MALI T760MP4 L2 cache hit ratio low (64%)

GPU caches are relatively small (an MP4 probably has only 256KB L2 cache), so for a data plane workload where the data size is large relative to the cache a significant proportion of cache misses are expected (i.e. every cache access will miss at least once). Note that you may have multiple L1 hits per L2 line load, so your effective hit rate from the shader core load instruction point of view may be much higher than this.

It's hard to give specific advice on data repacking without knowing specifics of the algorithm - but in general making your data types as narrow as possible, avoiding redundant data interleaved with used data, and ensuring good temporal reuse locality are all important.

For compute workloads how you split your work groups can also impact access patterns, so it's not always only about the data - the work group partitioning defines the data access pattern and the data layout defines how that maps to memory structures.

HTH,
Pete

GPU caches are relatively small (an MP4 probably has only 256KB L2 cache), so for a data plane workload where the data size is large relative to the cache a significant proportion of cache misses are expected (i.e. every cache access will miss at least once). Note that you may have multiple L1 hits per L2 line load, so your effective hit rate from the shader core load instruction point of view may be much higher than this.

It's hard to give specific advice on data repacking without knowing specifics of the algorithm - but in general making your data types as narrow as possible, avoiding redundant data interleaved with used data, and ensuring good temporal reuse locality are all important.

For compute workloads how you split your work groups can also impact access patterns, so it's not always only about the data - the work group partitioning defines the data access pattern and the data layout defines how that maps to memory structures.

HTH,
Pete