From 1cb47d549b8ef06bd9081fdedb2cab04ff08a08f Mon Sep 17 00:00:00 2001
From: Seungha Yang <seungha@centricular.com>
Date: Tue, 20 Dec 2022 01:23:42 +0900
Subject: [PATCH] cudaconverter: Don't sync per conversion

Caller should take care of synchronization

Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/3629>
---
 .../sys/nvcodec/gstcudaconverter.c            | 21 ++++++++++++++-----
 .../sys/nvcodec/gstcudaconverter.h            |  3 ++-
 .../sys/nvcodec/gstcudaconvertscale.c         |  2 +-
 3 files changed, 19 insertions(+), 7 deletions(-)

diff --git a/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconverter.c b/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconverter.c
index ec8e4e7f41..1183664665 100644
--- a/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconverter.c
+++ b/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconverter.c
@@ -2069,12 +2069,15 @@ ensure_fallback_buffer (GstCudaConverter * self, gint width_in_bytes,
 static CUtexObject
 gst_cuda_converter_create_texture (GstCudaConverter * self,
     CUdeviceptr src, gint width, gint height, gint stride, CUfilter_mode mode,
-    CUarray_format format, guint channles, gint plane, CUstream stream)
+    CUarray_format format, guint channles, gint plane, CUstream stream,
+    gboolean * need_sync)
 {
   GstCudaConverterPrivate *priv = self->priv;
   CUresult ret;
   CUdeviceptr src_ptr;
 
+  *need_sync = FALSE;
+
   src_ptr = src;
 
   if (priv->texture_align > 0 && (src_ptr % priv->texture_align) != 0) {
@@ -2096,12 +2099,14 @@ gst_cuda_converter_create_texture (GstCudaConverter * self,
         * GST_VIDEO_INFO_COMP_PSTRIDE (&priv->in_info, plane),
         params.Height = GST_VIDEO_INFO_COMP_HEIGHT (&priv->in_info, plane);
 
-    ret = CuMemcpy2D (&params);
+    ret = CuMemcpy2DAsync (&params, stream);
     if (!gst_cuda_result (ret)) {
       GST_ERROR_OBJECT (self, "Couldn't copy to fallback buffer");
       return 0;
     }
 
+    *need_sync = TRUE;
+
     src_ptr = priv->fallback_buffer[plane].ptr;
     stride = priv->fallback_buffer[plane].stride;
   }
@@ -2145,7 +2150,8 @@ gst_cuda_converter_unpack_rgb (GstCudaConverter * self,
 
 gboolean
 gst_cuda_converter_convert_frame (GstCudaConverter * converter,
-    GstVideoFrame * src_frame, GstVideoFrame * dst_frame, CUstream stream)
+    GstVideoFrame * src_frame, GstVideoFrame * dst_frame, CUstream stream,
+    gboolean * synchronized)
 {
   GstCudaConverterPrivate *priv;
   const TextureFormat *format;
@@ -2159,6 +2165,7 @@ gst_cuda_converter_convert_frame (GstCudaConverter * converter,
   gpointer args[] = { &texture[0], &texture[1], &texture[2], &texture[3],
     &dst[0], &dst[1], &dst[2], &dst[3], &stride[0], &stride[1]
   };
+  gboolean need_sync = FALSE;
 
   g_return_val_if_fail (GST_IS_CUDA_CONVERTER (converter), FALSE);
   g_return_val_if_fail (src_frame != NULL, FALSE);
@@ -2196,7 +2203,7 @@ gst_cuda_converter_convert_frame (GstCudaConverter * converter,
           GST_VIDEO_FRAME_COMP_HEIGHT (src_frame, i),
           GST_VIDEO_FRAME_PLANE_STRIDE (src_frame, i),
           priv->filter_mode[i], format->array_format[i], format->channels[i], i,
-          stream);
+          stream, &need_sync);
       if (!texture[i]) {
         GST_ERROR_OBJECT (converter, "Couldn't create texture %d", i);
         goto out;
@@ -2223,7 +2230,11 @@ gst_cuda_converter_convert_frame (GstCudaConverter * converter,
     goto out;
   }
 
-  CuStreamSynchronize (stream);
+  if (need_sync)
+    CuStreamSynchronize (stream);
+
+  if (synchronized)
+    *synchronized = need_sync;
 
   ret = TRUE;
 
diff --git a/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconverter.h b/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconverter.h
index b19c03ef14..1be9d8a9b3 100644
--- a/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconverter.h
+++ b/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconverter.h
@@ -93,7 +93,8 @@ GstCudaConverter *  gst_cuda_converter_new (const GstVideoInfo * in_info,
 gboolean            gst_cuda_converter_convert_frame (GstCudaConverter * converter,
                                                       GstVideoFrame * src_frame,
                                                       GstVideoFrame * dst_frame,
-                                                      CUstream cuda_stream);
+                                                      CUstream cuda_stream,
+                                                      gboolean * synchronized);
 
 G_END_DECLS
 
diff --git a/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconvertscale.c b/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconvertscale.c
index 6d5ed1a226..b779c98329 100644
--- a/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconvertscale.c
+++ b/subprojects/gst-plugins-bad/sys/nvcodec/gstcudaconvertscale.c
@@ -1380,7 +1380,7 @@ gst_cuda_base_convert_transform (GstBaseTransform * trans,
   }
 
   if (!gst_cuda_converter_convert_frame (self->converter, &in_frame, &out_frame,
-          gst_cuda_stream_get_handle (btrans->stream))) {
+          gst_cuda_stream_get_handle (btrans->stream), NULL)) {
     GST_ERROR_OBJECT (self, "Failed to convert frame");
     ret = GST_FLOW_ERROR;
   }