[FFmpeg-devel] [PATCH] Allow using primary CUDA device context

[Oleg Dobkin]

Add AVCUDADeviceContextFlags to control the creation of CUDA device
context for the hardware CUDA decoder.

The current values are 0 (default behavior) - new context will be
created for each decoder, and 1 - primary CUDA context will be used.

There are several reasons for using primary device context instead of
creating a new one:

 - This is the recommended way to handle device contexts (see
https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__CTX.html#group__CUDA__CTX_1g65dc0012348bc84810e2103a40d8e2cf)

 - Memory allocations, kernels and other state are associated with the
current device context. Currently, the context is not accessible from
FFmpeg API, so, technically, the memory created by the hardware decoder
(the video frame) can't be safely read.

Signed-off-by: Oleg Dobkin <olegd at anyvision.co>
---
 configure                           |  8 ++++----
 libavutil/hwcontext_cuda.c          | 22 +++++++++++++++++-----
 libavutil/hwcontext_cuda_internal.h |  2 ++
 3 files changed, 23 insertions(+), 9 deletions(-)

diff --git a/configure b/configure
index 8f4f2884cf..2519e6421f 100755
--- a/configure
+++ b/configure
@@ -6139,10 +6139,10 @@ fi
 
 if ! disabled ffnvcodec; then
     ffnv_hdr_list="ffnvcodec/nvEncodeAPI.h ffnvcodec/dynlink_cuda.h ffnvcodec/dynlink_cuviddec.h ffnvcodec/dynlink_nvcuvid.h"
-    check_pkg_config ffnvcodec "ffnvcodec >= 9.0.18.0" "$ffnv_hdr_list" "" || \
-      check_pkg_config ffnvcodec "ffnvcodec >= 8.2.15.8 ffnvcodec < 8.3" "$ffnv_hdr_list" "" || \
-      check_pkg_config ffnvcodec "ffnvcodec >= 8.1.24.9 ffnvcodec < 8.2" "$ffnv_hdr_list" "" || \
-      check_pkg_config ffnvcodec "ffnvcodec >= 8.0.14.9 ffnvcodec < 8.1" "$ffnv_hdr_list" ""
+    check_pkg_config ffnvcodec "ffnvcodec >= 9.1.23.1" "$ffnv_hdr_list" "" || \
+      check_pkg_config ffnvcodec "ffnvcodec >= 9.0.18.3 ffnvcodec < 9.1" "$ffnv_hdr_list" "" || \
+      check_pkg_config ffnvcodec "ffnvcodec >= 8.2.15.10 ffnvcodec < 8.3" "$ffnv_hdr_list" "" || \
+      check_pkg_config ffnvcodec "ffnvcodec >= 8.1.24.11 ffnvcodec < 8.2" "$ffnv_hdr_list" ""
 fi
 
 check_cpp_condition winrt windows.h "!WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)"
diff --git a/libavutil/hwcontext_cuda.c b/libavutil/hwcontext_cuda.c
index cca39e9fc7..e72efbe5af 100644
--- a/libavutil/hwcontext_cuda.c
+++ b/libavutil/hwcontext_cuda.c
@@ -29,6 +29,8 @@
 
 #define CUDA_FRAME_ALIGNMENT 256
 
+#define USE_PRIMARY_CONTEXT 1
+
 typedef struct CUDAFramesContext {
     int shift_width, shift_height;
 } CUDAFramesContext;
@@ -281,8 +283,12 @@ static void cuda_device_uninit(AVHWDeviceContext *device_ctx)
     if (hwctx->internal) {
         CudaFunctions *cu = hwctx->internal->cuda_dl;
         if (hwctx->internal->is_allocated && hwctx->cuda_ctx) {
-            CHECK_CU(cu->cuCtxDestroy(hwctx->cuda_ctx));
+            if (hwctx->internal->flags & USE_PRIMARY_CONTEXT)
+                CHECK_CU(cu->cuDevicePrimaryCtxRelease(hwctx->internal->cuda_device));
+            else
+                CHECK_CU(cu->cuCtxDestroy(hwctx->cuda_ctx));
             hwctx->cuda_ctx = NULL;
+            hwctx->internal->cuda_device = NULL;
         }
         cuda_free_functions(&hwctx->internal->cuda_dl);
     }
@@ -322,7 +328,6 @@ static int cuda_device_create(AVHWDeviceContext *device_ctx,
 {
     AVCUDADeviceContext *hwctx = device_ctx->hwctx;
     CudaFunctions *cu;
-    CUdevice cu_device;
     CUcontext dummy;
     int ret, device_idx = 0;
 
@@ -338,18 +343,25 @@ static int cuda_device_create(AVHWDeviceContext *device_ctx,
     if (ret < 0)
         goto error;
 
-    ret = CHECK_CU(cu->cuDeviceGet(&cu_device, device_idx));
+    ret = CHECK_CU(cu->cuDeviceGet(&hwctx->internal->cuda_device, device_idx));
     if (ret < 0)
         goto error;
 
-    ret = CHECK_CU(cu->cuCtxCreate(&hwctx->cuda_ctx, CU_CTX_SCHED_BLOCKING_SYNC, cu_device));
+    hwctx->internal->flags = flags;
+
+    if (flags & USE_PRIMARY_CONTEXT)
+        ret = CHECK_CU(cu->cuDevicePrimaryCtxRetain(&hwctx->cuda_ctx, hwctx->internal->cuda_device));
+    else
+        ret = CHECK_CU(cu->cuCtxCreate(&hwctx->cuda_ctx, CU_CTX_SCHED_BLOCKING_SYNC, hwctx->internal->cuda_device));
+
     if (ret < 0)
         goto error;
 
     // Setting stream to NULL will make functions automatically use the default CUstream
     hwctx->stream = NULL;
 
-    CHECK_CU(cu->cuCtxPopCurrent(&dummy));
+    if (!(flags & USE_PRIMARY_CONTEXT))
+        CHECK_CU(cu->cuCtxPopCurrent(&dummy));
 
     hwctx->internal->is_allocated = 1;
 
diff --git a/libavutil/hwcontext_cuda_internal.h b/libavutil/hwcontext_cuda_internal.h
index e1bc6ff350..d5633c58d5 100644
--- a/libavutil/hwcontext_cuda_internal.h
+++ b/libavutil/hwcontext_cuda_internal.h
@@ -31,6 +31,8 @@
 struct AVCUDADeviceContextInternal {
     CudaFunctions *cuda_dl;
     int is_allocated;
+    CUdevice cuda_device;
+    int flags;
 };
 
 #endif /* AVUTIL_HWCONTEXT_CUDA_INTERNAL_H */
-- 
2.17.1