DSP_INIT can be called from multiple threads

[deven96]

I believe I ran into a case where multi-threaded programs get into VP8LDspInit concurrently and put the critical function pointers in a bad state which affects subsequent usage until program restart.

On this x86_64 system we have SSE and AVX2 available and ended up in a state where both VP8LPredictorsAdd[11] and VP8LPredictorsAdd_SSE[11] point to the same PredictorAdd11_AVX2 function leading to strange recursion and CPU saturation

A good case with a function pointers in a reasonable state looks like

sudo gdb -p ${pid} --batch \
-ex "set pagination off" \
-ex "print VP8LPredictorsAdd[11]" \
-ex "print VP8LPredictorsAdd_SSE[11]" \
-ex "print VP8LPredictorsAdd[11] == VP8LPredictorsAdd_SSE[11]" \
2>/dev/null

$1 = (VP8LPredictorsAddSubFunc) 0x55dc3dff7650 <PredictorAdd11_AVX2>
$2 = (VP8LPredictorsAddSubFunc) 0x55dc3dff8950 <PredictorAdd11_SSE2>
$3 = 0

Meanwhile a bad case looks like this

sudo gdb -p ${pid} --batch \
-ex "set pagination off" \
-ex "print VP8LPredictorsAdd[11]" \
-ex "print VP8LPredictorsAdd_SSE[11]" \
-ex "print VP8LPredictorsAdd[11] == VP8LPredictorsAdd_SSE[11]" \
2>/dev/null

$1 = (VP8LPredictorsAddSubFunc) 0x561e65724650 <PredictorAdd11_AVX2>
$2 = (VP8LPredictorsAddSubFunc) 0x561e65724650 <PredictorAdd11_AVX2>
$3 = 1

According to perf it seems to be stuck in PredictorAdd11_AVX2. So it seems to now call VP8LPredictorsAdd_SSE[11] for the remainder of the pixels which is just a pointer back to itself. I see that there exists the flag WEBP_USE_THREAD here supposed to enable threading support by enabling the pthread_mutex path during initialization.