j2k_cuda_runtime/
lib.rs

1// SPDX-License-Identifier: MIT OR Apache-2.0
2
3//! CUDA codec engine and Driver API runtime used by J2K CUDA adapter crates.
4
5#![deny(unsafe_op_in_unsafe_fn)]
6#![deny(missing_docs)]
7#![warn(unreachable_pub)]
8
9macro_rules! cuda_kernel_params {
10    ($($arg:ident),+ $(,)?) => {
11        [$(cuda_kernel_param(&mut $arg)),+]
12    };
13}
14
15macro_rules! impl_cuda_htj2k_encoded_status_accessors {
16    () => {
17        /// HTJ2K cleanup segment length in bytes.
18        pub fn cleanup_length(&self) -> u32 {
19            htj2k_encoded_cleanup_length(self.status)
20        }
21
22        /// HTJ2K refinement segment length in bytes.
23        pub fn refinement_length(&self) -> u32 {
24            htj2k_encoded_refinement_length(self.status)
25        }
26
27        /// Number of coding passes in the encoded payload.
28        pub fn num_coding_passes(&self) -> u8 {
29            htj2k_encoded_num_coding_passes(self.status)
30        }
31
32        /// Number of missing most-significant bitplanes.
33        pub fn num_zero_bitplanes(&self) -> u8 {
34            htj2k_encoded_num_zero_bitplanes(self.status)
35        }
36
37        /// Kernel status row downloaded after dispatch.
38        pub fn status(&self) -> CudaHtj2kEncodeStatus {
39            self.status
40        }
41
42        /// CUDA execution counters for the encode dispatch.
43        pub fn execution(&self) -> CudaExecutionStats {
44            self.execution
45        }
46
47        /// CUDA event timings for the encode dispatch.
48        pub fn stage_timings(&self) -> CudaHtj2kEncodeStageTimings {
49            self.stage_timings
50        }
51    };
52}
53
54mod allocation;
55mod build_flags;
56mod bytes;
57mod context;
58mod driver;
59mod error;
60mod execution;
61mod htj2k_decode;
62mod htj2k_encode;
63mod htj2k_packetize;
64mod j2k_decode;
65mod j2k_encode;
66mod jpeg;
67mod kernels;
68mod memory;
69#[cfg(test)]
70mod tests;
71mod transcode;
72
73pub use build_flags::transcode_kernels_built;
74#[cfg(test)]
75pub(crate) use context::CudaKernelName;
76pub use context::{
77    CudaContext, CudaExternalHostOwner, CudaExternalHostReservation,
78    CudaHtj2kCompactEncodedCodeBlock, CudaHtj2kCompactEncodedCodeBlocks,
79};
80pub use error::CudaError;
81#[cfg(test)]
82pub(crate) use execution::CudaEvent;
83pub use execution::{
84    CudaExecutionStats, CudaKernelBatchOutput, CudaKernelContiguousBatchOutput, CudaKernelOutput,
85    CudaPooledKernelOutput, CudaQueuedExecution,
86};
87pub use htj2k_decode::{
88    CudaHtj2kCleanupTarget, CudaHtj2kCodeBlockJob, CudaHtj2kDecodeOutput, CudaHtj2kDecodeResources,
89    CudaHtj2kDecodeStageTimings, CudaHtj2kDecodeTableResources, CudaHtj2kDecodeTables,
90    CudaHtj2kDequantizeTarget, CudaHtj2kStatus, CudaPooledHtj2kDecodeOutput,
91    CudaQueuedHtj2kCleanup,
92};
93pub use htj2k_encode::{
94    CudaHtj2kEncodeCodeBlockJob, CudaHtj2kEncodeCodeBlockRegionJob, CudaHtj2kEncodeResidentTarget,
95    CudaHtj2kEncodeResources, CudaHtj2kEncodeStageTimings, CudaHtj2kEncodeStatus,
96    CudaHtj2kEncodeTables, CudaHtj2kEncodedCodeBlock, CudaHtj2kEncodedCodeBlocks,
97};
98pub use htj2k_packetize::{
99    CudaHtj2kPacketizationBlock, CudaHtj2kPacketizationPacket, CudaHtj2kPacketizationStageTimings,
100    CudaHtj2kPacketizationStatus, CudaHtj2kPacketizationSubband,
101    CudaHtj2kPacketizationSubbandTagState, CudaHtj2kPacketizationTagNodeState,
102    CudaHtj2kPacketizedTile,
103};
104pub use j2k_decode::{
105    CudaJ2kIdwtJob, CudaJ2kIdwtTarget, CudaJ2kInverseMctJob, CudaJ2kRect, CudaJ2kStoreGray16Job,
106    CudaJ2kStoreGray8Job, CudaJ2kStoreRgb16Job, CudaJ2kStoreRgb16MctJob, CudaJ2kStoreRgb8Job,
107    CudaJ2kStoreRgb8MctJob, CudaJ2kStoreRgb8MctTarget, CudaJ2kStridedInterleavedPixels,
108};
109pub use j2k_encode::{
110    CudaDwt53LevelShape, CudaDwt53Output, CudaDwt97BatchStageTimings, CudaDwt97Output,
111    CudaJ2kDeinterleavedComponents, CudaJ2kQuantizeJob, CudaJ2kQuantizeSubbandRegionJob,
112    CudaJ2kQuantizedSubband, CudaJ2kResidentComponents, CudaJ2kResidentQuantizedSubband,
113    CudaResidentDwt53Output, CudaResidentDwt97Output,
114};
115pub use jpeg::{
116    CudaJpegBaselineEncodeFormat, CudaJpegBaselineEncodeHuffmanTable, CudaJpegBaselineEncodeParams,
117    CudaJpegBaselineEntropyEncodeBatchJob, CudaJpegBaselineEntropyEncodeJob,
118    CudaJpegChunkedEntropyConfig, CudaJpegChunkedEntropyPlan, CudaJpegChunkedEntropyReport,
119    CudaJpegEntropyCheckpoint, CudaJpegEntropyOverflowState, CudaJpegEntropySyncState,
120    CudaJpegHuffmanTable, CudaJpegRgb8DecodePlan, CudaJpegRgb8Sampling,
121};
122pub use memory::{
123    CudaBufferPool, CudaBufferPoolDiagnostics, CudaBufferPoolLimits, CudaBufferPoolTakeTrace,
124    CudaDeviceBuffer, CudaDeviceBufferRange, CudaDeviceBufferView, CudaDeviceBufferViewMut,
125    CudaPinnedUploadOperationGuard, CudaPinnedUploadStagingCheckout,
126    CudaPinnedUploadStagingPoolDiagnostics, CudaPinnedUploadStagingPoolLimits,
127    CudaPooledDeviceBuffer,
128};
129pub use transcode::{
130    CudaDwt97BatchGeometry, CudaDwt97BatchWithPoolRequest, CudaHtj2k97CodeblockBands,
131    CudaHtj2k97CodeblockBatchWithPoolRequest, CudaHtj2k97DeviceCodeblockBands,
132    CudaHtj2k97I16CodeblockBatchWithPoolRequest, CudaHtj2k97QuantizeParams,
133    CudaTranscodeDwt97Bands, CudaTranscodeReversible53Bands,
134};
135
136#[cfg(test)]
137pub(crate) use bytes::{
138    f32_slice_as_bytes, f32_slice_as_bytes_mut, htj2k_cleanup_multi_jobs_as_bytes,
139    i32_slice_as_bytes, i32_slice_as_bytes_mut,
140};
141#[cfg(test)]
142pub(crate) use context::HTJ2K_UVLC_ENCODE_TABLE_BYTES;
143#[cfg(test)]
144pub(crate) use htj2k_decode::{
145    htj2k_decode_multi_cleanup_dequant_kernel_for_jobs, htj2k_decode_multi_kernel_for_jobs,
146    CudaHtj2kCleanupMultiKernelJob, HTJ2K_STATUS_OK, HTJ2K_STATUS_UNSUPPORTED,
147};
148#[cfg(test)]
149pub(crate) use htj2k_encode::{
150    htj2k_encode_compact_jobs, htj2k_encode_compact_jobs_multi_input, CudaHtj2kEncodeCompactJob,
151    CudaHtj2kEncodeKernelJob, CudaHtj2kEncodeMultiInputKernelJob, HTJ2K_ENCODE_OUTPUT_CAPACITY,
152};
153#[cfg(test)]
154pub(crate) use j2k_decode::{
155    checked_f32_words_byte_len, format_idwt_batch_trace_row, idwt_batch_kernel_mode,
156    idwt_batch_trace_row, idwt_batch_uses_cooperative_53, CudaJ2kIdwtBatchKernelMode,
157    CudaJ2kIdwtMultiKernelJob,
158};
159#[cfg(test)]
160pub(crate) use jpeg::jpeg_entropy_overflow_count;
161#[cfg(test)]
162pub(crate) use memory::{copy_pooled_bytes_to_vec_uninit, pool_fit_buffer_index_by_len};
163#[cfg(test)]
164pub(crate) use transcode::{should_use_pinned_pooled_i16_upload, validate_dct_block_grid};
j2k_cuda_runtime/lib.rs

j2k_cuda_runtime/
lib.rs
