Non-owning implementation view of a DDL sketch. More...

Public Types
using	register_type = uint32_t

using	layout_type = Layout

Public Member Functions
__host__ __device__ constexpr	sketch_view (device_span< register_type > registers, cuda::std::optional< blacklist_view > blacklist=cuda::std::nullopt) noexcept
	Constructs a reference over `registers`.

__host__ __device__ constexpr device_span< register_type >	data () const noexcept
	Device registers.

Result< void >	clear_async (cuda::stream_ref stream) const noexcept
	Resets every register to zero.

Result< void >	add_async (device_span< uint64_t const > input, cuda::stream_ref stream) const noexcept
	Accumulates packed device k-mers into the existing sketch without clearing it.

Result< void >	add_sequence_async (device_span< char const > sequence, cuda::stream_ref stream) const noexcept
	Accumulates device-resident raw ASCII bases without clearing.

template<bool IncludeCardinality = false>
Result< void >	summary_async (sketch_view other, pairwise_summary &output, cuda::stream_ref stream) const noexcept
	Computes the raw pairwise summary into caller-owned device storage.

Result< void >	cardinality_async (uint64_t empty_out, double estimate_out, cuda::stream_ref stream) const noexcept
	Computes this sketch's cardinality reduction on the GPU.

Result< void >	hybrid_cardinality_async (hybrid_cardinality_estimates *output, cuda::stream_ref stream) const noexcept
	Computes BBTools and paper-style HybridDDL estimates in one GPU register scan.

Static Public Member Functions
static constexpr size_t	bucket_count () noexcept
	Number of registers in the sketch.

static constexpr uint32_t	kmer_length () noexcept
	Compile-time k-mer length.

Detailed Description

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>
class cuddl::detail::sketch_view< K, BucketCount, Layout >

Non-owning implementation view of a DDL sketch.

Provides allocation-free, stream-ordered operations on an external allocation of BucketCount uint32_t registers, each holding its bucket's winning score. Pass by value into device or host code.

Definition at line 25 of file sketch_view.cuh.

Member Typedef Documentation

◆ layout_type

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

using cuddl::detail::sketch_view< K, BucketCount, Layout >::layout_type = Layout

Definition at line 32 of file sketch_view.cuh.

◆ register_type

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

using cuddl::detail::sketch_view< K, BucketCount, Layout >::register_type = uint32_t

Definition at line 31 of file sketch_view.cuh.

Constructor & Destructor Documentation

◆ sketch_view()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

__host__ __device__ constexpr cuddl::detail::sketch_view< K, BucketCount, Layout >::sketch_view	(	device_span< register_type >	registers,
		cuda::std::optional< blacklist_view >	blacklist = `cuda::std::nullopt`
	)

inlineexplicitconstexprnoexcept

Constructs a reference over registers.

Definition at line 35 of file sketch_view.cuh.

39 : registers_(registers), blacklist_(blacklist) {}

cuddl::detail::blacklist

__global__ size_t uint32_t bool blacklist_view blacklist

Definition kernels.cuh:96

cuddl::detail::registers

__global__ size_t uint32_t * registers

Definition kernels.cuh:94

Member Function Documentation

◆ add_async()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

Result< void > cuddl::detail::sketch_view< K, BucketCount, Layout >::add_async	(	device_span< uint64_t const >	input,
		cuda::stream_ref	stream
	)		const

inlinenoexcept

Accumulates packed device k-mers into the existing sketch without clearing it.

The input must remain valid until stream completes.

Definition at line 67 of file sketch_view.cuh.

                                                                                       {
        return detail::launch_construction<BucketCount, Layout>(
            input, registers_, stream, blacklist_
        );
    }

◆ add_sequence_async()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

Result< void > cuddl::detail::sketch_view< K, BucketCount, Layout >::add_sequence_async	(	device_span< char const >	sequence,
		cuda::stream_ref	stream
	)		const

inlinenoexcept

Accumulates device-resident raw ASCII bases without clearing.

Expects raw contiguous ASCII (no FASTA/newline stripping). Canonicalises case-insensitively with the packed path; ambiguity breaks windows and emits no k-mer across it. Accumulates without reset; short input below K is a no-op. Windows per call are capped at UINT32_MAX, so size must not exceed UINT32_MAX + K - 1. No hidden carry: callers supply the K-1 overlap within one record and no overlap between distinct records. No host copies. The input must remain valid until stream completes.

Definition at line 82 of file sketch_view.cuh.

                                                                                               {
        return detail::launch_sequence_add<BucketCount, Layout>(
            sequence, K, registers_, stream, blacklist_
        );
    }

◆ bucket_count()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

static constexpr size_t cuddl::detail::sketch_view< K, BucketCount, Layout >::bucket_count ( )

inlinestaticconstexprnoexcept

Number of registers in the sketch.

Definition at line 47 of file sketch_view.cuh.

                                                                  {
        return BucketCount;
    }

◆ cardinality_async()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

Result< void > cuddl::detail::sketch_view< K, BucketCount, Layout >::cardinality_async	(	uint64_t *	empty_out,
		double *	estimate_out,
		cuda::stream_ref	stream
	)		const

inlinenoexcept

Computes this sketch's cardinality reduction on the GPU.

empty_out receives the empty-register count and estimate_out the cardinality estimate.

Definition at line 107 of file sketch_view.cuh.

                     {
        detail::cardinality_kernel<BucketCount, Layout><<<1, detail::block_size, 0, stream.get()>>>(
            registers_.data(), empty_out, estimate_out
        );
        return cuda_try(cudaGetLastError());
    }

Here is the call graph for this function:

Here is the caller graph for this function:

◆ clear_async()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

Result< void > cuddl::detail::sketch_view< K, BucketCount, Layout >::clear_async ( cuda::stream_ref stream ) const

inlinenoexcept

Resets every register to zero.

Definition at line 57 of file sketch_view.cuh.

                                                                                 {
        return cuda_try([&] {
            cuda::fill_bytes(stream, cuda::std::span{registers_.data(), registers_.size()}, 0);
        });
    }

Here is the call graph for this function:

◆ data()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

__host__ __device__ constexpr device_span< register_type > cuddl::detail::sketch_view< K, BucketCount, Layout >::data ( ) const

inlineconstexprnoexcept

Device registers.

Definition at line 42 of file sketch_view.cuh.

                                                                                                 {
        return registers_;
    }

◆ hybrid_cardinality_async()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

Result< void > cuddl::detail::sketch_view< K, BucketCount, Layout >::hybrid_cardinality_async	(	hybrid_cardinality_estimates *	output,
		cuda::stream_ref	stream
	)		const

inlinenoexcept

Computes BBTools and paper-style HybridDDL estimates in one GPU register scan.

Definition at line 119 of file sketch_view.cuh.

                     {
        detail::hybrid_cardinality_kernel<BucketCount, Layout>
            <<<1, detail::block_size, 0, stream.get()>>>(registers_.data(), output);
        return cuda_try(cudaGetLastError());
    }

Here is the call graph for this function:

◆ kmer_length()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

static constexpr uint32_t cuddl::detail::sketch_view< K, BucketCount, Layout >::kmer_length ( )

inlinestaticconstexprnoexcept

Compile-time k-mer length.

Definition at line 52 of file sketch_view.cuh.

                                                                   {
        return K;
    }

◆ summary_async()

template<uint32_t K, size_t BucketCount, typename Layout = default_register_layout>

template<bool IncludeCardinality = false>

Result< void > cuddl::detail::sketch_view< K, BucketCount, Layout >::summary_async	(	sketch_view< K, BucketCount, Layout >	other,
		pairwise_summary &	output,
		cuda::stream_ref	stream
	)		const

inlinenoexcept

Computes the raw pairwise summary into caller-owned device storage.

output must point at device memory valid until stream completes.

Definition at line 92 of file sketch_view.cuh.

                     {
        detail::summary_kernel<BucketCount, IncludeCardinality, Layout>
            <<<1, detail::block_size, 0, stream.get()>>>(
                registers_.data(), other.registers_.data(), output
            );
        return cuda_try(cudaGetLastError());
    }

Here is the call graph for this function:

The documentation for this class was generated from the following file:

sketch_view.cuh

Public Types

Public Member Functions

Static Public Member Functions

Detailed Description

Member Typedef Documentation

◆ layout_type

◆ register_type

Constructor & Destructor Documentation

◆ sketch_view()

Member Function Documentation

◆ add_async()

◆ add_sequence_async()

◆ bucket_count()

◆ cardinality_async()

◆ clear_async()

◆ data()

◆ hybrid_cardinality_async()

◆ kmer_length()

◆ summary_async()