Struct OccupancyFromPtx 

pub struct OccupancyFromPtx { /* private fields */ }

Parses register usage out of a PTX string and wires it into the CPU-side occupancy model.

Construct with OccupancyFromPtx::parse, then either inspect per-kernel register counts via OccupancyFromPtx::kernel or compute an occupancy estimate directly with OccupancyFromPtx::estimate_for.

Implementations

impl OccupancyFromPtx

pub fn parse(ptx: &str) -> CudaResult<Self>

Parse all kernels and their .reg directives out of a PTX module.

Register declarations are attributed to the most recently seen .entry. Directives that appear before any .entry (module-scope .reg, which PTX permits for globals) are ignored for occupancy purposes because they do not occupy per-thread registers.

Errors

Returns CudaError::InvalidValue if the PTX contains no .entry directive at all (there is nothing whose occupancy could be modelled).

pub fn kernels(&self) -> &[PtxKernel]

All kernels discovered in the PTX, in source order.

pub fn kernel(&self, name: &str) -> Option<&PtxKernel>

Look up a kernel by its entry name.

pub fn register_usage(&self, name: &str) -> Option<PtxRegisterUsage>

Register usage for the named kernel, or None if it is absent.

pub fn estimate_for( &self, kernel_name: &str, sm_major: u32, sm_minor: u32, block_size: u32, shared_memory: u32, ) -> CudaResult<OccupancyEstimate>

Estimate occupancy for one kernel on a synthetic device of the given compute capability, using the register count parsed from the PTX.

This is the offline equivalent of cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags: the parsed .reg count is substituted for the figure ptxas would otherwise supply.

Parameters

• kernel_name — entry name to model.
• sm_major / sm_minor — target compute capability (e.g. 9, 0).
• block_size — threads per block.
• shared_memory — dynamic shared memory per block in bytes.

Errors

Returns CudaError::NotFound if kernel_name is not present.

pub fn estimate_with_info( &self, kernel_name: &str, info: DeviceOccupancyInfo, block_size: u32, shared_memory: u32, ) -> CudaResult<OccupancyEstimate>

Estimate occupancy for one kernel against an explicit device description, for callers that already hold a DeviceOccupancyInfo.

Errors

Returns CudaError::NotFound if kernel_name is not present.

Trait Implementations

impl Clone for OccupancyFromPtx

fn clone(&self) -> OccupancyFromPtx

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for OccupancyFromPtx

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for OccupancyFromPtx

fn default() -> OccupancyFromPtx

Returns the “default value” for a type.