff --predictable test.js
To test in a browser, you can just output html from emcc, run a server in the folder and open the page in a browser:
$ emcc -O3 test.c halfsample.c -o test.html
The code can also run natively. I'm on MacOS and have the latest Xcode command line tools installed.
The WebAssembly compilers in current browsers don't do any sort of auto-vectorization, so to make it a somewhat fairer comparison I also tend to disable that for the native build:
$ clang -O3 -fno-vectorize -fno-slp-vectorize test.c halfsample.c -o test
$ ./test
There is also an optional WebAssembly SIMD implementation. It compiles successfully with the following command line, using the latest-upstream emscripten:
$ emcc -O3 -msimd128 -fno-vectorize -fno-slp-vectorize -DUSE_WASM_SIMD test.c halfsample.c -o test.js
The SIMD shift instructions aren't supported by the v7.7 branch of d8, but do work with the latest master build (and in Chrome Canary).
$ path/to/v8/out/x64.release/d8 --no-wasm-tier-up -no-liftoff --predictable --experimental-wasm-simd test.js