I am trying to understand the meaning of a set of performance counters: l2_subp queries, misses and hits.
Is (l2_subp0_read_sector_queries + l2_subp1_read_sector_queries) the total number of 32 bit memory
requests made to L2 cache ?
If this is correct I would assume that:
i.e. the total number of L2 cache requests should be equal to the hits plus the misses.
In the application I am working on the relationship does not hold, the two members of the
equation are close but not equal.
Is this simply due to uncertainty during profiling or the two numbers should not be
the same (in which case I do not understand the meaning of the profiler counter) ?
Listing the counters of interest and their description -
l2_subp0_read_sector_queries- Number of read requests from L1 to slice 0 of L2 cache. This increments by 1 for each 32-byte access.
l2_subp0_read_hit_sectors- Number of read requests from L1 that hit in slice 0 of L2 cache. This increments by 1 for each 32-byte access.
l2_subp0_read_sector_misses- Number of read misses in slice 0 of L2 cache. This increments by 1 for each 32-byte access.
l2_subp0_total_read_sector_queries - Total read requests to slice 0 of L2 cache. This includes requests from L1, Texture cache, system memory. This increments by 1 for each 32-byte access.
*_read_sector_queries and *_read_hit_sectors are read requests from L1 to L2 while *_read_sector_misses and *_total_read_sector_queries are from all units to L2. So we really can not establish the relation mentioned in the previous comment.
Equality can be replaced with “less than or equal to” to make it correct -
l2_subp0_read_sector_queries + l2_subp1_read_sector_queries <=
l2_subp0_read_sector_misses + l2_subp1_read_sector_misses +
l2_subp0_read_hit_sectors + l2_subp1_read_hit_sectors