#1413:

* build all kernel variants we may need
* ignore generated nvenc7 pyd
* remove nvapi code and use pynvml instead
* better CUDA kernel loading error messages

git-svn-id: https://xpra.org/svn/Xpra/trunk@14825 3bb7dfac-3a0b-4e04-842a-767bc560f471

src/setup.py

@@ -195,7 +195,6 @@ def is_msvc():
 # * wheezy: 53.35
 csc_swscale_ENABLED     = DEFAULT and pkg_config_ok("--exists", "libswscale", fallback=WIN32)
 nvenc7_ENABLED = DEFAULT and BITS==64 and pkg_config_ok("--exists", "nvenc7")
-nvapi_ENABLED = DEFAULT and BITS==64 and pkg_config_ok("--exists", "nvapi")
 csc_libyuv_ENABLED      = DEFAULT and pkg_config_ok("--exists", "libyuv", fallback=WIN32)
 
 #Cython / gcc / packaging build options:
@@ -211,7 +210,7 @@ def is_msvc():
 
 #allow some of these flags to be modified on the command line:
 SWITCHES = ["enc_x264", "enc_x265", "enc_ffmpeg",
-            "nvenc7", "nvapi",
+            "nvenc7",
             "vpx", "pillow",
             "v4l2",
             "dec_avcodec2", "csc_swscale",
@@ -2175,11 +2174,6 @@ def osx_pkgconfig(*pkgs_options, **ekw):
 
 toggle_packages(nvenc7_ENABLED, "xpra.codecs.nvenc7")
 toggle_packages(nvenc7_ENABLED, "xpra.codecs.cuda_common", "xpra.codecs.nv_util")
-if nvapi_ENABLED:
-    cython_add(Extension("xpra.codecs.nvapi_version",
-                ["xpra/codecs/nvapi_version.pyx"],
-                **pkgconfig("nvapi")
-                ))
 
 if nvenc7_ENABLED:
     #find nvcc:

src/win32/BUILD_CUDA_KERNEL.BAT

@@ -10,9 +10,14 @@ SET CUDA_SRC=xpra\codecs\cuda_common\%KERNEL%.cu
 SET CUDA_BIN=xpra\codecs\cuda_common\%KERNEL%.fatbin
 
 REM nvcc -fatbin -c %CUDA_SRC% -o %CUDA_BIN% --use-local-env --cl-version %CL_VERSION% -gencode=arch=compute_62,code=sm_62 --machine 64 -v
-ECHO nvcc -fatbin -c %CUDA_SRC% -o %CUDA_BIN% --use-local-env --cl-version %CL_VERSION% -gencode=arch=compute_62,code=sm_62 --machine 64
-nvcc -fatbin -c %CUDA_SRC% -o %CUDA_BIN% --use-local-env --cl-version %CL_VERSION% -gencode=arch=compute_62,code=sm_62 --machine 64
-REM done
+nvcc -fatbin -c %CUDA_SRC% -o %CUDA_BIN% ^
+	--use-local-env --cl-version %CL_VERSION% --machine 64 ^
+	-gencode=arch=compute_50,code=sm_50 ^
+	-gencode=arch=compute_52,code=sm_52 ^
+	-gencode=arch=compute_53,code=sm_53 ^
+	-gencode=arch=compute_60,code=sm_60 ^
+	-gencode=arch=compute_61,code=sm_61 ^
+	-gencode=arch=compute_62,code=sm_62
 
 ENDLOCAL
 SET ERRORLEVEL=0

src/win32/PY27_MINGW_BUILD.sh

@@ -58,8 +58,9 @@ fi
 
 if [ "${DO_CUDA}" == "1" ]; then
 	echo "* Building CUDA kernels"
-	cmd.exe //c "win32\\BUILD_CUDA_KERNEL" BGRA_to_NV12
-	cmd.exe //c "win32\\BUILD_CUDA_KERNEL" BGRA_to_YUV444
+	cmd.exe //c "win32\\BUILD_CUDA_KERNEL" BGRA_to_NV12 || exit 1
+	cmd.exe //c "win32\\BUILD_CUDA_KERNEL" BGRA_to_YUV444 || exit 1
+	echo
 fi
 
 echo "* Building Python 2.7 Cython modules"

src/xpra/codecs/cuda_common/cuda_context.py

@@ -338,13 +338,17 @@ def get_CUDA_function(device_id, function_name):
         log("get_CUDA_function(%s, %s) cubin file=%s", device_id, function_name, cubin_file)
         data = load_binary_file(cubin_file)
         if not data:
-            log.error("failed to load CUDA bin file %s", cubin_file)
+            log.error("Error: failed to load CUDA bin file '%s'", cubin_file)
             return None
         log(" loaded %s bytes", len(data))
         KERNELS[function_name] = data
     #now load from cubin:
     start = time.time()
-    mod = driver.module_from_buffer(data)
+    try:
+        mod = driver.module_from_buffer(data)
+    except Exception as e:
+        log.error("Error: failed to load module from buffer for '%s'", function_name)
+        return None
     log("get_CUDA_function(%s, %s) module=%s", device_id, function_name, mod)
     try:
         CUDA_function = mod.get_function(function_name)

src/xpra/codecs/nv_util.py

@@ -8,7 +8,6 @@
 import os
 from xpra.log import Logger
 log = Logger("encoder", "nvenc")
-from xpra.os_util import WIN32
 from xpra.util import pver, print_nested_dict, engs, envbool
 
 
@@ -53,21 +52,7 @@ def get_proc_driver_version():
 
 
 def identify_nvidia_module_version():
-    if os.name!="posix":
-        if not WIN32:
-            log.warn("Warning: unable to identify the NVidia driver version on this platform")
-            return None
-        #try the nvapi call:
-        try:
-            from xpra.codecs.nvapi_version import get_driver_version    #@UnresolvedImport
-            v = get_driver_version()
-            log("NVAPI get_driver_version()=%s", v)
-        except Exception as e:
-            log.warn("Warning: failed to get the driver version through NVAPI:")
-            log.warn(" %s", e)
-            v = []
-    else:
-        v = get_nvml_driver_version() or get_proc_driver_version()
+    v = get_nvml_driver_version() or get_proc_driver_version()
     #only keep numeric values:
     numver = []
     try: