Skip to main content

j2k_cuda_runtime/
build_flags.rs

1use crate::driver::CuResult;
2#[cfg(j2k_cuda_oxide_enabled)]
3use crate::error::CudaError;
4use std::{os::raw::c_uint, sync::OnceLock};
5
6pub(crate) const CUDA_SUCCESS: CuResult = 0;
7
8pub(crate) const PINNED_POOLED_I16_UPLOAD_MAX_BYTES: usize = 4 * 1024 * 1024;
9
10pub(crate) const DWT97_ROW_LIFT_MAX_WIDTH: i32 = 1024;
11
12pub(crate) const DWT97_ROW_LIFT_COOP_THREADS_X: c_uint = 128;
13
14pub(crate) const DWT97_ROW_LIFT_COOP_ROWS_PER_BLOCK: c_uint = 4;
15
16pub(crate) const CUDA_IDWT_TRACE_ENV_VAR: &str = "J2K_CUDA_IDWT_TRACE";
17
18#[cfg(j2k_cuda_oxide_enabled)]
19pub(crate) const REQUIRE_CUDA_OXIDE_BUILD_ENV_VAR: &str = "J2K_REQUIRE_CUDA_OXIDE_BUILD";
20
21pub(crate) const DWT97_FUSED_COLUMN_QUANTIZE_DISABLE_ENV_VAR: &str =
22    "J2K_CUDA_DISABLE_DWT97_FUSED_COLUMN_QUANTIZE";
23
24pub(crate) static CUDA_STAGE_TIMINGS_DISABLED: OnceLock<bool> = OnceLock::new();
25
26pub(crate) static DWT97_FUSED_COLUMN_QUANTIZE_DISABLED: OnceLock<bool> = OnceLock::new();
27
28pub(crate) fn cuda_stage_timings_disabled() -> bool {
29    *CUDA_STAGE_TIMINGS_DISABLED
30        .get_or_init(|| std::env::var_os("J2K_CUDA_DISABLE_STAGE_TIMINGS").is_some())
31}
32
33pub(crate) fn dwt97_fused_column_quantize_disabled() -> bool {
34    *DWT97_FUSED_COLUMN_QUANTIZE_DISABLED
35        .get_or_init(|| std::env::var_os(DWT97_FUSED_COLUMN_QUANTIZE_DISABLE_ENV_VAR).is_some())
36}
37
38#[cfg(j2k_cuda_oxide_enabled)]
39fn ensure_cuda_oxide_ptx_built(built: bool, display_name: &str) -> Result<(), CudaError> {
40    if built {
41        Ok(())
42    } else {
43        Err(CudaError::InvalidArgument {
44            message: format!(
45                "{display_name} PTX was not built; set {REQUIRE_CUDA_OXIDE_BUILD_ENV_VAR} on a Linux cuda-oxide host to require it"
46            ),
47        })
48    }
49}
50
51macro_rules! cuda_oxide_ptx_guard {
52    (feature = $feature:literal, $ensure_fn:ident, $built_const:ident, $display_name:literal, $built_cfg:meta) => {
53        #[cfg(feature = $feature)]
54        pub(crate) fn $ensure_fn() -> Result<(), CudaError> {
55            ensure_cuda_oxide_ptx_built($built_const, $display_name)
56        }
57
58        #[cfg(feature = $feature)]
59        pub(crate) const $built_const: bool = cfg!($built_cfg);
60    };
61}
62
63cuda_oxide_ptx_guard!(
64    feature = "cuda-oxide-copy-u8",
65    ensure_cuda_oxide_copy_u8_ptx_built,
66    CUDA_OXIDE_COPY_U8_PTX_BUILT,
67    "cuda-oxide CopyU8",
68    j2k_cuda_oxide_copy_u8_built
69);
70cuda_oxide_ptx_guard!(
71    feature = "cuda-oxide-j2k-encode",
72    ensure_cuda_oxide_j2k_encode_ptx_built,
73    CUDA_OXIDE_J2K_ENCODE_PTX_BUILT,
74    "cuda-oxide J2K encode",
75    j2k_cuda_oxide_j2k_encode_built
76);
77cuda_oxide_ptx_guard!(
78    feature = "cuda-oxide-j2k-decode-store",
79    ensure_cuda_oxide_j2k_decode_store_ptx_built,
80    CUDA_OXIDE_J2K_DECODE_STORE_PTX_BUILT,
81    "cuda-oxide J2K decode store",
82    j2k_cuda_oxide_j2k_decode_store_built
83);
84cuda_oxide_ptx_guard!(
85    feature = "cuda-oxide-j2k-classic-decode",
86    ensure_cuda_oxide_j2k_classic_decode_ptx_built,
87    CUDA_OXIDE_J2K_CLASSIC_DECODE_PTX_BUILT,
88    "cuda-oxide classic J2K decode",
89    j2k_cuda_oxide_j2k_classic_decode_built
90);
91cuda_oxide_ptx_guard!(
92    feature = "cuda-oxide-j2k-dequantize",
93    ensure_cuda_oxide_j2k_dequantize_ptx_built,
94    CUDA_OXIDE_J2K_DEQUANTIZE_PTX_BUILT,
95    "cuda-oxide J2K dequantize",
96    j2k_cuda_oxide_j2k_dequantize_built
97);
98cuda_oxide_ptx_guard!(
99    feature = "cuda-oxide-j2k-idwt",
100    ensure_cuda_oxide_j2k_idwt_ptx_built,
101    CUDA_OXIDE_J2K_IDWT_PTX_BUILT,
102    "cuda-oxide J2K IDWT",
103    j2k_cuda_oxide_j2k_idwt_built
104);
105cuda_oxide_ptx_guard!(
106    feature = "cuda-oxide-j2k-ml",
107    ensure_cuda_oxide_j2k_ml_ptx_built,
108    CUDA_OXIDE_J2K_ML_PTX_BUILT,
109    "cuda-oxide j2k-ml",
110    j2k_cuda_oxide_j2k_ml_built
111);
112cuda_oxide_ptx_guard!(
113    feature = "cuda-oxide-transcode",
114    ensure_cuda_oxide_transcode_ptx_built,
115    CUDA_OXIDE_TRANSCODE_PTX_BUILT,
116    "cuda-oxide transcode",
117    j2k_cuda_oxide_transcode_built
118);
119cuda_oxide_ptx_guard!(
120    feature = "cuda-oxide-htj2k-decode",
121    ensure_cuda_oxide_htj2k_decode_ptx_built,
122    CUDA_OXIDE_HTJ2K_DECODE_PTX_BUILT,
123    "cuda-oxide HTJ2K decode",
124    j2k_cuda_oxide_htj2k_decode_built
125);
126cuda_oxide_ptx_guard!(
127    feature = "cuda-oxide-htj2k-encode",
128    ensure_cuda_oxide_htj2k_encode_ptx_built,
129    CUDA_OXIDE_HTJ2K_ENCODE_PTX_BUILT,
130    "cuda-oxide HTJ2K encode",
131    j2k_cuda_oxide_htj2k_encode_built
132);
133cuda_oxide_ptx_guard!(
134    feature = "cuda-oxide-jpeg-decode",
135    ensure_cuda_oxide_jpeg_decode_ptx_built,
136    CUDA_OXIDE_JPEG_DECODE_PTX_BUILT,
137    "cuda-oxide JPEG decode",
138    j2k_cuda_oxide_jpeg_decode_built
139);
140cuda_oxide_ptx_guard!(
141    feature = "cuda-oxide-jpeg-encode",
142    ensure_cuda_oxide_jpeg_encode_ptx_built,
143    CUDA_OXIDE_JPEG_ENCODE_PTX_BUILT,
144    "cuda-oxide JPEG encode",
145    j2k_cuda_oxide_jpeg_encode_built
146);
147
148/// Whether the coefficient-domain transcode CUDA Oxide kernels were compiled.
149/// Backends check this to return a structured unavailable error on non-strict
150/// non-CUDA/doc builds instead of attempting a device launch.
151#[must_use]
152pub fn transcode_kernels_built() -> bool {
153    #[cfg(feature = "cuda-oxide-transcode")]
154    {
155        CUDA_OXIDE_TRANSCODE_PTX_BUILT
156    }
157    #[cfg(not(feature = "cuda-oxide-transcode"))]
158    {
159        false
160    }
161}