Include cuda kernel in my project - c++

Question

I have this c++ project in which I call a cuda kernel by means of a wrapper function.
My c++ file looks like this (this is extern.cc):
#include "extern.h"
#include "qc/operator.h"
#include "qc/quStates.h"
#include "gpu.h"
...
ROUTINE(ext_bit) {
int i;
quState *qbit;
PAR_QUSTATE(q,"q");
opBit *op;
tComplex I(0,1);
tComplex sg= inv ? -1 : 1;
char c=(def->id())[0];
if(def->id().length()!=1) c='?';
switch(c) {
case 'H': op=new opBit(1,1,1,-1,sqrt(0.5)); break;
case 'X': op=new opBit(0,1,1,0); break;
case 'Y': op=new opBit(0,-I,I,0); break;
case 'Z': op=new opBit(1,0,0,-1); break;
case 'S': op=new opBit(1,0,0,sg*I); break;
case 'T': op=new opBit(1,0,0,sqrt(0.5)+sg*sqrt(0.5)*I); break;
case '?':
default: EXTERR("unknown single qubit operator "+def->id());
}
// This is where I call my wrapper function
// the error that I get is: expected primary-expression before ',' token
gpucaller(opBit, q);
qcl_delete(op);
return 0;
}
where "gpucaller" is my wrapper function that calls the kernel, both defined in cuda_kernel.cu:
/* compiling with:
nvcc -arch sm_11 -c -I"/home/glu/NVIDIA_GPU_Computing_SDK/C/common/inc" -I"." -I"./qc" -I"/usr/local/cuda/include" -o cuda_kernel.o cuda_kernel.cu
*/
#ifndef _CUDA_KERNEL_H_
#define _CUDA_KERNEL_H_
#define MAX_QUBITS 25
#define BLOCKDIM 512
#define MAX_TERMS_PER_BLOCK (2*BLOCKDIM)
#define THREAD_MASK (~0ul << 1)
// includes
#include <cutil_inline.h>
#include "gpu.h"
__constant__ float devOpBit[2][2];
__global__ void qcl1(cuFloatComplex *a, int N, int qbCount, int blockGrpSize, int k)
{
//int idx = blockIdx.x * BLOCKDIM + threadIdx.x;
//int tx = threadIdx.x;
cuFloatComplex t0_0, t0_1, t1_0, t1_1;
int x0_idx, x1_idx;
int i, grpSize, b0_idx, b1_idx;
__shared__ cuFloatComplex aS[MAX_TERMS_PER_BLOCK];
...
}
void gpucaller(opBit* op, quBaseState* q) {
// make an operator copy
float** myOpBit = (float**)op->getDeviceReadyOpBit();
unsigned int timer = 0;
cuFloatComplex *a_d;
long int N = 1 << q->mapbits();
int size = sizeof(cuFloatComplex) * N;
// start timer
cutilCheckError( cutCreateTimer( &timer));
cutilCheckError( cutStartTimer( timer));
// allocate device memory
cudaMalloc((void**)&a_d,size);
// copy host memory to device
cudaMemcpy(a_d, q->termsarray, size, cudaMemcpyHostToDevice);
// copy quantic operator to constant memory
cutilSafeCall( cudaMemcpyToSymbol(devOpBit, myOpBit, 2*sizeof(float[2]), 0) );
printf("Cuda errors: %s\n", cudaGetErrorString( cudaGetLastError() ) );
// setup execution parameters
dim3 dimBlock(BLOCKDIM, 1, 1);
int n_blocks = N/MAX_TERMS_PER_BLOCK + (N%MAX_TERMS_PER_BLOCK == 0 ? 0:1);
dim3 dimGrid(n_blocks, 1, 1);
...
// execute the kernel
qcl1<<< dimGrid, dimBlock >>>(a_d, N, gates, blockGrpSize, k);
// check if kernel execution generated and error
cutilCheckMsg("Kernel execution failed");
...
// copy result from device to host
cudaMemcpy(q->termsarray, a_d, size, cudaMemcpyDeviceToHost);
// stop timer
cutilCheckError( cutStopTimer( timer));
//printf( "GPU Processing time: %f (ms)\n", cutGetTimerValue( timer));
cutilCheckError( cutDeleteTimer( timer));
// cleanup memory on device
cudaFree(a_d);
cudaThreadExit();
}
#endif // #ifndef _CUDA_KERNEL_H_
and "gpu.h" has the following content:
#ifndef _GPU_H_
#define _GPU_H_
#include "qc/operator.h"
#include "qc/qustates.h"
void gpucaller(opBit* op, quBaseState* q);
#endif // #ifndef _GPU_H_
I don't include the .cu file in my c++ file, I only include the .h file (gpu.h - contains the prototype of my kernel caller function) in both the c++ and the .cu files.
I compile the .cu file with nvcc, and link the resulting .o file in my project's Makefile.
Also, I didn't forget to add the "-lcudart" flag to the Makefile.
The problem is that when I compile my main project, I get this error:
expected primary-expression before ',' token
and is referring to the line in extern.cc where I call the "gpucaller" function.
Does anyone know how to get this right?
EDIT: I've tried compiling again, this time removing the arguments from the gpucaller's function definition (and obviously not passing any arguments to the function, which is wrong because I need to pass arguments). It compiled just fine.
So the problem is that gpucaller's argument types aren't recognized, I have no idea why (I've included the headers where the arguments' types are declared, ie "qc/operator.h" and "qc/quStates.h"). Does anyone have a solution to this?
My project's Makefile is this:
VERSION=0.6.3
# Directory for Standard .qcl files
QCLDIR = /usr/local/lib/qcl
# Path for qcl binaries
QCLBIN = /usr/local/bin
ARCH = `g++ -dumpmachine || echo bin`
# Comment out if you want to compile for a different target architecture
# To build libqc.a, you will also have to edit qc/Makefile!
#ARCH = i686-linux
#ARCHOPT = -m32 -march=i686
# Debugging and optimization options
#DEBUG = -g -pg -DQCL_DEBUG -DQC_DEBUG
#DEBUG = -g -DQCL_DEBUG -DQC_DEBUG
DEBUG = -O2 -g -DQCL_DEBUG -DQC_DEBUG
#DEBUG = -O2
# Plotting support
#
# Comment out if you don't have GNU libplotter and X
PLOPT = -DQCL_PLOT
PLLIB = -L/usr/X11/lib -lplotter
# Readline support
#
# Comment out if you don't have GNU readline on your system
# explicit linking against libtermcap or libncurses may be required
RLOPT = -DQCL_USE_READLINE
#RLLIB = -lreadline
RLLIB = -lreadline -lncurses
# Interrupt support
#
# Comment out if your system doesn't support ANSI C signal handling
IRQOPT = -DQCL_IRQ
# Replace with lex and yacc on non-GNU systems (untested)
LEX = flex
YACC = bison
INSTALL = install
##### You shouldn't have to edit the stuff below #####
DATE = `date +"%y.%m.%d-%H%M"`
QCDIR = qc
QCLIB = $(QCDIR)/libqc.a
QCLINC = lib
#CXX = g++
#CPP = $(CC) -E
CXXFLAGS = -c $(ARCHOPT) -Wall $(DEBUG) $(PLOPT) $(RLOPT) $(IRQOPT) -I$(QCDIR) -DDEF_INCLUDE_PATH="\"$(QCLDIR)\""
LDFLAGS = $(ARCHOPT) -L$(QCDIR) $(DEBUG) $(PLLIB) -lm -lfl -lqc $(RLLIB) -L"/usr/local/cuda/lib" -lcudart
FILESCC = $(wildcard *.cc)
FILESH = $(wildcard *.h)
SOURCE = $(FILESCC) $(FILESH) qcl.lex qcl.y Makefile
OBJECTS = types.o syntax.o typcheck.o symbols.o error.o \
lex.o yacc.o print.o quheap.o extern.o eval.o exec.o \
parse.o options.o debug.o cond.o dump.o plot.o format.o cuda_kernel.o
all: do-it-all
ifeq (.depend,$(wildcard .depend))
include .depend
do-it-all: build
else
do-it-all: dep
$(MAKE)
endif
#### Rules for depend
dep: lex.cc yacc.cc yacc.h $(QCLIB)
for i in *.cc; do \
$(CPP) -I$(QCDIR) -MM $$i; \
done > .depend
lex.cc: qcl.lex yacc.h
$(LEX) -olex.cc qcl.lex
yacc.cc: qcl.y
$(YACC) -t -d -o yacc.cc qcl.y
yacc.h: yacc.cc
mv yacc.*?h yacc.h
$(QCLIB):
cd $(QCDIR) && $(MAKE) libqc.a
#### Rules for build
build: qcl $(QCLINC)/default.qcl
qcl: $(OBJECTS) qcl.o $(QCLIB)
$(CXX) $(OBJECTS) qcl.o $(LDFLAGS) -o qcl
$(QCLINC)/default.qcl: extern.cc
grep "^//!" extern.cc | cut -c5- > $(QCLINC)/default.qcl
checkinst:
[ -f ./qcl -a -f $(QCLINC)/default.qcl ] || $(MAKE) build
install: checkinst
$(INSTALL) -m 0755 -d $(QCLBIN) $(QCLDIR)
$(INSTALL) -m 0755 ./qcl $(QCLBIN)
$(INSTALL) -m 0644 ./$(QCLINC)/*.qcl $(QCLDIR)
uninstall:
-rm -f $(QCLBIN)/qcl
-rm -f $(QCLDIR)/*.qcl
-rmdir $(QCLDIR)
#### Other Functions
edit:
nedit $(SOURCE) &
clean:
rm -f *.o lex.* yacc.*
cd $(QCDIR) && $(MAKE) clean
clear: clean
rm -f qcl $(QCLINC)/default.qcl .depend
cd $(QCDIR) && $(MAKE) clear
dist-src: dep
mkdir qcl-$(VERSION)
cp README CHANGES COPYING .depend $(SOURCE) qcl-$(VERSION)
mkdir qcl-$(VERSION)/qc
cp qc/Makefile qc/*.h qc/*.cc qcl-$(VERSION)/qc
cp -r lib qcl-$(VERSION)
tar czf qcl-$(VERSION).tgz --owner=0 --group=0 qcl-$(VERSION)
rm -r qcl-$(VERSION)
dist-bin: build
mkdir qcl-$(VERSION)-$(ARCH)
cp Makefile README CHANGES COPYING qcl qcl-$(VERSION)-$(ARCH)
cp -r lib qcl-$(VERSION)-$(ARCH)
tar czf qcl-$(VERSION)-$(ARCH).tgz --owner=0 --group=0 qcl-$(VERSION)-$(ARCH)
rm -r qcl-$(VERSION)-$(ARCH)
upload: dist-src
scp qcl-$(VERSION)*.tgz oemer#tph.tuwien.ac.at:html/tgz
scp: dist-src
scp qcl-$(VERSION).tgz oemer#tph.tuwien.ac.at:bak/qcl-$(DATE).tgz
The only changes that I've added to the original Makefile is adding "cuda_kernel.o" to the OBJECTS' line, and adding the "-lcudart" flag to LDFLAGS.
UPDATE: Thanks harrism for helping me out. I was passing a type as a parameter.

Answer 1

gpucaller(opBit, q);
You are passing a type name (opBit) as a function parameter, which is not valid C or C++. It looks like you need to do this instead:
gpucaller(op, q);

Answer 2

Does it literally look like this in cuda.h?
void gpucaller(type1 param1, type2 param2);
Are type1 and type2 declared anywhere so that your regular C++ compiler know what these types are? If not, then you'd get an error like you're saying you're getting.

Answer 3

Your problem code is pretty big and complex at the moment. Try to strip it down to a more simple failure case and update your question once you have that. It'll make it easier to attempt reproduction. Strip out the cuda timer code, the switch case, replace implementation details with ... where it doesn't matter, etc.
I compile with msvc and nvcc then link with icl; so if you can make a simple example I can see if it compiles with a totally different compiler setup. That should narrow down the problem.
Even though renaming your own header cuda.h to somethingspecific.h didn't help, I don't think it's a good idea to leave it as cuda.h. It's confusing and a potential source of problems.