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 classic_decode;
58mod context;
59mod driver;
60mod error;
61mod execution;
62mod htj2k_decode;
63mod htj2k_encode;
64mod htj2k_packetize;
65mod j2k_decode;
66mod j2k_encode;
67mod jpeg;
68mod kernels;
69mod memory;
70mod ml;
71#[cfg(test)]
72mod tests;
73mod transcode;
74
75pub use build_flags::transcode_kernels_built;
76pub use classic_decode::{
77    CudaClassicCodeBlockJob, CudaClassicDecodeStageTimings, CudaClassicDecodeTarget,
78    CudaClassicSegment, CudaClassicStatus,
79};
80#[cfg(test)]
81pub(crate) use context::CudaKernelName;
82pub use context::{
83    CudaContext, CudaExternalHostOwner, CudaExternalHostReservation,
84    CudaHtj2kCompactEncodedCodeBlock, CudaHtj2kCompactEncodedCodeBlocks,
85};
86pub use error::CudaError;
87#[cfg(test)]
88pub(crate) use execution::CudaEvent;
89pub use execution::{
90    CudaExecutionStats, CudaKernelBatchOutput, CudaKernelContiguousBatchOutput, CudaKernelOutput,
91    CudaPooledKernelOutput, CudaQueuedExecution,
92};
93pub use htj2k_decode::{
94    CudaHtj2kCleanupTarget, CudaHtj2kCodeBlockJob, CudaHtj2kDecodeOutput, CudaHtj2kDecodeResources,
95    CudaHtj2kDecodeStageTimings, CudaHtj2kDecodeTableResources, CudaHtj2kDecodeTables,
96    CudaHtj2kDequantizeTarget, CudaHtj2kStatus, CudaPooledHtj2kDecodeOutput,
97    CudaQueuedHtj2kCleanup,
98};
99pub use htj2k_encode::{
100    CudaHtj2kEncodeCodeBlockJob, CudaHtj2kEncodeCodeBlockRegionJob, CudaHtj2kEncodeResidentTarget,
101    CudaHtj2kEncodeResources, CudaHtj2kEncodeStageTimings, CudaHtj2kEncodeStatus,
102    CudaHtj2kEncodeTables, CudaHtj2kEncodedCodeBlock, CudaHtj2kEncodedCodeBlocks,
103};
104pub use htj2k_packetize::{
105    CudaHtj2kPacketizationBlock, CudaHtj2kPacketizationPacket, CudaHtj2kPacketizationStageTimings,
106    CudaHtj2kPacketizationStatus, CudaHtj2kPacketizationSubband,
107    CudaHtj2kPacketizationSubbandTagState, CudaHtj2kPacketizationTagNodeState,
108    CudaHtj2kPacketizedTile,
109};
110pub use j2k_decode::{
111    CudaJ2kIdwtJob, CudaJ2kIdwtTarget, CudaJ2kInverseMctJob, CudaJ2kRect, CudaJ2kStoreGray16Job,
112    CudaJ2kStoreGray8Job, CudaJ2kStoreRgb16Job, CudaJ2kStoreRgb16MctJob, CudaJ2kStoreRgb8Job,
113    CudaJ2kStoreRgb8MctJob, CudaJ2kStoreRgb8MctTarget, CudaJ2kStridedInterleavedPixels,
114};
115pub use j2k_encode::{
116    CudaDwt53LevelShape, CudaDwt53Output, CudaDwt97BatchStageTimings, CudaDwt97Output,
117    CudaJ2kDeinterleavedComponents, CudaJ2kQuantizeJob, CudaJ2kQuantizeSubbandRegionJob,
118    CudaJ2kQuantizedSubband, CudaJ2kResidentComponents, CudaJ2kResidentQuantizedSubband,
119    CudaResidentDwt53Output, CudaResidentDwt97Output,
120};
121pub use jpeg::{
122    CudaJpegBaselineEncodeFormat, CudaJpegBaselineEncodeHuffmanTable, CudaJpegBaselineEncodeParams,
123    CudaJpegBaselineEntropyEncodeBatchJob, CudaJpegBaselineEntropyEncodeJob,
124    CudaJpegChunkedEntropyConfig, CudaJpegChunkedEntropyPlan, CudaJpegChunkedEntropyReport,
125    CudaJpegEntropyCheckpoint, CudaJpegEntropyOverflowState, CudaJpegEntropySyncState,
126    CudaJpegHuffmanTable, CudaJpegRgb8DecodePlan, CudaJpegRgb8Sampling,
127};
128pub use memory::{
129    CudaBufferPool, CudaBufferPoolDiagnostics, CudaBufferPoolLimits, CudaBufferPoolTakeTrace,
130    CudaDeviceBuffer, CudaDeviceBufferRange, CudaDeviceBufferView, CudaDeviceBufferViewMut,
131    CudaExternalDeviceBufferViewMut, CudaPinnedUploadOperationGuard,
132    CudaPinnedUploadStagingCheckout, CudaPinnedUploadStagingPoolDiagnostics,
133    CudaPinnedUploadStagingPoolLimits, CudaPooledDeviceBuffer,
134};
135pub use ml::{CudaJ2kMlKernelConfig, CudaJ2kMlLayout, CudaJ2kMlNormalization, CudaJ2kMlSample};
136pub use transcode::{
137    CudaDwt97BatchGeometry, CudaDwt97BatchWithPoolRequest, CudaHtj2k97CodeblockBands,
138    CudaHtj2k97CodeblockBatchWithPoolRequest, CudaHtj2k97DeviceCodeblockBands,
139    CudaHtj2k97I16CodeblockBatchWithPoolRequest, CudaHtj2k97QuantizeParams,
140    CudaTranscodeDwt97Bands, CudaTranscodeReversible53Bands,
141};
142
143#[cfg(test)]
144pub(crate) use bytes::{
145    f32_slice_as_bytes, f32_slice_as_bytes_mut, htj2k_cleanup_multi_jobs_as_bytes,
146    i32_slice_as_bytes, i32_slice_as_bytes_mut,
147};
148#[cfg(test)]
149pub(crate) use context::HTJ2K_UVLC_ENCODE_TABLE_BYTES;
150#[cfg(test)]
151pub(crate) use htj2k_decode::{
152    htj2k_decode_multi_cleanup_dequant_kernel_for_jobs, htj2k_decode_multi_kernel_for_jobs,
153    CudaHtj2kCleanupMultiKernelJob, HTJ2K_STATUS_OK, HTJ2K_STATUS_UNSUPPORTED,
154};
155#[cfg(test)]
156pub(crate) use htj2k_encode::{
157    htj2k_encode_compact_jobs, htj2k_encode_compact_jobs_multi_input, CudaHtj2kEncodeCompactJob,
158    CudaHtj2kEncodeKernelJob, CudaHtj2kEncodeMultiInputKernelJob, HTJ2K_ENCODE_OUTPUT_CAPACITY,
159};
160#[cfg(test)]
161pub(crate) use j2k_decode::{
162    checked_f32_words_byte_len, format_idwt_batch_trace_row, idwt_batch_kernel_mode,
163    idwt_batch_trace_row, idwt_batch_uses_cooperative_53, CudaJ2kIdwtBatchKernelMode,
164    CudaJ2kIdwtMultiKernelJob,
165};
166#[cfg(test)]
167pub(crate) use jpeg::jpeg_entropy_overflow_count;
168#[cfg(test)]
169pub(crate) use memory::{copy_pooled_bytes_to_vec_uninit, pool_fit_buffer_index_by_len};
170#[cfg(test)]
171pub(crate) use transcode::{should_use_pinned_pooled_i16_upload, validate_dct_block_grid};
j2k_cuda_runtime/lib.rs

j2k_cuda_runtime/
lib.rs
