Computing the maximum value in an array is slow with SIMD.

[juj]

Here's a commit that tests using SSE1 to compute the maximum value in a large array: juj@c62525e

SIMD support in Emscripten is still very early, but writing this down so that it can follow the development as it progresses and gets better.

Running natively on a Macbook Pro built with clang++ benchmark_sse1.cpp -O3 -o a, the results are

N: 16777216
Block scalar took 0.027743 msecs. Result: 1.000000.
Block scalar 4 unroll took 0.029196 msecs (1.052x of scalar) Result: 1.000000.
Block SSE1 no unroll took 0.004371 msecs (0.158x of scalar) Result: 1.000000.
Block SSE1 Unroll 2 took 0.005398 msecs (0.195x of scalar) Result: 1.000000.
Block SSE1 Unroll 4 took 0.004678 msecs (0.169x of scalar) Result: 1.000000.
Block SSE1 Unroll 4 pf took 0.003755 msecs (0.135x of scalar) Result: 1.000000.
Block SSE1 Unroll 16 took 0.003934 msecs (0.142x of scalar) Result: 1.000000.
Block SSE1 Unroll 16 pf took 0.004002 msecs (0.144x of scalar) Result: 1.000000.

Running in FF Nightly from today, 35.0a1 (2014-09-24), with the command line em++ -O3 tests/benchmark_sse1.cpp -o a.html -s TOTAL_MEMORY=268435456, the output is

N: 16777216
Block scalar took 0.032746 msecs. Result: 1.000000.
Block scalar 4 unroll took 0.030144 msecs (0.921x of scalar) Result: 1.000000.
Block SSE1 no unroll took 1.347598 msecs (41.153x of scalar) Result: 1.000000.
Block SSE1 Unroll 2 took 1.327801 msecs (40.548x of scalar) Result: 1.000000.
Block SSE1 Unroll 4 took 1.333699 msecs (40.728x of scalar) Result: 1.000000.
Block SSE1 Unroll 16 took 1.432141 msecs (43.734x of scalar) Result: 1.000000.

so we see that running natively, the SSE1 version is 5-7 times faster than scalar, whereas the Emscripten run is 40-43 times slower than scalar. For one, the JS version is not yet asm.js-validating, so slow performance is to be expected, but that's a good start!

[juj]

Testing on current incoming and Nightly from today, I'm seeing

N: 16777216
Block scalar took 0.014260 msecs. Result: 1.000000.
Block scalar 4 unroll took 0.009430 msecs (0.661x of scalar) Result: 1.000000.
Block SSE1 no unroll took 0.013255 msecs (0.930x of scalar) Result: 1.000000.
Block SSE1 Unroll 2 took 0.010540 msecs (0.739x of scalar) Result: 1.000000.
Block SSE1 Unroll 4 took 0.009025 msecs (0.633x of scalar) Result: 1.000000.
Block SSE1 Unroll 16 took 0.007760 msecs (0.544x of scalar) Result: 1.000000.

so using min & max roughly doubles performance. Much better than 40-43x from before. Running natively, I get about 7x performance with the SIMD version (and 2x compared to SIMD.js). Closing, since this is now in the correct ballpark at least.