將舊的cuda代碼移植到7.5

Question

與cuda 3.5兼容的舊源代碼與新的cuda版本不兼容，例如， 7.5。例如，

cuda_runtime_api.cc: In function ‘cudaError_t cudaMemcpyToSymbol(const char*, const void*, size_t, size_t, cudaMemcpyKind)’: 
cuda_runtime_api.cc:556:141: error: declaration of C function ‘cudaError_t cudaMemcpyToSymbol(const char*, const void*, size_t, size_t, cudaMemcpyKind)’ conflicts with 
__host__ cudaError_t CUDARTAPI cudaMemcpyToSymbol(const char *symbol, const void *src, size_t count, size_t offset, enum cudaMemcpyKind kind) 
                                     ^
In file included from cuda_runtime_api.cc:155:0: 
/usr/local/cuda/include/cuda_runtime_api.h:4335:39: error: previous declaration ‘cudaError_t cudaMemcpyToSymbol(const void*, const void*, size_t, size_t, cudaMemcpyKind)’ here 
extern __host__ cudaError_t CUDARTAPI cudaMemcpyToSymbol(const void *symbol, const void *src, size_t count, size_t offset __dv(0), enum cudaMemcpyKind kind __dv(cudaMemcpyHostToDevice)); 

所以說，代碼中的cudaMemcpyToSymbol（）與7.5實現不兼容。源代碼是

__host__ cudaError_t CUDARTAPI cudaMemcpyToSymbol(const char *symbol, const void *src, size_t count, size_t offset, enum cudaMemcpyKind kind) 
{....} 

和7.5版本的/usr/local/cuda/include/cuda_runtime_api.h是

extern __host__ cudaError_t CUDARTAPI cudaMemcpyToSymbol(const void *symbol, const void *src, size_t count, size_t offset __dv(0), enum cudaMemcpyKind kind __dv(cudaMemcpyHostToDevice)); 

我該如何解決呢？

更新：錯誤文件中的cuda_runtime_api.cc指向用戶定義的錯誤文件here。

mahmood@u1604:~/gem5-gpu/benchmarks/libcuda$ g++ -std=c++0x -O3 -g -Wall -fPIC -DCUDART_VERSION=7050 -I./ -I/usr/local/cuda/include -I../../gem5/util/m5 -c cuda_runtime_api.cc -o cuda_runtime_api.o 
cuda_runtime_api.cc: In function ‘cudaError_t cudaMemcpyToSymbol(const char*, const void*, size_t, size_t, cudaMemcpyKind)’: 
cuda_runtime_api.cc:556:143: error: declaration of C function ‘cudaError_t cudaMemcpyToSymbol(const char*, const void*, size_t, size_t, cudaMemcpyKind)’ conflicts with 
__host__ cudaError_t CUDARTAPI cudaMemcpyToSymbol(const char *symbol, const void *src, size_t count, size_t offset , enum cudaMemcpyKind kind) 
                                      ^
In file included from cuda_runtime_api.cc:155:0: 
/usr/local/cuda/include/cuda_runtime_api.h:4335:39: error: previous declaration ‘cudaError_t cudaMemcpyToSymbol(const void*, const void*, size_t, size_t, cudaMemcpyKind)’ here 
extern __host__ cudaError_t CUDARTAPI cudaMemcpyToSymbol(const void *symbol, const void *src, size_t count, size_t offset __dv(0), enum cudaMemcpyKind kind __dv(cudaMemcpyHostToDevice)); 
           ^

Answer 1

在註釋之後，符號本身必須被傳遞，而不是字符串中的名稱。下面是關於如何使用這個新的API簽名的說明例子：

#include "cuda_runtime.h" 
#include "device_launch_parameters.h" 

#include <stdio.h> 

__constant__ int a[1]; 

__global__ void kernel() 
{ 
    printf("a = %d\n", a[0]); 
} 

int main() 
{ 
    int value = 42; 
    cudaMemcpyToSymbol(a, &value, sizeof(int)); 
    kernel << <1, 1 >> >(); 
    return cudaDeviceSynchronize(); 
} 

應該打印

a = 42