cuda::experimental::places::localized_stats#
-
struct localized_stats#
Statistics describing how a localized allocation - or a dry-run evaluation of one - distributes a tensor over data places.
Produced by evaluate_localized_placement() and by localized_array (see localized_array::get_stats()). This is the returnable form of the report previously only printed to stderr under CUDASTF_LOCALIZED_ALLOC_STATS.
Public Functions
-
inline double accuracy() const#
Fraction of bytes local to their owner once ownership is quantized to placement blocks: exact on the analytic tiers, a sampled estimate on the fallback tier (see total_samples).
Public Members
-
size_t total_bytes = 0#
requested payload size in bytes
-
size_t vm_bytes = 0#
block-rounded virtual reservation size in bytes
-
size_t block_size = 0#
placement granularity in bytes
-
size_t nblocks = 0#
number of placement blocks
-
size_t nallocs = 0#
physical allocations after merging same-owner runs
-
size_t total_samples = 0#
Placement-evidence counters.
Units depend on how owners were computed: sampled tier — probes drawn / probes agreeing with the chosen owner; analytic tiers (exact runs, census majority) — total bytes / bytes placed on their true owner (closed form, so accuracy() is exact).
-
size_t matching_samples = 0#
-
::std::unordered_map<::std::string, size_t> bytes_per_place#
Bytes backed by each place, keyed by data_place::to_string()
-
::std::unordered_map<size_t, size_t> bytes_per_grid_index#
Bytes owned by each grid position, keyed by the position’s linear index (dim4::get_index of the pos4; friendlier than strings across FFI)
-
inline double accuracy() const#