Compiling CUDA files with Qt 5.12

Ailurus

Hi, I'm trying to compile my Qt project containing a CUDA file (.cu). In QtCreator, I've edited the .pro file based on http://blog.qt.io/blog/2015/03/03/qt-weekly-28-qt-and-cuda-on-the-jetson-tk1/. The relevant part is as follows —

# File(s) containing my CUDA code
CUDA_SOURCES = CodeCUDA.cu

# Location of CUDA on my system (Arch Linux)
CUDA_DIR = /opt/cuda

INCLUDEPATH += $$CUDA_DIR/include
LIBS += -L$$CUDA_DIR/lib64 -lcuda -lcudart

# Compute capability of my GPU
CUDA_ARCH = sm_30

cuda.input = CUDA_SOURCES
cuda.output = ${QMAKE_FILE_BASE}_cuda.o
cuda.commands = $$CUDA_DIR/bin/nvcc -c -arch=$$CUDA_ARCH -o ${QMAKE_FILE_OUT} ${QMAKE_FILE_NAME}
cuda.dependency_type = TYPE_C
cuda.depend_command = $$CUDA_DIR/bin/nvcc -M ${QMAKE_FILE_NAME}

QMAKE_EXTRA_COMPILERS += cuda

Yet when I try to build this, I get the error

:-1: error: No rule to make target 'CodeCUDA.o', needed by 'CodeCUDA_cuda.o'.  Stop.

[Edit]

In the automatically generated Makefile, I see this —

compiler_cuda_make_all: CodeCUDA_cuda.o
compiler_cuda_clean:
	-$(DEL_FILE) CodeCUDA_cuda.o
CodeCUDA_cuda.o: ../OpenGL/CodeCUDA.cu \
		CodeCUDA.o \
		\: \
		../OpenGL/CodeCUDA.cu \
		/ \
		/usr/include/stdc-predef.h \
		/ \

Which is then followed by a long list of CUDA headers. If I manually change this to the following, my project compiles and runs correctly —

compiler_cuda_make_all: CodeCUDA_cuda.o
compiler_cuda_clean:
	-$(DEL_FILE) CodeCUDA_cuda.o
CodeCUDA_cuda.o: ../OpenGL/CodeCUDA.cu \
		/ \
		/usr/include/stdc-predef.h \
		/ \

Any ideas what is causing the 'incorrect' generation of my Makefile?

SGaist

Hi,

What platform are you on ?

Ailurus

I'm on (Arch) Linux, 64-bit.

kshegunov

I don't know why it tries to generate that dependency, but as a long shot can you try without changing the objects file name, i.e.:

cuda.output = ${QMAKE_FILE_BASE}.o

SGaist

Tested it on macOS and it worked correctly.

Can you show the rest of your .pro file ?

Ailurus

Thanks guys for looking into it. @kshegunov, in that case the error changes to

:-1: error: No rule to make target ':', needed by 'CodeCUDA.o'.  Stop.

Part of the Makefile then looks like this —

		../OpenGL/CodeCUDA.cu \
		\: \
		../OpenGL/CodeCUDA.cu \
		/ \
		/usr/include/stdc-predef.h \
		/ \

This is the entire .pro file —

#-------------------------------------------------
#
# Project created by QtCreator 2019-01-07T11:13:21
#
#-------------------------------------------------

QT += core gui

greaterThan(QT_MAJOR_VERSION, 4): QT += widgets

TARGET = TemplateShowMeshOpenGL
TEMPLATE = app

# The following define makes your compiler emit warnings if you use
# any feature of Qt which has been marked as deprecated (the exact warnings
# depend on your compiler). Please consult the documentation of the
# deprecated API in order to know how to port your code away from it.
DEFINES += QT_DEPRECATED_WARNINGS

# You can also make your code fail to compile if you use deprecated APIs.
# In order to do so, uncomment the following line.
# You can also select to disable deprecated APIs only up to a certain version of Qt.
#DEFINES += QT_DISABLE_DEPRECATED_BEFORE=0x060000    # disables all the APIs deprecated before Qt 6.0.0

CONFIG += c++11

SOURCES += \
        main.cpp \
        mainwindow.cpp \
    mainview.cpp

HEADERS += \
        mainwindow.h \
    mainview.h

FORMS += \
        mainwindow.ui

# File(s) containing my CUDA code
CUDA_SOURCES = CodeCUDA.cu \
             KernelVBO.cu

# Location of CUDA on my system (Arch Linux)
CUDA_DIR = /opt/cuda

INCLUDEPATH += $$CUDA_DIR/include
LIBS += -L$$CUDA_DIR/lib64 -lcuda -lcudart

# Compute capability of my GPU
CUDA_ARCH = sm_30

cuda.input = CUDA_SOURCES
cuda.output = ${QMAKE_FILE_BASE}_cuda.o
cuda.commands = $$CUDA_DIR/bin/nvcc -c -arch=$$CUDA_ARCH -o ${QMAKE_FILE_OUT} ${QMAKE_FILE_NAME}
cuda.dependency_type = TYPE_C
cuda.depend_command = $$CUDA_DIR/bin/nvcc -M ${QMAKE_FILE_NAME}

QMAKE_EXTRA_COMPILERS += cuda

# Default rules for deployment.
qnx: target.path = /tmp/$${TARGET}/bin
else: unix:!android: target.path = /opt/$${TARGET}/bin
!isEmpty(target.path): INSTALLS += target

RESOURCES += \
    resources.qrc

I'm using Qt Creator 4.8.1, QMake version 3.1, and Qt version 5.12.0.

SGaist

@Ailurus said in Compiling CUDA files with Qt 5.12:

CUDA_SOURCES

What if you try with:

CUDA_SOURCES += \
             $$PWD/CodeCUDA.cu \
             $$PWD/KernelVBO.cu

Ailurus

Unfortunately that does not make a difference. Hmm, could it be a setting somewhere in Build Settings? Effective qmake call is set to

/usr/bin/qmake /home/pieter/Dropbox/CUDA/OpenGL/TemplateShowMeshOpenGL.pro -spec linux-g++ CONFIG+=debug CONFIG+=qml_debug && /usr/bin/make qmake_all

Manually changing the Makefile is not a big issue, at this point I'm just curious why it is generated like it is.

SGaist

That version of qmake is the one from your distribution and likely not the one from your 5.12 installation.

Use the full path to it to ensure the correct one is used.

Ailurus

Hmm, it's the only qmake I have on my machine. Info —

$ which qmake
/usr/bin/qmake
$ /usr/bin/qmake -v
QMake version 3.1
Using Qt version 5.12.0 in /usr/lib

SGaist

Good to know.

In that case, can you try with the pre-built version ? Just to compare the output.

BrandonL

This is definitely a problem when upgrading from an older version of qmake to qmake 3.1 which is part of Qt 5.12.3.
For some reason qmake 3.1 parses the output of a compile depend_command differently than before, which breaks existing .pro files that use a custom compiler for CUDA. Qmake ends up constructing a dependency list that looks like this:

...
cuda/foo.cuh \
cuda/foo.cu \
buildDir/MyCudaLib.o \ 
\: \               <--------- THIS IS WRONG
cuda/bar.cu \
/ \
/usr/local/cuda/include/cuda_runtime.h \
...

The solution is to pipe the output of nvcc to:

sed \"s/^.*: //\"

to filter out the garbage, ie:

cuda.depend_command = $$CUDA_DIR/bin/nvcc -g -M $$CUDA_INC $$NVCCFLAGS ${QMAKE_FILE_NAME} | sed \"s/^.*: //\"