Question

I’m trying to create a small ffmpeg “hack” which enables parallel executin of the yadif filter.

I think I have found a solution, however there can only be one concurrent instance of it. This is because the “scalable_yadif_context” is local to the function “scalable_yadif_filter_line1” which replaces the original yadif “filter_line” function. I could make the “scalable_yadif_context” thread local, however since this function is called often it would have a quite high overhead.

Any ideas as how to solve this issue?

// We need the context description in order to access the original filter_line function. Just redefine it here and hope that it is not changed inside of libavfilter.
typedef struct {
    int mode;
    int parity;
    int frame_pending;
    int auto_enable;
    AVFilterBufferRef *cur;
    AVFilterBufferRef *next;
    AVFilterBufferRef *prev;
    AVFilterBufferRef *out;
    void (*filter_line)(uint8_t *dst,
                        uint8_t *prev, uint8_t *cur, uint8_t *next,
                        int w, int prefs, int mrefs, int parity, int mode);
    const AVPixFmtDescriptor *csp;
} YADIFContext;

struct scalable_yadif_context
{
    std::vector<std::function<void()>> calls;
    int end_prefs;

    scalable_yadif_context() : end_prefs(std::numeric_limits<int>::max()){}
};

void (*org_yadif_filter_line)(uint8_t *dst, uint8_t *prev, uint8_t *cur, uint8_t *next, int w, int prefs, int mrefs, int parity, int mode) = 0;

void scalable_yadif_filter_line(scalable_yadif_context& ctx, uint8_t *dst, uint8_t *prev, uint8_t *cur, uint8_t *next, int w, int prefs, int mrefs, int parity, int mode)
{
    if(ctx.end_prefs == std::numeric_limits<int>::max())
        ctx.end_prefs = -prefs;  // Last call to filter_line will have negative pref

    ctx.calls.push_back([=]
    {
        org_yadif_filter_line(dst, prev, cur, next, w, prefs, mrefs, parity, mode);
    });    

    if(prefs == ctx.end_prefs)
    {       
        tbb::parallel_for(tbb::blocked_range<size_t>(0, ctx.calls.size()), [=](const tbb::blocked_range<size_t>& r)
        {
            for(auto n = r.begin(); n != r.end(); ++n)
                ctx.calls[n]();
        });
        ctx.calls.clear();
        ctx.end_prefs = std::numeric_limits<int>::max();
    }
}

void scalable_yadif_filter_line1(uint8_t *dst, uint8_t *prev, uint8_t *cur, uint8_t *next, int w, int prefs, int mrefs, int parity, int mode)
{
     // local to the current function, making this thread local would incur heavy overhead.
    static scalable_yadif_context ctx;
    scalable_yadif_filter_line(ctx, dst, prev, cur, next, w, prefs, mrefs, parity, mode);
}

void make_scalable_yadif(AVFilterContext* ctx)
{
    YADIFContext* yadif = (YADIFContext*)ctx->priv;

    // Data race should not be problem since we are always writing the same value
    org_yadif_filter_line = yadif->filter_line;

    // hmm, will only work for one concurrent instance... 
    // I need a unique "scalable_yadif_filter_line1" for each call...
    yadif->filter_line = scalable_yadif_filter_line1; 
}

I’ve created an extremely ugly solution that works for up to 18 concurrent instances.

#define RENAME(a) f ## a

#define ff(x) \
void RENAME(x)(uint8_t *dst, uint8_t *prev, uint8_t *cur, uint8_t *next, int w, int prefs, int mrefs, int parity, int mode) \
{\
    static scalable_yadif_context ctx;\
    scalable_yadif_filter_line(ctx, dst, prev, cur, next, w, prefs, mrefs, parity, mode);\
}

ff(0); ff(1); ff(2); ff(3); ff(4); ff(5); ff(6); ff(7); ff(8); ff(9); ff(10); ff(11); ff(12); ff(13); ff(14); ff(15); ff(16); ff(17);

void (*fs[])(uint8_t *dst, uint8_t *prev, uint8_t *cur, uint8_t *next, int w, int prefs, int mrefs, int parity, int mode) = 

{f0, f1, f2, f3, f4, f5, f6, f7, f8, f9, f10, f11, f12, f13, f14, f15, f16, f17};

namespace caspar {

void init()
{
    for(int n = 0; n < 18; ++n)
        tags.push(n);
}

int make_scalable_yadif(AVFilterContext* ctx)
{
    static boost::once_flag flag = BOOST_ONCE_INIT;
    boost::call_once(&init, flag);

    YADIFContext* yadif = (YADIFContext*)ctx->priv;
    org_yadif_filter_line = yadif->filter_line;

    int tag;
    if(!tags.try_pop(tag))
    {
        LOG(warning) << "Not enough scalable-yadif instances. Running non-scalable";
        return -1;
    }

    yadif->filter_line = fs[tag];
    return tag;
}

void release_scalable_yadif(int tag)
{
    if(tag != -1)
        tags.push(tag);
}

Answer 1

Why not just pass a per-thread buffer to the scalable_yadif_filter_line1 function? It might require a bit of reorganization, but it’s far better than using statics or thread-locals (after all, what happens to your thread local buffer if the thread goes on to do something else?)

If you cannot pass a buffer to the function (due to a fixed ffmpeg API), TLS is probably your only choice. Overhead isn’t as bad as you might think, but it’s still not that good. I’d highly recommend looking into modifying ffmpeg to add a context parameter.