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 CUDA_ERROR_NOT_READY: CuResult = 600;
9
10pub(crate) const PINNED_POOLED_I16_UPLOAD_MAX_BYTES: usize = 4 * 1024 * 1024;
11
12pub(crate) const DWT97_ROW_LIFT_MAX_WIDTH: i32 = 1024;
13
14pub(crate) const DWT97_ROW_LIFT_COOP_THREADS_X: c_uint = 128;
15
16pub(crate) const DWT97_ROW_LIFT_COOP_ROWS_PER_BLOCK: c_uint = 4;
17
18pub(crate) const CUDA_IDWT_TRACE_ENV_VAR: &str = "J2K_CUDA_IDWT_TRACE";
19
20#[cfg(j2k_cuda_oxide_enabled)]
21pub(crate) const REQUIRE_CUDA_OXIDE_BUILD_ENV_VAR: &str = "J2K_REQUIRE_CUDA_OXIDE_BUILD";
22
23pub(crate) const DWT97_FUSED_COLUMN_QUANTIZE_DISABLE_ENV_VAR: &str =
24    "J2K_CUDA_DISABLE_DWT97_FUSED_COLUMN_QUANTIZE";
25
26pub(crate) static CUDA_STAGE_TIMINGS_DISABLED: OnceLock<bool> = OnceLock::new();
27
28pub(crate) static DWT97_FUSED_COLUMN_QUANTIZE_DISABLED: OnceLock<bool> = OnceLock::new();
29
30pub(crate) fn cuda_stage_timings_disabled() -> bool {
31    *CUDA_STAGE_TIMINGS_DISABLED
32        .get_or_init(|| std::env::var_os("J2K_CUDA_DISABLE_STAGE_TIMINGS").is_some())
33}
34
35pub(crate) fn dwt97_fused_column_quantize_disabled() -> bool {
36    *DWT97_FUSED_COLUMN_QUANTIZE_DISABLED
37        .get_or_init(|| std::env::var_os(DWT97_FUSED_COLUMN_QUANTIZE_DISABLE_ENV_VAR).is_some())
38}
39
40#[cfg(j2k_cuda_oxide_enabled)]
41fn ensure_cuda_oxide_ptx_built(built: bool, display_name: &str) -> Result<(), CudaError> {
42    if built {
43        Ok(())
44    } else {
45        Err(CudaError::InvalidArgument {
46            message: format!(
47                "{display_name} PTX was not built; set {REQUIRE_CUDA_OXIDE_BUILD_ENV_VAR} on a Linux cuda-oxide host to require it"
48            ),
49        })
50    }
51}
52
53macro_rules! cuda_oxide_ptx_guard {
54    (feature = $feature:literal, $ensure_fn:ident, $built_const:ident, $display_name:literal, $built_cfg:meta) => {
55        #[cfg(feature = $feature)]
56        pub(crate) fn $ensure_fn() -> Result<(), CudaError> {
57            ensure_cuda_oxide_ptx_built($built_const, $display_name)
58        }
59
60        #[cfg(feature = $feature)]
61        pub(crate) const $built_const: bool = cfg!($built_cfg);
62    };
63}
64
65cuda_oxide_ptx_guard!(
66    feature = "cuda-oxide-copy-u8",
67    ensure_cuda_oxide_copy_u8_ptx_built,
68    CUDA_OXIDE_COPY_U8_PTX_BUILT,
69    "cuda-oxide CopyU8",
70    j2k_cuda_oxide_copy_u8_built
71);
72cuda_oxide_ptx_guard!(
73    feature = "cuda-oxide-j2k-encode",
74    ensure_cuda_oxide_j2k_encode_ptx_built,
75    CUDA_OXIDE_J2K_ENCODE_PTX_BUILT,
76    "cuda-oxide J2K encode",
77    j2k_cuda_oxide_j2k_encode_built
78);
79cuda_oxide_ptx_guard!(
80    feature = "cuda-oxide-j2k-decode-store",
81    ensure_cuda_oxide_j2k_decode_store_ptx_built,
82    CUDA_OXIDE_J2K_DECODE_STORE_PTX_BUILT,
83    "cuda-oxide J2K decode store",
84    j2k_cuda_oxide_j2k_decode_store_built
85);
86cuda_oxide_ptx_guard!(
87    feature = "cuda-oxide-j2k-classic-decode",
88    ensure_cuda_oxide_j2k_classic_decode_ptx_built,
89    CUDA_OXIDE_J2K_CLASSIC_DECODE_PTX_BUILT,
90    "cuda-oxide classic J2K decode",
91    j2k_cuda_oxide_j2k_classic_decode_built
92);
93cuda_oxide_ptx_guard!(
94    feature = "cuda-oxide-j2k-dequantize",
95    ensure_cuda_oxide_j2k_dequantize_ptx_built,
96    CUDA_OXIDE_J2K_DEQUANTIZE_PTX_BUILT,
97    "cuda-oxide J2K dequantize",
98    j2k_cuda_oxide_j2k_dequantize_built
99);
100cuda_oxide_ptx_guard!(
101    feature = "cuda-oxide-j2k-idwt",
102    ensure_cuda_oxide_j2k_idwt_ptx_built,
103    CUDA_OXIDE_J2K_IDWT_PTX_BUILT,
104    "cuda-oxide J2K IDWT",
105    j2k_cuda_oxide_j2k_idwt_built
106);
107cuda_oxide_ptx_guard!(
108    feature = "cuda-oxide-j2k-ml",
109    ensure_cuda_oxide_j2k_ml_ptx_built,
110    CUDA_OXIDE_J2K_ML_PTX_BUILT,
111    "cuda-oxide j2k-ml",
112    j2k_cuda_oxide_j2k_ml_built
113);
114cuda_oxide_ptx_guard!(
115    feature = "cuda-oxide-transcode",
116    ensure_cuda_oxide_transcode_ptx_built,
117    CUDA_OXIDE_TRANSCODE_PTX_BUILT,
118    "cuda-oxide transcode",
119    j2k_cuda_oxide_transcode_built
120);
121cuda_oxide_ptx_guard!(
122    feature = "cuda-oxide-htj2k-decode",
123    ensure_cuda_oxide_htj2k_decode_ptx_built,
124    CUDA_OXIDE_HTJ2K_DECODE_PTX_BUILT,
125    "cuda-oxide HTJ2K decode",
126    j2k_cuda_oxide_htj2k_decode_built
127);
128cuda_oxide_ptx_guard!(
129    feature = "cuda-oxide-htj2k-encode",
130    ensure_cuda_oxide_htj2k_encode_ptx_built,
131    CUDA_OXIDE_HTJ2K_ENCODE_PTX_BUILT,
132    "cuda-oxide HTJ2K encode",
133    j2k_cuda_oxide_htj2k_encode_built
134);
135cuda_oxide_ptx_guard!(
136    feature = "cuda-oxide-jpeg-decode",
137    ensure_cuda_oxide_jpeg_decode_ptx_built,
138    CUDA_OXIDE_JPEG_DECODE_PTX_BUILT,
139    "cuda-oxide JPEG decode",
140    j2k_cuda_oxide_jpeg_decode_built
141);
142cuda_oxide_ptx_guard!(
143    feature = "cuda-oxide-jpeg-encode",
144    ensure_cuda_oxide_jpeg_encode_ptx_built,
145    CUDA_OXIDE_JPEG_ENCODE_PTX_BUILT,
146    "cuda-oxide JPEG encode",
147    j2k_cuda_oxide_jpeg_encode_built
148);
149
150/// Whether the coefficient-domain transcode CUDA Oxide kernels were compiled.
151/// Backends check this to return a structured unavailable error on non-strict
152/// non-CUDA/doc builds instead of attempting a device launch.
153#[must_use]
154pub fn transcode_kernels_built() -> bool {
155    #[cfg(feature = "cuda-oxide-transcode")]
156    {
157        CUDA_OXIDE_TRANSCODE_PTX_BUILT
158    }
159    #[cfg(not(feature = "cuda-oxide-transcode"))]
160    {
161        false
162    }
163}