I have 2 files that form a small CUDA library from my previous program (which works well, btw) written on C++.
The header for this library is:
#ifndef __cudaLU__
#define __cudaLU__
#include <assert.h>
#include <cuda_runtime.h>
#include <cusolverDn.h>
#include <cusolverSp.h>
#include <cusparse.h>
#include <cuComplex.h>
#include <stdlib.h>
void denseLS(int dim,
std::complex<float> * A,
std::complex<float> * b );
void sparseLS(int dim,
std::complex<float> *csrVal,
int *csrRowPtr,
int *csrColInd,
std::complex<float> *vecVal);
#endif
And I want to use this library in my old-as-the-hills C program just by setting procedure in the head of my main.c file:
extern void denseLS(int dim, float complex *A, float complex *b);
And it fails with a bunch of similar errors. Few of them are:
..NA/cudaLS.cu(115): error: namespace "std" has no member "complex"
..NA/cudaLS.cu(115): error: expected a ")"
..NA/cudaLS.cu(137): error: identifier "csrRowPtr" is undefined
..NA/cudaLS.cu(169): error: identifier "csrColInd" is undefined
..NA/cudaLS.cu(170): error: identifier "csrVal" is undefined
..NA/cudaLS.cu(171): error: identifier "vecVal" is undefined
I tried to make a change std::complex -> float complex and nothing works. Still same errors (without std error, ofc).
The cmake instructions file
cmake_minimum_required(VERSION 3.8)
project(NA)
set(CMAKE_C_STANDARD 11)
find_package(GSL REQUIRED)
find_package(CUDA REQUIRED)
include_directories("${CUDA_INCLUDE_DIRS}")
cuda_add_library(solvers STATIC
cudaLS.cu
cudaLS.h)
target_link_libraries(solvers ${CUDA_LIBRARIES} ${CUDA_cusparse_LIBRARY} ${CUDA_cusolver_LIBRARY})
target_compile_features(solvers PUBLIC cxx_std_11)
set_target_properties( solvers
PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
add_executable(NA main.c)
set_target_properties(NA PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
target_link_libraries(NA PRIVATE GSL::gsl m solvers)
What am I doing wrong pals?
UPD:
g++/gcc - 7.3
Linux
Well, I found what exactly I did in a wrong way.
Cmake is OK. But headers in the .h file have to be modified to
extern "C" void denseLS(int dim, cuComplex *A, cuComplex *b );
The cuda functions in .c have to be decleared in the head (or separate .h-file) as
void denseLS(int dim, float complex *A, float complex *b);
Related
I want statically declare a global variable with __device__ qualifier. In the same time I want to store functions intended to GPU in a separate file.
However, if I do so, the variable value is not transferred to GPU -- there are no errors in compilation or execution time, but memcpy functions do nothing.
When I move kernel function into the file with the host code, everything works.
I am sure, that it should be possible to split host and device functions into separate files in this case, but how to do this? I have seen just examples, when kernels and host code are in the same file.
I would be also very thankful, if somebody explained, why does it behaves so.
A sample code is listed below.
Thank you in advance.
Working directory:
$ ls
functionsGPU.cu functionsGPU.cuh staticGlobalMemory.cu
staticGlobalMemory.cu:
#include "functionsGPU.cuh"
#if VARIANT == 2
__global__ void checkGlobalVariable(){
printf("Old value (dev): %f\n", devData);
devData += 2.0f;
printf("New value (dev): %f\n", devData);
}
#endif
int main(int argc, char **argv){
int dev = 0;
float val = 3.2;
cudaSetDevice(dev);
printf("---------\nVARIANT %i\n---------\n", VARIANT);
printf("Old value (host): %f\n", val);
cudaMemcpyToSymbol(devData, &val, sizeof(float));
checkGlobalVariable <<<1, 1>>> ();
cudaMemcpyFromSymbol(&val, devData, sizeof(float));
printf("New value (host): %f\n", val);
cudaDeviceReset();
return 0;
}
functionsGPU.cuh:
#ifndef FUNCTIONSGPU_CUH
#define FUNCTIONSGPU_CUH
#include <cuda_runtime.h>
#include <stdio.h>
#define VARIANT 1
__device__ float devData;
#if VARIANT == 1
__global__ void checkGlobalVariable();
#endif
#endif
functionsGPU.cu:
#include "functionsGPU.cuh"
#if VARIANT == 1
__global__ void checkGlobalVariable(){
printf("Old value (dev): %f\n", devData);
devData += 2.0f;
printf("New value (dev): %f\n", devData);
}
#endif
This is compiled as
$ nvcc -arch=sm_61 staticGlobalMemory.cu functionsGPU.cu -o staticGlobalMemory
Output if the kernel and host code are in separate files (incorrect):
---------
VARIANT 1
---------
Old value (host): 3.200000
Old value (dev): 0.000000
New value (dev): 2.000000
New value (host): 3.200000
Output if the kernel and host code are in the same file (correct):
---------
VARIANT 2
---------
Old value (host): 3.200000
Old value (dev): 3.200000
New value (dev): 5.200000
New value (host): 5.200000
Your code structure, where device code in one compilation unit references device code or device entities in another compilation unit, will require CUDA relocatable device code compilation and linking.
In the case of __device__ variables such as what you have here:
Add -rdc=true to enable this, to your nvcc compilation command line
Add extern in front of the definition of devData, in functionsGPU.cuh
Add __device__ float devData; to staticGlobalMemory.cu
In the case of linking to a __device__ function in a separate file, along with providing the prototype typically via a header file like you would with any function in C++, you also need to add -rdc=true to your nvcc compilation command line, to enable device code linking. Steps 2 and 3 above are not needed.
That should fix the issue. Step 1 provides the necessary cross-module linkage, and steps 2 and 3 will fix the duplicate definition problem you would have, since you are including the same variable via a header file in separate compilation units.
For a reference of how to do the device code compilation setting in windows visual studio, see here.
I'm writing a CUDA kernel that is compiled at runtime using NVRTC (CUDA version 9.2 with NVRTC version 7.5), which needs the stdint.h header, in order to have the int32_t etc. types.
If I write the kernel source code without the include, it works correctly. For example the kernel
extern "C" __global__ void f() { ... }
Compiles to PTX code where f is defined as .visible .entry f.
But if the kernel source code is
#include <stdint.h>
extern "C" __global__ void f() { ... }
it reports A function without execution space annotations (__host__/__device__/__global__) is considered a host function, and host functions are not allowed in JIT mode. (also without extern "C").
Passing -default-device makes the PTX code .visible .func f, so the function cannot be called from the host.
Is there a way to include headers in the source code, and still have a __global__ entry function? Or alternately, a way to know which integer size convention is used on the by the NVRTC compiler, so that the int32_t etc. types can be manually defined?
Edit:
Example program that shows the problem:
#include <cstdlib>
#include <string>
#include <vector>
#include <memory>
#include <cassert>
#include <iostream>
#include <cuda.h>
#include <cuda_runtime.h>
#include <nvrtc.h>
[[noreturn]] void fail(const std::string& msg, int code) {
std::cerr << "error: " << msg << " (" << code << ')' << std::endl;
std::exit(EXIT_FAILURE);
}
std::unique_ptr<char[]> compile_to_ptx(const char* program_source) {
nvrtcResult rv;
// create nvrtc program
nvrtcProgram prog;
rv = nvrtcCreateProgram(
&prog,
program_source,
"program.cu",
0,
nullptr,
nullptr
);
if(rv != NVRTC_SUCCESS) fail("nvrtcCreateProgram", rv);
// compile nvrtc program
std::vector<const char*> options = {
"--gpu-architecture=compute_30"
};
//options.push_back("-default-device");
rv = nvrtcCompileProgram(prog, options.size(), options.data());
if(rv != NVRTC_SUCCESS) {
std::size_t log_size;
rv = nvrtcGetProgramLogSize(prog, &log_size);
if(rv != NVRTC_SUCCESS) fail("nvrtcGetProgramLogSize", rv);
auto log = std::make_unique<char[]>(log_size);
rv = nvrtcGetProgramLog(prog, log.get());
if(rv != NVRTC_SUCCESS) fail("nvrtcGetProgramLog", rv);
assert(log[log_size - 1] == '\0');
std::cerr << "Compile error; log:\n" << log.get() << std::endl;
fail("nvrtcCompileProgram", rv);
}
// get ptx code
std::size_t ptx_size;
rv = nvrtcGetPTXSize(prog, &ptx_size);
if(rv != NVRTC_SUCCESS) fail("nvrtcGetPTXSize", rv);
auto ptx = std::make_unique<char[]>(ptx_size);
rv = nvrtcGetPTX(prog, ptx.get());
if(rv != NVRTC_SUCCESS) fail("nvrtcGetPTX", rv);
assert(ptx[ptx_size - 1] == '\0');
nvrtcDestroyProgram(&prog);
return ptx;
}
const char program_source[] = R"%%%(
//#include <stdint.h>
extern "C" __global__ void f(int* in, int* out) {
out[threadIdx.x] = in[threadIdx.x];
}
)%%%";
int main() {
CUresult rv;
// initialize CUDA
rv = cuInit(0);
if(rv != CUDA_SUCCESS) fail("cuInit", rv);
// compile program to ptx
auto ptx = compile_to_ptx(program_source);
std::cout << "PTX code:\n" << ptx.get() << std::endl;
}
When //#include <stdint.h> in the kernel source is uncommented it no longer compiles. When //options.push_back("-default-device"); is uncommented it compiles but does not mark the function f as .entry.
CMakeLists.txt to compile it (needs CUDA driver API + NVRTC)
cmake_minimum_required(VERSION 3.4)
project(cudabug CXX)
find_package(CUDA REQUIRED)
set(CMAKE_CXX_STANDARD 14)
set(CMAKE_CXX_STANDARD_REQUIRED 14)
add_executable(cudabug cudabug.cc)
include_directories(SYSTEM ${CUDA_INCLUDE_DIRS})
link_directories(${CUDA_LIBRARY_DIRS})
target_link_libraries(cudabug PUBLIC ${CUDA_LIBRARIES} nvrtc cuda)
[Preface: this is a very hacky answer, and is specific to the GNU toolchain (although I suspect the problem in the question is also specific to the GNU toolchain)].
It would appear that the problem here is with the GNU standard header features.h, which gets pulled into stdint.h and then winds up defining a lot of stub functions which have the default __host__ compilation space. This causes nvrtc to blow up. It also seems that the -default-device option will result in a resolved glibC compiler feature set which makes the whole nvrtc compiler fail.
You can defeat this (in a very hacky way) by predefining a feature set for the standard library which excludes all the host functions. Changing your JIT kernel code to
const char program_source[] = R"%%%(
#define __ASSEMBLER__
#define __extension__
#include <stdint.h>
extern "C" __global__ void f(int32_t* in, int32_t* out) {
out[threadIdx.x] = in[threadIdx.x];
}
)%%%";
got me this:
$ nvcc -std=c++14 -ccbin=g++-7 jit_header.cu -o jitheader -lnvrtc -lcuda
$ ./jitheader
PTX code:
//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-24330188
// Cuda compilation tools, release 9.2, V9.2.148
// Based on LLVM 3.4svn
//
.version 6.2
.target sm_30
.address_size 64
// .globl f
.visible .entry f(
.param .u64 f_param_0,
.param .u64 f_param_1
)
{
.reg .b32 %r<3>;
.reg .b64 %rd<8>;
ld.param.u64 %rd1, [f_param_0];
ld.param.u64 %rd2, [f_param_1];
cvta.to.global.u64 %rd3, %rd2;
cvta.to.global.u64 %rd4, %rd1;
mov.u32 %r1, %tid.x;
mul.wide.u32 %rd5, %r1, 4;
add.s64 %rd6, %rd4, %rd5;
ld.global.u32 %r2, [%rd6];
add.s64 %rd7, %rd3, %rd5;
st.global.u32 [%rd7], %r2;
ret;
}
Big caveat: This worked on the glibC system I tried it on. It probably won't work with other toolchains or libC implementations (if, indeed, they have this problem).
Another alternative is creating stand-ins, for some of the standard library headers. NVRTC's API supports your specifying header file contents as strings, associated with header names - before it will go looking through the filesystem for you. This approach is adopted in NVIDIA JITify, and I've adopted it myself working on something else which may or may not be released.
The easy way to do this You can just take the JITify header stubs for stdint.h, limits.h , from here, which I'm also attaching since it's not very long. Alternatively, you can generate this stub yourself to make sure you're not missing out on anything that's relevant from the standard. Here's how that works:
Start with your stdint.h file (or cstdint file as the case may be);
For each include directive in the file (and recursively, for each include in an include etc):
2.1 Figure out whether you can skip including the file altogether (possibly by making a few defines which are known to hold on the GPU).
2.2 If you're not sure you can skip the file - include it entirely and recurse to (2.), or keep it as its own separate header (and apply the whole process in (1.) to it).
You now have a header file which only includes device-safe header files (or none at all)
Partially-preprocess the file, dropping everything that won't be used on a GPU
Remove the lines which might be problematic on a GPU (e.g. #pragma's), and add __device__ __host__ or just __host__ as appropriate to each function declaration.
Important note: Doing this requires paying attention to licenses and copyrights. You would be creating a "derivative work" of glibc and/or JITify and/or StackOverflow contributions etc.
Now, the stdint.h and limits.h from NVIDIA JITify I promised. I've adapted them to not have namespaces:
stdint.h:
#pragma once
#include <limits.h>
typedef signed char int8_t;
typedef signed short int16_t;
typedef signed int int32_t;
typedef signed long long int64_t;
typedef signed char int_fast8_t;
typedef signed short int_fast16_t;
typedef signed int int_fast32_t;
typedef signed long long int_fast64_t;
typedef signed char int_least8_t;
typedef signed short int_least16_t;
typedef signed int int_least32_t;
typedef signed long long int_least64_t;
typedef signed long long intmax_t;
typedef signed long intptr_t; //optional
typedef unsigned char uint8_t;
typedef unsigned short uint16_t;
typedef unsigned int uint32_t;
typedef unsigned long long uint64_t;
typedef unsigned char uint_fast8_t;
typedef unsigned short uint_fast16_t;
typedef unsigned int uint_fast32_t;
typedef unsigned long long uint_fast64_t;
typedef unsigned char uint_least8_t;
typedef unsigned short uint_least16_t;
typedef unsigned int uint_least32_t;
typedef unsigned long long uint_least64_t;
typedef unsigned long long uintmax_t;
#define INT8_MIN SCHAR_MIN
#define INT16_MIN SHRT_MIN
#if defined _WIN32 || defined _WIN64
#define WCHAR_MIN SHRT_MIN
#define WCHAR_MAX SHRT_MAX
typedef unsigned long long uintptr_t; //optional
#else
#define WCHAR_MIN INT_MIN
#define WCHAR_MAX INT_MAX
typedef unsigned long uintptr_t; //optional
#endif
#define INT32_MIN INT_MIN
#define INT64_MIN LLONG_MIN
#define INT8_MAX SCHAR_MAX
#define INT16_MAX SHRT_MAX
#define INT32_MAX INT_MAX
#define INT64_MAX LLONG_MAX
#define UINT8_MAX UCHAR_MAX
#define UINT16_MAX USHRT_MAX
#define UINT32_MAX UINT_MAX
#define UINT64_MAX ULLONG_MAX
#define INTPTR_MIN LONG_MIN
#define INTMAX_MIN LLONG_MIN
#define INTPTR_MAX LONG_MAX
#define INTMAX_MAX LLONG_MAX
#define UINTPTR_MAX ULONG_MAX
#define UINTMAX_MAX ULLONG_MAX
#define PTRDIFF_MIN INTPTR_MIN
#define PTRDIFF_MAX INTPTR_MAX
#define SIZE_MAX UINT64_MAX
limits.h:
#pragma once
#if defined _WIN32 || defined _WIN64
#define __WORDSIZE 32
#else
#if defined __x86_64__ && !defined __ILP32__
#define __WORDSIZE 64
#else
#define __WORDSIZE 32
#endif
#endif
#define MB_LEN_MAX 16
#define CHAR_BIT 8
#define SCHAR_MIN (-128)
#define SCHAR_MAX 127
#define UCHAR_MAX 255
enum {
_JITIFY_CHAR_IS_UNSIGNED = (char)-1 >= 0,
CHAR_MIN = _JITIFY_CHAR_IS_UNSIGNED ? 0 : SCHAR_MIN,
CHAR_MAX = _JITIFY_CHAR_IS_UNSIGNED ? UCHAR_MAX : SCHAR_MAX,
};
#define SHRT_MIN (-32768)
#define SHRT_MAX 32767
#define USHRT_MAX 65535
#define INT_MIN (-INT_MAX - 1)
#define INT_MAX 2147483647
#define UINT_MAX 4294967295U
#if __WORDSIZE == 64
# define LONG_MAX 9223372036854775807L
#else
# define LONG_MAX 2147483647L
#endif
#define LONG_MIN (-LONG_MAX - 1L)
#if __WORDSIZE == 64
#define ULONG_MAX 18446744073709551615UL
#else
#define ULONG_MAX 4294967295UL
#endif
#define LLONG_MAX 9223372036854775807LL
#define LLONG_MIN (-LLONG_MAX - 1LL)
#define ULLONG_MAX 18446744073709551615ULL
I have written cuda function for dense sampling of feature points but i am getting error. My cuda code is given below. I am using cuda 7.5 toolkit.
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include "opencv2/imgproc/imgproc.hpp"
#include <opencv2/gpu/gpu.hpp>
#include <opencv2/opencv.hpp>
using namespace cv::gpu;
using namespace cv;
using namespace std;
__global__ void densefun(std::vector<int>* d_counters,std::vector<Point2f>* d_points,int d_x_max,int d_y_max,int width, int min_distance)
{
int i = blockDim.x * blockIdx.x + threadIdx.x;
Point2f point = (*d_points)[i];
int x = cvFloor(point.x);
int y = cvFloor(point.y);
//if(x >= d_x_max || y >= d_y_max)
//continue;
x /= min_distance;
y /= min_distance;
(*d_counters)[y*width+x]++;
}
void dense(std::vector<int>& counters,std::vector<Point2f>& points,int x_max,int y_max,int width)
{
std::vector<int>* d_counters;
std::vector<Point2f>* d_points;
int min_distance=5;
cudaMalloc(&d_counters,counters.size());
cudaMalloc(&d_points,points.size());
cudaMemcpy(d_points, &points, points.size(), cudaMemcpyHostToDevice);
densefun<<<1,points.size()>>>(d_counters,d_points,x_max,y_max,width,min_distance);
cudaMemcpy(&counters, d_counters, counters.size(), cudaMemcpyDeviceToHost);
cudaFree(d_counters);
cudaFree(d_points);
}
Output:
/home/supriya/Desktop/5Dec/CalculateFV_merged_gpu_old/build/denseCuda.cu(28):
error: calling a host function("cv::Point_ ::Point_") from
a global function("densefun") is not allowed
/home/supriya/Desktop/5Dec/CalculateFV_merged_gpu_old/build/denseCuda.cu(28):
error: calling a host function("std::vector ,
std::allocator > > ::operator []") from a global
function("densefun") is not allowed
/home/supriya/Desktop/5Dec/CalculateFV_merged_gpu_old/build/denseCuda.cu(29)
(col. 7): error: calling a host function("cvFloor") from a
global function("densefun") is not allowed
/home/supriya/Desktop/5Dec/CalculateFV_merged_gpu_old/build/denseCuda.cu(30)
(col. 7): error: calling a host function("cvFloor") from a
global function("densefun") is not allowed
/home/supriya/Desktop/5Dec/CalculateFV_merged_gpu_old/build/denseCuda.cu(35):
error: calling a host function("std::vector > ::operator []") from a global
function("densefun") is not allowed
5 errors detected in the compilation of
"/tmp/tmpxft_00000c85_00000000-7_denseCuda.cpp1.ii". CMake Error at
testVideo_generated_denseCuda.cu.o.cmake:260 (message): Error
generating file
/home/supriya/Desktop/5Dec/CalculateFV_merged_gpu_old/build/CMakeFiles/testVideo.dir//./testVideo_generated_denseCuda.cu.o
CMakeFiles/testVideo.dir/build.make:392: recipe for target
'CMakeFiles/testVideo.dir/./testVideo_generated_denseCuda.cu.o' failed
make[2]: *
[CMakeFiles/testVideo.dir/./testVideo_generated_denseCuda.cu.o] Error
1 CMakeFiles/Makefile2:130: recipe for target
'CMakeFiles/testVideo.dir/all' failed make[1]: *
[CMakeFiles/testVideo.dir/all] Error 2 Makefile:76: recipe for target
'all' failed make: *** [all] Error 2
You cannot use C++ standard library, OpenCV or any other non-CUDA specific library inside a CUDA kernel.
Instead of std::vector you need to use raw pointers to arrays allocated on the device, instead of Point2f you need to use CUDA specific vector type float2, instead of cvFloor you need to use __device__ floorf() and so on.
I am trying to run the cuSolver library available in cuda 7.0. I have an issue with using the cuSolver library that must be very simple to fix, but here I am asking for some help.
I have looked at quite a few examples posted around and I chose in particular this one from JackOLantern:
Parallel implementation for multiple SVDs using CUDA
I have just reduced it to a kernel_0.cu:
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include<iostream>
#include<iomanip>
#include<stdlib.h>
#include<stdio.h>
#include<assert.h>
#include<math.h>
#include <cusolverDn.h>
#include <cuda_runtime_api.h>
#include "Utilities.cuh"
/********/
/* MAIN */
/********/
int main(){
// --- gesvd only supports Nrows >= Ncols
// --- column major memory ordering
// --- cuSOLVE input/output parameters/arrays
int *devInfo; gpuErrchk(cudaMalloc(&devInfo, sizeof(int)));
// --- CUDA solver initialization
cusolverDnHandle_t solver_handle;
cusolverDnCreate(&solver_handle);
cusolverDnDestroy(solver_handle);
return 0;
}
I use the same Utilities.cuh and Utilities.cu as JackOlantern. I compile it as (to be explicit):
/usr/local/cuda-7.0/bin/nvcc kernel_0.cu Utilities.cu
And what I get is:
Utilities.cu(27): warning: conversion from a string literal to "char *" is deprecated
Utilities.cu(27): warning: conversion from a string literal to "char *" is deprecated
/tmp/tmpxft_00007e1d_00000000-22_kernel_0.o: In function `main':
tmpxft_00007e1d_00000000-4_kernel_0.cudafe1.cpp:(.text+0x3d): undefined reference to `cusolverDnCreate'
tmpxft_00007e1d_00000000-4_kernel_0.cudafe1.cpp:(.text+0x49): undefined reference to `cusolverDnDestroy'
collect2: error: ld returned 1 exit status
If I comment out the cusolverDnCreate and cusolverDnDestroy, it compiles fine, so the library is apparently well included.
What simple and basic point am I missing? I have searched around, but I could not fix it. Thanks there.
What simple and basic point am I missing?
You have to link against the cusolver library:
/usr/local/cuda-7.0/bin/nvcc kernel_0.cu Utilities.cu -lcusolver
I am having trouble using cudaMemset on a device variable. Is it possible to use the reference to the device variable for cudaMemset, or is it just a matter of missing compiler flags, or libraries.. I am using cuda 4.1, and
NVRM version: NVIDIA UNIX x86_64 Kernel Module 285.05.33 Thu Jan 19
14:07:02 PST 2012
This is my sample code:
#include <stdio.h>
#include <stdlib.h>
#include <cuda_runtime.h>
// device variable and kernel
__device__ float d_test;
int main() {
if (cudaMemset(&d_test,0,sizeof(float)) !=cudaSuccess)
printf("Error!\n");
}
which outputs:
Error!
Your problem is that d_test (as it appears in the host symbol table) isn't a valid device address and the runtime cannot access it directly. The solution is to use the cudaGetSymbolAddress API function to read the address of the device symbol from the context at runtime. Here is a slightly expanded version of your demonstration case which should work correctly:
#include <stdio.h>
#include <stdlib.h>
#include <cuda_runtime.h>
// device variable and kernel
__device__ float d_test;
inline void gpuAssert(cudaError_t code, char * file, int line, bool Abort=true)
{
if (code != cudaSuccess) {
fprintf(stderr, "GPUassert: %s %s %d\n", cudaGetErrorString(code),file,line);
if (Abort) exit(code);
}
}
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
int main()
{
float * _d_test;
gpuErrchk( cudaFree(0) );
gpuErrchk( cudaGetSymbolAddress((void **)&_d_test, "d_test") );
gpuErrchk( cudaMemset(_d_test,0,sizeof(float)) );
gpuErrchk( cudaThreadExit() );
return 0;
}
Here, we read the address of the device symbol d_test from the context into a host pointer _d_test. This can then be passed to host side API functions like cudaMemset, cudaMemcpy, etc.
Edit to note that the form of cudaGetSymbolAddress shown in this answer has been deprecated and removed from the CUDA runtime API. For modern CUDA, the call would be:
gpuErrchk( cudaGetSymbolAddress((void **)&_d_test, d_test) );
I believe you can also use cudaMemcpyFromSymbol:
A function, such as the following kernel, can change the value of the variable declared in global memory (outside of the main function)
__global__ void kernel1() { d_test = 1.0; }
Inside your main, you can obtain the value using cudaMemcpyFromSymbol
cudaMemcpyFromSymbol(&h_test,"d_test",sizeof(float),0,cudaMemcpyDeviceToHost);
Of course, there is also cudaMemcpyToSymbol to change the value of the global variable.
The idea came from here: Having problem assigning a device variable in CUDA