From d40ca8814ca6c4d1014c4558f69110150c2e539c Mon Sep 17 00:00:00 2001
From: dijunkun <junkun.di@hotmail.com>
Date: Thu, 10 Apr 2025 17:10:40 +0800
Subject: [PATCH] [fix] fix nvcodec crash due to cuda context not being
 released

---
 .../decode/nvcodec/nvidia_video_decoder.cpp   | 23 +++++++++++--------
 .../decode/nvcodec/nvidia_video_decoder.h     |  2 ++
 .../video/decode/video_decoder_factory.cpp    |  3 +--
 .../encode/nvcodec/nvidia_video_encoder.cpp   |  6 ++++-
 .../video/encode/video_encoder_factory.cpp    |  5 ++--
 5 files changed, 23 insertions(+), 16 deletions(-)

diff --git a/src/media/video/decode/nvcodec/nvidia_video_decoder.cpp b/src/media/video/decode/nvcodec/nvidia_video_decoder.cpp
index 3dfd3b0..5b1d0d9 100644
--- a/src/media/video/decode/nvcodec/nvidia_video_decoder.cpp
+++ b/src/media/video/decode/nvcodec/nvidia_video_decoder.cpp
@@ -1,7 +1,6 @@
 #include "nvidia_video_decoder.h"
 
 #include "log.h"
-#include "nvcodec_api.h"
 
 // #define SAVE_DECODED_NV12_STREAM
 // #define SAVE_RECEIVED_H264_STREAM
@@ -24,29 +23,33 @@ NvidiaVideoDecoder::~NvidiaVideoDecoder() {
     file_h264_ = nullptr;
   }
 #endif
+
+  if (cuda_context_) {
+    cuCtxDestroy(cuda_context_);
+    cuda_context_ = nullptr;
+  }
 }
 
 int NvidiaVideoDecoder::Init() {
-  ck(cuInit_ld(0));
+  ck(cuInit(0));
   int nGpu = 0;
   int iGpu = 0;
 
-  ck(cuDeviceGetCount_ld(&nGpu));
+  ck(cuDeviceGetCount(&nGpu));
   if (nGpu < 1) {
     return -1;
   }
 
-  CUdevice cuDevice;
-  cuDeviceGet_ld(&cuDevice, iGpu);
+  cuDeviceGet(&cuda_device_, iGpu);
 
-  CUcontext cuContext = NULL;
-  cuCtxCreate_ld(&cuContext, 0, cuDevice);
-  if (!cuContext) {
+  cuCtxCreate(&cuda_context_, 0, cuda_device_);
+  if (!cuda_context_) {
     return -1;
   }
 
-  decoder = new NvDecoder(cuContext, false, cudaVideoCodec_H264, true, false,
-                          nullptr, nullptr, false, 4096, 2160, 1000, false);
+  decoder =
+      new NvDecoder(cuda_context_, false, cudaVideoCodec_H264, true, false,
+                    nullptr, nullptr, false, 4096, 2160, 1000, false);
 
 #ifdef SAVE_DECODED_NV12_STREAM
   file_nv12_ = fopen("decoded_nv12_stream.yuv", "w+b");
diff --git a/src/media/video/decode/nvcodec/nvidia_video_decoder.h b/src/media/video/decode/nvcodec/nvidia_video_decoder.h
index 354426d..ef9c4f1 100644
--- a/src/media/video/decode/nvcodec/nvidia_video_decoder.h
+++ b/src/media/video/decode/nvcodec/nvidia_video_decoder.h
@@ -22,6 +22,8 @@ class NvidiaVideoDecoder : public VideoDecoder {
  private:
   std::shared_ptr<SystemClock> clock_ = nullptr;
   NvDecoder* decoder = nullptr;
+  CUcontext cuda_context_ = NULL;
+  CUdevice cuda_device_ = 0;
   bool get_first_keyframe_ = false;
   bool skip_frame_ = false;
 
diff --git a/src/media/video/decode/video_decoder_factory.cpp b/src/media/video/decode/video_decoder_factory.cpp
index 382610d..a32a399 100644
--- a/src/media/video/decode/video_decoder_factory.cpp
+++ b/src/media/video/decode/video_decoder_factory.cpp
@@ -7,7 +7,6 @@
 #if __APPLE__
 #else
 #include "nvcodec/nvidia_video_decoder.h"
-#include "nvcodec_api.h"
 #endif
 
 #include "log.h"
@@ -49,7 +48,7 @@ bool VideoDecoderFactory::CheckIsHardwareAccerlerationSupported() {
 #else
   CUresult cuResult;
   CUvideoctxlock cudaCtxLock;
-  cuResult = cuvidCtxLockCreate_ld(&cudaCtxLock, 0);
+  cuResult = cuvidCtxLockCreate(&cudaCtxLock, 0);
   if (cuResult != CUDA_SUCCESS) {
     LOG_WARN(
         "System not support hardware accelerated decode, use default software "
diff --git a/src/media/video/encode/nvcodec/nvidia_video_encoder.cpp b/src/media/video/encode/nvcodec/nvidia_video_encoder.cpp
index 6c692ad..9cc65eb 100644
--- a/src/media/video/encode/nvcodec/nvidia_video_encoder.cpp
+++ b/src/media/video/encode/nvcodec/nvidia_video_encoder.cpp
@@ -3,7 +3,6 @@
 #include <chrono>
 
 #include "log.h"
-#include "nvcodec_api.h"
 #include "nvcodec_common.h"
 
 // #define SAVE_RECEIVED_NV12_STREAM
@@ -36,6 +35,11 @@ NvidiaVideoEncoder::~NvidiaVideoEncoder() {
   if (encoder_) {
     encoder_->DestroyEncoder();
   }
+
+  if (cuda_context_) {
+    cuCtxDestroy(cuda_context_);
+    cuda_context_ = nullptr;
+  }
 }
 
 int NvidiaVideoEncoder::Init() {
diff --git a/src/media/video/encode/video_encoder_factory.cpp b/src/media/video/encode/video_encoder_factory.cpp
index 36d2019..dd44027 100644
--- a/src/media/video/encode/video_encoder_factory.cpp
+++ b/src/media/video/encode/video_encoder_factory.cpp
@@ -6,7 +6,6 @@
 #else
 #include "aom/aom_av1_encoder.h"
 #include "nvcodec/nvidia_video_encoder.h"
-#include "nvcodec_api.h"
 #include "openh264/openh264_encoder.h"
 #endif
 
@@ -49,7 +48,7 @@ bool VideoEncoderFactory::CheckIsHardwareAccerlerationSupported() {
   CUresult cuResult;
   NV_ENCODE_API_FUNCTION_LIST functionList = {NV_ENCODE_API_FUNCTION_LIST_VER};
 
-  cuResult = cuInit_ld(0);
+  cuResult = cuInit(0);
   if (cuResult != CUDA_SUCCESS) {
     LOG_WARN(
         "System not support hardware accelerated encode, use default software "
@@ -57,7 +56,7 @@ bool VideoEncoderFactory::CheckIsHardwareAccerlerationSupported() {
     return false;
   }
 
-  NVENCSTATUS nvEncStatus = NvEncodeAPICreateInstance_ld(&functionList);
+  NVENCSTATUS nvEncStatus = NvEncodeAPICreateInstance(&functionList);
   if (nvEncStatus != NV_ENC_SUCCESS) {
     LOG_WARN(
         "System not support hardware accelerated encode, use default software "