Operators

cuPQC-BigInt operators are used to describe a big-integer instance and configure its execution. They are combined via addition (+) to form a cuPQC-BigInt descriptor.

Description Operators

BitWidth

template<unsigned int BW>
BitWidth()

Sets the bit width BW of a big-integer instance. BW must be a multiple of 32 and greater than 0. The number of 32-bit limbs is BW / 32.

Not every width satisfying those constraints is built: the set depends on the chosen TPI, and one outside it fails to link. See Supported Widths.

Default: BitWidth<256>.

Read back from a composed descriptor as BI::bit_width, and the resulting limb count as BI::num_limbs.

SM

template<unsigned int Architecture>
SM()

Sets the target GPU architecture Architecture to generate code for, expressed as major * 100 + minor * 10 (for example, 800 for compute capability 8.0).

Default: SM<800>.

Read back from a composed descriptor as BI::sm_arch.

TPI

template<unsigned int T>
TPI()

Sets the number of Threads Per Instance T: the number of consecutive warp lanes that cooperate to hold and operate on a single big-integer instance. T must be a power of two, 1 <= T <= 32, and must not exceed the instance’s limb count (BW / 32).

Required when Warp() is used; must not be combined with Thread().

Default: none. A descriptor without this operator reports BI::tpi as 1, which is the Thread() case; 1 cannot be requested explicitly, because TPI<1>() alongside Thread() is still rejected.

Read back from a composed descriptor as BI::tpi.

Execution Operators

Exactly one of Thread() or Warp() must appear in a descriptor.

Thread

Thread()

Selects single-thread execution: one CUDA thread owns and operates on the entire big-integer instance. Cannot be combined with TPI().

Warp

Warp()

Selects warp-cooperative execution: TPI() consecutive warp lanes cooperate on one instance, each lane holding a contiguous slice of the limbs (lowest-index lane holds the least-significant limbs). Requires TPI() to be specified.

Not all operations are available under warp execution; see Warp Execution Coverage.

Error-Handling Operators

OnErrorNone

OnErrorNone()

Selects the default on-error policy: failures are not reported by the error policy itself. Callers of div_rem and inv_mod must inspect the returned bigint_error.

Default on-error policy.

OnErrorTrap

OnErrorTrap()

On any operation failure, calls __trap() on the offending thread(s), halting the kernel.

OnErrorPrintTrap

OnErrorPrintTrap()

On any operation failure, prints a diagnostic message identifying the failing block and thread indices and the specific bigint_error, then calls __trap().