fix(cuda): release resources on init failure, and close three stale state.md rows

CHANGELOG.md

@@ -20369,6 +20369,21 @@ uses the kernel-template readback pair: `integer_psnr_cuda`, `integer_ssim_cuda`
   were already correct; `speed_chroma_cuda.c` was the only remaining file.
 
 
+- **`speed_chroma_cuda` and `speed_temporal_cuda` leaked their CUDA module and stream on init failure paths.**
+  `cuModuleLoadData()` and `cuStreamCreate()` both jump to failure labels on error,
+  but those labels (`fail_pop`, `fail_after_pop`, and `free_cpu`/`free_all`) only
+  called `free_cuda_buffers` / `free_cuda_buffers_st`, which only release device
+  and host memory and never touch `s->module` or `s->stream`.
+  Any failure after module loading leaked the module, and any failure after stream
+  creation leaked both. In `speed_chroma_cuda.c`, PR #1007 had originally added
+  inline cleanup which was inadvertently removed during the PR #1029 merge;
+  factoring the teardown into `release_cuda_module_and_stream` (and
+  `release_cuda_module_and_stream_st` in `speed_temporal_cuda.c`) prevents future
+  drift and unifies cleanup with `close_fex_cuda` / `close_fex_st`.
+  `integer_ms_ssim_cuda.c` and `integer_psnr_hvs_cuda.c` failure exits are
+  hardened to route through `close_fex_cuda` as well.
+
+
 - **fix(cuda)**: add `static` to 12 TU-internal helper functions across
   `integer_vif_cuda.c`, `integer_adm_cuda.c`, and `integer_motion_cuda.c` —
   eliminates ODR name-clash risk in static-lib builds; restores parity with

changelog.d/fixed/cuda-speed-init-leak-regression.md

@@ -0,0 +1,13 @@
+- **`speed_chroma_cuda` and `speed_temporal_cuda` leaked their CUDA module and stream on init failure paths.**
+  `cuModuleLoadData()` and `cuStreamCreate()` both jump to failure labels on error,
+  but those labels (`fail_pop`, `fail_after_pop`, and `free_cpu`/`free_all`) only
+  called `free_cuda_buffers` / `free_cuda_buffers_st`, which only release device
+  and host memory and never touch `s->module` or `s->stream`.
+  Any failure after module loading leaked the module, and any failure after stream
+  creation leaked both. In `speed_chroma_cuda.c`, PR #1007 had originally added
+  inline cleanup which was inadvertently removed during the PR #1029 merge;
+  factoring the teardown into `release_cuda_module_and_stream` (and
+  `release_cuda_module_and_stream_st` in `speed_temporal_cuda.c`) prevents future
+  drift and unifies cleanup with `close_fex_cuda` / `close_fex_st`.
+  `integer_ms_ssim_cuda.c` and `integer_psnr_hvs_cuda.c` failure exits are
+  hardened to route through `close_fex_cuda` as well.

core/src/feature/cuda/integer_ms_ssim_cuda.c

@@ -311,7 +311,7 @@ static int init_fex_cuda(VmafFeatureExtractor *fex, enum VmafPixelFormat pix_fmt
     if (ctx_pushed)
         (void)cu_f->cuCtxPopCurrent(NULL);
 fail_after_pop:
-    (void)vmaf_cuda_kernel_lifecycle_close(&s->lc, fex->cu_state);
+    (void)close_fex_cuda(fex);
     return _cuda_err;
 }
 
@@ -628,8 +628,10 @@ static int close_fex_cuda(VmafFeatureExtractor *fex)
     }
     ret |= vmaf_dictionary_free(&s->feature_name_dict);
     const CudaFunctions *cu_f = fex->cu_state->f;
-    if (cu_f && s->module)
+    if (cu_f && s->module) {
         (void)cu_f->cuModuleUnload(s->module);
+        s->module = NULL;
+    }
     return ret;
 }
 

core/src/feature/cuda/integer_psnr_hvs_cuda.c

@@ -240,7 +240,7 @@ static int init_fex_cuda(VmafFeatureExtractor *fex, enum VmafPixelFormat pix_fmt
     if (ctx_pushed)
         (void)cu_f->cuCtxPopCurrent(NULL);
 fail_after_pop:
-    (void)vmaf_cuda_kernel_lifecycle_close(&s->lc, fex->cu_state);
+    (void)close_fex_cuda(fex);
     return _cuda_err;
 }
 
@@ -521,8 +521,10 @@ static int close_fex_cuda(VmafFeatureExtractor *fex)
             (void)vmaf_cuda_buffer_host_free(fex->cu_state, s->h_uint_dist[p]);
     }
     ret |= vmaf_dictionary_free(&s->feature_name_dict);
-    if (cu_f && s->module)
+    if (cu_f && s->module) {
         (void)cu_f->cuModuleUnload(s->module);
+        s->module = NULL;
+    }
     return ret;
 }
 

core/src/feature/cuda/speed_chroma_cuda.c

@@ -835,6 +835,7 @@ static int init_fex_cuda(VmafFeatureExtractor *fex, enum VmafPixelFormat pix_fmt
     return 0;
 
 free_cpu:
+    release_cuda_module_and_stream(s, cu_f);
     free_cuda_buffers(s, cu_f);
     return err;
 
@@ -914,14 +915,12 @@ static int extract_fex_cuda(VmafFeatureExtractor *fex, VmafPicture *ref_pic,
 static int close_fex_cuda(VmafFeatureExtractor *fex)
 {
     SpeedChromaCudaState *s = fex->priv;
-    if (fex->cu_state && fex->cu_state->f)
+    if (fex->cu_state && fex->cu_state->f) {
         free_cuda_buffers(s, fex->cu_state->f);
+        release_cuda_module_and_stream(s, fex->cu_state->f);
+    }
     if (s->feature_name_dict)
         vmaf_dictionary_free(&s->feature_name_dict);
-    if (s->stream && fex->cu_state && fex->cu_state->f)
-        (void)fex->cu_state->f->cuStreamDestroy(s->stream);
-    if (fex->cu_state && fex->cu_state->f && s->module)
-        (void)fex->cu_state->f->cuModuleUnload(s->module);
     return 0;
 }
 

core/src/feature/cuda/speed_temporal_cuda.c

@@ -442,6 +442,20 @@ static int run_score_st(SpeedTemporalCudaState *s, CudaFunctions *cu_f, float *s
 /* Lifecycle                                                           */
 /* ------------------------------------------------------------------ */
 
+static void release_cuda_module_and_stream_st(SpeedTemporalCudaState *s, CudaFunctions *cu_f)
+{
+    if (!s || !cu_f)
+        return;
+    if (s->stream) {
+        (void)cu_f->cuStreamDestroy(s->stream);
+        s->stream = 0;
+    }
+    if (s->module) {
+        (void)cu_f->cuModuleUnload(s->module);
+        s->module = NULL;
+    }
+}
+
 static int init_fex_st(VmafFeatureExtractor *fex, enum VmafPixelFormat pix_fmt, unsigned bpc,
                        unsigned w, unsigned h)
 {
@@ -550,10 +564,12 @@ static int init_fex_st(VmafFeatureExtractor *fex, enum VmafPixelFormat pix_fmt,
     return 0;
 
 free_all:
+    release_cuda_module_and_stream_st(s, cu_f);
     free_cuda_buffers_st(s, cu_f);
     return err;
 fail_pop:
     (void)cu_f->cuCtxPopCurrent(NULL);
+    release_cuda_module_and_stream_st(s, cu_f);
     free_cuda_buffers_st(s, cu_f);
     return _cuda_err;
 fail:
@@ -715,13 +731,10 @@ static int close_fex_st(VmafFeatureExtractor *fex)
     SpeedTemporalCudaState *s = fex->priv;
     if (fex->cu_state && fex->cu_state->f) {
         free_cuda_buffers_st(s, fex->cu_state->f);
-        if (s->stream)
-            (void)fex->cu_state->f->cuStreamDestroy(s->stream);
+        release_cuda_module_and_stream_st(s, fex->cu_state->f);
     }
     if (s->feature_name_dict)
         vmaf_dictionary_free(&s->feature_name_dict);
-    if (fex->cu_state && fex->cu_state->f && s->module)
-        (void)fex->cu_state->f->cuModuleUnload(s->module);
     return 0;
 }
 

docs/backends/cuda/overview.md

@@ -159,6 +159,18 @@ Adding a new CUDA extractor: see [`/add-feature-extractor`](../../../.claude/ski
   that amortises `cuStreamSynchronize` over 8 frames rather than 1, superseding
   the engine-level optimization for this extractor.
 
+### Error-path resource release (PR #1279)
+
+Every CUDA feature extractor releases what it has already acquired when `init`
+or `submit` fails part-way: pinned host buffers, device buffers, the module and
+the stream, in the reverse order of acquisition and each one NULL-guarded. The
+`speed_chroma` / `speed_temporal` twins share one `release_cuda_module_and_stream()`
+helper so the two failure labels cannot drift apart again (PR #1007 added the
+calls inline and PR #1029 removed them the same day); `integer_ms_ssim` and
+`integer_psnr_hvs` release their pinned buffers on the same labels. There is no
+user-visible behaviour change: a failed init still returns the same error code,
+it just no longer leaks (docs/state.md `T-CUDA-INIT-SUBMIT-LEAKS-2026-06-19`).
+
 ## Profiling
 
 See [nvtx/profiling.md](../nvtx/profiling.md) for Nsight Systems recipes that

docs/rebase-notes.md

@@ -1,6 +1,17 @@
 <!-- markdownlint-disable MD001 MD003 MD004 MD007 MD013 MD018 MD022 MD024 MD025 MD026 MD028 MD029 MD031 MD032 MD033 MD036 MD037 MD038 MD040 MD041 MD046 MD049 MD050 MD051 MD052 MD053 MD055 MD056 MD058 MD059 -->
 # Rebase notes
 
+## fix/gpu-init-leaks-and-hip-mirror — fix CUDA init error path leaks and close verified GPU state issues (2026-09-04)
+
+no rebase impact: fork-local CUDA and documentation files.
+
+- `core/src/feature/cuda/speed_chroma_cuda.c`, `core/src/feature/cuda/speed_temporal_cuda.c`,
+  `core/src/feature/cuda/integer_ms_ssim_cuda.c`, `core/src/feature/cuda/integer_psnr_hvs_cuda.c`:
+  wholly fork-added CUDA feature extractors (upstream Netflix/vmaf does not have GPU SpEED, and
+  upstream CUDA extractors lack these fork-specific teardown paths). No upstream rebase conflicts.
+- `docs/state.md`: closed out resolved audit tasks `T-HIP-MOTION-V2-MIRROR-OFF-BY-ONE-2026-06-13`,
+  `T-SYCL-INIT-LEAKS-EXC-2026-06-19`, `T-SPEED-GPU-REGISTRY-ORPHAN-2026-06-19`, and
+  `T-CUDA-INIT-SUBMIT-LEAKS-2026-06-19`.
 <!-- markdownlint-disable MD001 MD003 MD004 MD007 MD013 MD018 MD022 MD024 MD025 MD026 MD028 MD029 MD031 MD032 MD033 MD036 MD037 MD038 MD040 MD041 MD046 MD049 MD050 MD051 MD052 MD053 MD055 MD056 MD058 MD059 -->
 ## fix/vmaftune-state-bugs — libx264 two-pass CRF conflict fix (2026-09-03)