nvcodec: Refactor plugin initialization

Create CUDA context per device, instead of per codec and encoder/decoder.
Allocating CUDA context is heavy operation so we should reuse it
as much as possible.

Fixes: https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/issues/1130

sys/nvcodec/gstcudaloader.c

@@ -24,6 +24,9 @@
 #include "gstcudaloader.h"
 #include <gmodule.h>
 
+GST_DEBUG_CATEGORY_EXTERN (gst_nvcodec_debug);
+#define GST_CAT_DEFAULT gst_nvcodec_debug
+
 #ifndef G_OS_WIN32
 #define CUDA_LIBNAME "libcuda.so.1"
 #else

sys/nvcodec/gstnvbaseenc.c

@@ -28,6 +28,7 @@
 
 #include <string.h>
 
+GST_DEBUG_CATEGORY_EXTERN (gst_nvenc_debug);
 #define GST_CAT_DEFAULT gst_nvenc_debug
 
 #if HAVE_NVCODEC_GST_GL

sys/nvcodec/gstnvdec.c

@@ -34,8 +34,8 @@
 
 #include <string.h>
 
-GST_DEBUG_CATEGORY_STATIC (gst_nvdec_debug_category);
-#define GST_CAT_DEFAULT gst_nvdec_debug_category
+GST_DEBUG_CATEGORY_EXTERN (gst_nvdec_debug);
+#define GST_CAT_DEFAULT gst_nvdec_debug
 
 #ifdef HAVE_NVCODEC_GST_GL
 #define SUPPORTED_GL_APIS (GST_GL_API_OPENGL | GST_GL_API_OPENGL3 | GST_GL_API_GLES2)
@@ -1544,13 +1544,9 @@ typedef struct
 static void
 gst_nvdec_register (GstPlugin * plugin, GType type, cudaVideoCodec codec_type,
     const gchar * codec, const gchar * sink_caps_string, guint rank,
-    gint device_count)
+    gint device_idx, CUcontext cuda_ctx)
 {
-  gint i;
-
-  for (i = 0; i < device_count; i++) {
-    CUdevice cuda_device;
-    CUcontext cuda_ctx;
+  {
     CUresult cuda_ret;
     gint max_width = 0, min_width = G_MAXINT;
     gint max_height = 0, min_height = G_MAXINT;
@@ -1575,18 +1571,12 @@ gst_nvdec_register (GstPlugin * plugin, GType type, cudaVideoCodec codec_type,
     GValue format = G_VALUE_INIT;
     GValue profile_list = G_VALUE_INIT;
 
-    if (CuDeviceGet (&cuda_device, i) != CUDA_SUCCESS)
-      continue;
-
-    if (CuCtxCreate (&cuda_ctx, 0, cuda_device) != CUDA_SUCCESS)
-      continue;
-
     g_value_init (&format_list, GST_TYPE_LIST);
     g_value_init (&format, G_TYPE_STRING);
     g_value_init (&profile_list, GST_TYPE_LIST);
 
     if (CuCtxPushCurrent (cuda_ctx) != CUDA_SUCCESS)
-      goto cuda_free;
+      goto done;
 
     for (c_idx = 0; c_idx < G_N_ELEMENTS (chroma_list); c_idx++) {
       for (b_idx = 0; b_idx < G_N_ELEMENTS (bitdepth_minus8); b_idx++) {
@@ -1682,7 +1672,7 @@ gst_nvdec_register (GstPlugin * plugin, GType type, cudaVideoCodec codec_type,
 
     if (num_support == 0) {
       GST_INFO ("device can not support %s", codec);
-      goto cuda_free;
+      goto done;
     }
 
     src_templ = gst_caps_new_simple ("video/x-raw",
@@ -1717,16 +1707,14 @@ gst_nvdec_register (GstPlugin * plugin, GType type, cudaVideoCodec codec_type,
 
     CuCtxPopCurrent (NULL);
 
-  cuda_free:
-    CuCtxDestroy (cuda_ctx);
-
+  done:
     g_value_unset (&format_list);
     g_value_unset (&format);
     g_value_unset (&profile_list);
 
     if (sink_templ && src_templ) {
-      gst_nvdec_subclass_register (plugin, type, codec_type, codec, i, rank,
-          sink_templ, src_templ);
+      gst_nvdec_subclass_register (plugin, type, codec_type, codec, device_idx,
+          rank, sink_templ, src_templ);
     }
 
     gst_clear_caps (&sink_templ);
@@ -1773,14 +1761,10 @@ const GstNvCodecMap codec_map[] = {
 };
 
 void
-gst_nvdec_plugin_init (GstPlugin * plugin)
+gst_nvdec_plugin_init (GstPlugin * plugin, guint device_index,
+    CUcontext cuda_ctx)
 {
   gint i;
-  CUresult cuda_ret;
-  gint dev_count = 0;
-
-  GST_DEBUG_CATEGORY_INIT (gst_nvdec_debug_category, "nvdec", 0,
-      "Debug category for the nvdec element");
 
   if (!gst_cuvid_can_get_decoder_caps ()) {
     GstCaps *src_templ;
@@ -1804,27 +1788,16 @@ gst_nvdec_plugin_init (GstPlugin * plugin)
       sink_templ = gst_caps_from_string (codec_map[i].sink_caps_string);
 
       gst_nvdec_subclass_register (plugin, GST_TYPE_NVDEC, codec_map[i].codec,
-          codec_map[i].codec_name, 0, GST_RANK_PRIMARY, sink_templ, src_templ);
+          codec_map[i].codec_name, device_index, GST_RANK_PRIMARY,
+          sink_templ, src_templ);
     }
 
     return;
   }
 
-  cuda_ret = CuInit (0);
-  if (cuda_ret != CUDA_SUCCESS) {
-    GST_ERROR ("Failed to initialize CUDA API");
-    return;
-  }
-
-  cuda_ret = CuDeviceGetCount (&dev_count);
-  if (cuda_ret != CUDA_SUCCESS || dev_count == 0) {
-    GST_ERROR ("No CUDA devices detected");
-    return;
-  }
-
   for (i = 0; i < G_N_ELEMENTS (codec_map); i++) {
     gst_nvdec_register (plugin, GST_TYPE_NVDEC, codec_map[i].codec,
         codec_map[i].codec_name, codec_map[i].sink_caps_string,
-        GST_RANK_PRIMARY, dev_count);
+        GST_RANK_PRIMARY, device_index, cuda_ctx);
   }
 }

sys/nvcodec/gstnvdec.h

@@ -101,7 +101,9 @@ struct _GstNvDecClass
 
 GType gst_nvdec_get_type (void);
 
-void gst_nvdec_plugin_init (GstPlugin * plugin);
+void gst_nvdec_plugin_init (GstPlugin * plugin,
+                            guint device_index,
+                            CUcontext cuda_ctx);
 
 G_END_DECLS
 

sys/nvcodec/gstnvenc.c

@@ -44,7 +44,7 @@ typedef NVENCSTATUS NVENCAPI
 tNvEncodeAPICreateInstance (NV_ENCODE_API_FUNCTION_LIST * functionList);
 tNvEncodeAPICreateInstance *nvEncodeAPICreateInstance;
 
-GST_DEBUG_CATEGORY (gst_nvenc_debug);
+GST_DEBUG_CATEGORY_EXTERN (gst_nvenc_debug);
 #define GST_CAT_DEFAULT gst_nvenc_debug
 
 static NV_ENCODE_API_FUNCTION_LIST nvenc_api;
@@ -290,27 +290,6 @@ gst_nvenc_get_nv_buffer_format (GstVideoFormat fmt)
   return NV_ENC_BUFFER_FORMAT_UNDEFINED;
 }
 
-static gboolean
-load_nvenc_library (void)
-{
-  GModule *module;
-
-  module = g_module_open (NVENC_LIBRARY_NAME, G_MODULE_BIND_LAZY);
-  if (module == NULL) {
-    GST_WARNING ("Could not open library %s, %s",
-        NVENC_LIBRARY_NAME, g_module_error ());
-    return FALSE;
-  }
-
-  if (!g_module_symbol (module, "NvEncodeAPICreateInstance",
-          (gpointer *) & nvEncodeAPICreateInstance)) {
-    GST_ERROR ("%s", g_module_error ());
-    return FALSE;
-  }
-
-  return TRUE;
-}
-
 typedef struct
 {
   GstVideoFormat gst_format;
@@ -603,13 +582,9 @@ gst_nvenc_get_supported_codec_profiles (gpointer enc, GUID codec_id)
 
 static void
 gst_nv_enc_register (GstPlugin * plugin, GUID codec_id, const gchar * codec,
-    guint rank, gint device_count)
+    guint rank, gint device_index, CUcontext cuda_ctx)
 {
-  gint i;
-
-  for (i = 0; i < device_count; i++) {
-    CUdevice cuda_device;
-    CUcontext cuda_ctx, dummy;
+  {
     GValue *formats = NULL;
     GValue *profiles;
     GValue *interlace_modes;
@@ -626,23 +601,17 @@ gst_nv_enc_register (GstPlugin * plugin, GUID codec_id, const gchar * codec,
     gint j;
     GstNvEncDeviceCaps device_caps = { 0, };
 
-    if (CuDeviceGet (&cuda_device, i) != CUDA_SUCCESS)
-      continue;
-
-    if (CuCtxCreate (&cuda_ctx, 0, cuda_device) != CUDA_SUCCESS)
-      continue;
-
-    if (CuCtxPopCurrent (&dummy) != CUDA_SUCCESS) {
-      goto cuda_free;
-    }
-
     params.version = gst_nvenc_get_open_encode_session_ex_params_version ();
     params.apiVersion = gst_nvenc_get_api_version ();
     params.device = cuda_ctx;
     params.deviceType = NV_ENC_DEVICE_TYPE_CUDA;
 
+    if (CuCtxPushCurrent (cuda_ctx) != CUDA_SUCCESS)
+      goto done;
+
     if (NvEncOpenEncodeSessionEx (&params, &enc) != NV_ENC_SUCCESS) {
-      goto cuda_free;
+      CuCtxPopCurrent (NULL);
+      goto done;
     }
 
     if (NvEncGetEncodeGUIDs (enc, guids, G_N_ELEMENTS (guids),
@@ -692,7 +661,7 @@ gst_nv_enc_register (GstPlugin * plugin, GUID codec_id, const gchar * codec,
       device_caps.rc_modes = 0;
     } else {
       GST_DEBUG ("[device-%d %s] rate control modes: 0x%x",
-          i, codec, device_caps.rc_modes);
+          device_index, codec, device_caps.rc_modes);
 #define IS_SUPPORTED_RC(rc_modes,mode) \
       ((((rc_modes) & (mode)) == mode) ? "supported" : "not supported")
 
@@ -744,18 +713,20 @@ gst_nv_enc_register (GstPlugin * plugin, GUID codec_id, const gchar * codec,
       device_caps.bframes = 0;
     }
 
-    DEBUG_DEVICE_CAPS (i,
+    DEBUG_DEVICE_CAPS (device_index,
         codec, "weighted prediction", device_caps.weighted_prediction);
 
-    DEBUG_DEVICE_CAPS (i, codec, "custom vbv-buffer-size",
+    DEBUG_DEVICE_CAPS (device_index, codec, "custom vbv-buffer-size",
         device_caps.custom_vbv_bufsize);
 
-    DEBUG_DEVICE_CAPS (i, codec, "rc-loockahead", device_caps.lookahead);
+    DEBUG_DEVICE_CAPS (device_index, codec, "rc-loockahead",
+        device_caps.lookahead);
 
-    DEBUG_DEVICE_CAPS (i, codec, "temporal adaptive quantization",
+    DEBUG_DEVICE_CAPS (device_index, codec, "temporal adaptive quantization",
         device_caps.temporal_aq);
 
-    GST_DEBUG ("[device-%d %s] max bframes: %d", i, codec, device_caps.bframes);
+    GST_DEBUG ("[device-%d %s] max bframes: %d", device_index, codec,
+        device_caps.bframes);
 
     interlace_modes = gst_nvenc_get_interlace_modes (enc, codec_id);
 
@@ -806,18 +777,17 @@ gst_nv_enc_register (GstPlugin * plugin, GUID codec_id, const gchar * codec,
 
   enc_free:
     NvEncDestroyEncoder (enc);
+    CuCtxPopCurrent (NULL);
     /* fall-through */
 
-  cuda_free:
-    CuCtxDestroy (cuda_ctx);
-
+  done:
     if (sink_templ && src_templ) {
       if (gst_nvenc_cmp_guid (codec_id, NV_ENC_CODEC_H264_GUID)) {
-        gst_nv_h264_enc_register (plugin, i, rank, sink_templ, src_templ,
-            &device_caps);
+        gst_nv_h264_enc_register (plugin, device_index, rank, sink_templ,
+            src_templ, &device_caps);
       } else if (gst_nvenc_cmp_guid (codec_id, NV_ENC_CODEC_HEVC_GUID)) {
-        gst_nv_h265_enc_register (plugin, i, rank, sink_templ, src_templ,
-            &device_caps);
+        gst_nv_h265_enc_register (plugin, device_index, rank, sink_templ,
+            src_templ, &device_caps);
       } else {
         g_assert_not_reached ();
       }
@@ -837,19 +807,26 @@ gst_nv_enc_register (GstPlugin * plugin, GUID codec_id, const gchar * codec,
 
 static guint32 gst_nvenc_api_version = NVENCAPI_VERSION;
 
-void
-gst_nvenc_plugin_init (GstPlugin * plugin)
+gboolean
+gst_nvenc_load_library (void)
 {
-  NVENCSTATUS ret = NV_ENC_SUCCESS;
+  GModule *module;
+  NVENCSTATUS ret;
 
-  GST_DEBUG_CATEGORY_INIT (gst_nvenc_debug, "nvenc", 0, "Nvidia NVENC encoder");
-
-  nvenc_api.version = NV_ENCODE_API_FUNCTION_LIST_VER;
-  if (!load_nvenc_library ()) {
-    GST_INFO ("Failed to load nvenc library");
-    return;
+  module = g_module_open (NVENC_LIBRARY_NAME, G_MODULE_BIND_LAZY);
+  if (module == NULL) {
+    GST_WARNING ("Could not open library %s, %s",
+        NVENC_LIBRARY_NAME, g_module_error ());
+    return FALSE;
   }
 
+  if (!g_module_symbol (module, "NvEncodeAPICreateInstance",
+          (gpointer *) & nvEncodeAPICreateInstance)) {
+    GST_ERROR ("%s", g_module_error ());
+    return FALSE;
+  }
+
+  nvenc_api.version = NV_ENCODE_API_FUNCTION_LIST_VER;
   ret = nvEncodeAPICreateInstance (&nvenc_api);
 
   /* WARNING: Any developers who want to bump SDK version must ensure that
@@ -892,31 +869,17 @@ gst_nvenc_plugin_init (GstPlugin * plugin)
     ret = nvEncodeAPICreateInstance (&nvenc_api);
   }
 
-  if (ret == NV_ENC_SUCCESS) {
-    CUresult cuda_ret;
-    gint dev_count = 0;
+  return ret == NV_ENC_SUCCESS;
+}
 
-    GST_INFO ("Created NVEncodeAPI instance, got function table");
-
-    cuda_ret = CuInit (0);
-    if (cuda_ret != CUDA_SUCCESS) {
-      GST_ERROR ("Failed to initialize CUDA API");
-      return;
-    }
-
-    cuda_ret = CuDeviceGetCount (&dev_count);
-    if (cuda_ret != CUDA_SUCCESS || dev_count == 0) {
-      GST_ERROR ("No CUDA devices detected");
-      return;
-    }
-
-    gst_nv_enc_register (plugin, NV_ENC_CODEC_H264_GUID,
-        "h264", GST_RANK_PRIMARY * 2, dev_count);
-    gst_nv_enc_register (plugin, NV_ENC_CODEC_HEVC_GUID,
-        "h265", GST_RANK_PRIMARY * 2, dev_count);
-  } else {
-    GST_ERROR ("too old driver, could not load api vtable");
-  }
+void
+gst_nvenc_plugin_init (GstPlugin * plugin, guint device_index,
+    CUcontext cuda_ctx)
+{
+  gst_nv_enc_register (plugin, NV_ENC_CODEC_H264_GUID,
+      "h264", GST_RANK_PRIMARY * 2, device_index, cuda_ctx);
+  gst_nv_enc_register (plugin, NV_ENC_CODEC_HEVC_GUID,
+      "h265", GST_RANK_PRIMARY * 2, device_index, cuda_ctx);
 }
 
 guint32

sys/nvcodec/gstnvenc.h

@@ -26,8 +26,6 @@
 #include "gstcudaloader.h"
 #include "nvEncodeAPI.h"
 
-GST_DEBUG_CATEGORY_EXTERN (gst_nvenc_debug);
-
 G_GNUC_INTERNAL
 gboolean                gst_nvenc_cmp_guid (GUID g1, GUID g2);
 
@@ -48,7 +46,9 @@ GValue *                gst_nvenc_get_supported_codec_profiles (gpointer enc,
                                                                 GUID codec_id);
 
 G_GNUC_INTERNAL
-void                    gst_nvenc_plugin_init (GstPlugin * plugin);
+void                    gst_nvenc_plugin_init (GstPlugin * plugin,
+                                               guint device_index,
+                                               CUcontext cuda_ctx);
 
 G_GNUC_INTERNAL
 guint32                 gst_nvenc_get_api_version (void);
@@ -113,5 +113,7 @@ guint32                 gst_nvenc_get_event_params_version (void);
 G_GNUC_INTERNAL
 guint32                 gst_nvenc_get_open_encode_session_ex_params_version (void);
 
+G_GNUC_INTERNAL
+gboolean                gst_nvenc_load_library (void);
 
 #endif /* __GST_NVENC_H_INCLUDED__ */

sys/nvcodec/plugin.c

@@ -32,17 +32,81 @@
 #include "gstnvdec.h"
 #include "gstnvenc.h"
 
+GST_DEBUG_CATEGORY (gst_nvcodec_debug);
+GST_DEBUG_CATEGORY (gst_nvdec_debug);
+GST_DEBUG_CATEGORY (gst_nvenc_debug);
+
+#define GST_CAT_DEFAULT gst_nvcodec_debug
+
 static gboolean
 plugin_init (GstPlugin * plugin)
 {
-  if (!gst_cuda_load_library ())
-    return TRUE;
+  CUresult cuda_ret;
+  gint dev_count = 0;
+  gint i;
+  gboolean nvdec_available = TRUE;
+  gboolean nvenc_available = TRUE;
 
-  if (gst_cuvid_load_library ()) {
-    gst_nvdec_plugin_init (plugin);
+  GST_DEBUG_CATEGORY_INIT (gst_nvcodec_debug, "nvcodec", 0, "nvcodec");
+  GST_DEBUG_CATEGORY_INIT (gst_nvdec_debug, "nvdec", 0, "nvdec");
+  GST_DEBUG_CATEGORY_INIT (gst_nvenc_debug, "nvenc", 0, "nvenc");
+
+  if (!gst_cuda_load_library ()) {
+    GST_WARNING ("Failed to load cuda library");
+    return TRUE;
   }
 
-  gst_nvenc_plugin_init (plugin);
+  if (!gst_cuvid_load_library ()) {
+    GST_WARNING ("Failed to load nvdec library");
+    nvdec_available = FALSE;
+  }
+
+  if (!gst_nvenc_load_library ()) {
+    GST_WARNING ("Failed to load nvenc library");
+    nvenc_available = FALSE;
+  }
+
+  if (!nvdec_available && !nvenc_available)
+    return TRUE;
+
+  cuda_ret = CuInit (0);
+  if (cuda_ret != CUDA_SUCCESS) {
+    GST_WARNING ("Failed to init cuda, ret: 0x%x", (gint) cuda_ret);
+    return TRUE;
+  }
+
+  if (CuDeviceGetCount (&dev_count) != CUDA_SUCCESS || !dev_count) {
+    GST_WARNING ("No available device, ret: 0x%x", (gint) cuda_ret);
+    return TRUE;
+  }
+
+  for (i = 0; i < dev_count; i++) {
+    CUdevice cuda_device;
+    CUcontext cuda_ctx;
+
+    cuda_ret = CuDeviceGet (&cuda_device, i);
+    if (cuda_ret != CUDA_SUCCESS) {
+      GST_WARNING ("Failed to get device handle %d, ret: 0x%x", i,
+          (gint) cuda_ret);
+      continue;
+    }
+
+    cuda_ret = CuCtxCreate (&cuda_ctx, 0, cuda_device);
+    if (cuda_ret != CUDA_SUCCESS) {
+      GST_WARNING ("Failed to create cuda context, ret: 0x%x", (gint) cuda_ret);
+      continue;
+    }
+
+    CuCtxPopCurrent (NULL);
+
+    if (nvdec_available)
+      gst_nvdec_plugin_init (plugin, i, cuda_ctx);
+
+    if (nvenc_available)
+      gst_nvenc_plugin_init (plugin, i, cuda_ctx);
+
+    CuCtxDestroy (cuda_ctx);
+  }
 
   return TRUE;
 }