Related
In CUDA 9.2 I have something like this:
#ifdef __CUDA_ARCH__
struct Context { float n[4]; } context;
#else
typedef __m128 Context;
#endif
struct A { float k[2]; };
struct B { float q[4]; };
struct FTransform : thrust::unary_function<A, B>
{
const Context context;
FTransform(Context context) : context(context){}
__device__ __host__ B operator()(const A& a) const
{
B b{{a.k[0], a.k[1], a.k[0]*context.n[0], a.k[1]*context.n[1]}};
return b;
}
};
void DoThrust(B* _bs, const Context& context, A* _as, uint32_t count)
{
thrust::device_ptr<B> bs = thrust::device_pointer_cast(_bs);
thrust::device_ptr<A> as = thrust::device_pointer_cast(_as);
FTransform fTransform(context);
auto first = thrust::make_transform_iterator(as, fTransform);
auto last = thrust::make_transform_iterator(as + count, fTransform);
thrust::copy(first, last, bs);
}
int main(int c, char **argv)
{
const uint32_t Count = 4;
Context context;
A* as;
B* bs;
cudaMalloc(&as, Count*sizeof(A));
cudaMalloc(&bs, Count*sizeof(B));
A hostAs[Count];
cudaMemcpy(as, hostAs, Count * sizeof(A), cudaMemcpyHostToDevice);
DoThrust(bs, context, as, Count);
B hostBs[Count];
cudaMemcpy(hostBs, bs, Count * sizeof(B), cudaMemcpyDeviceToHost);//crash
return 0;
}
Then when I call a standard cudaMemcpy() call later on the results I get the exception "an illegal memory access was encountered".
If I replace the thrust code with a non-thrust equivalent there is no error and everything works fine. Various combinations of trying to copy to device_vectors etc I get different crashes that seem to be thrust trying to release the device_ptr's for some reason - so maybe it is here for some reason?
== UPDATE ==
Ok that was confusing it appears it's due to the functor FTransform context member variable in my actual more complicated case. This specifically:
struct FTransform : thrust::unary_function<A, B>
{
#ifdef __CUDA_ARCH__
struct Context { float v[4]; } context;
#else
__m128 context;
#endif
...
};
So I guess it's an alignment problem somehow => in fact it is, as this works:
#ifdef __CUDA_ARCH__
struct __align__(16) Context { float v[4]; } context;
#else
__m128 context;
#endif
The solution is to ensure that if you use aligned types in thrust functor members (such as __m128 SSE types) that are copied to the GPU, that they are defined as aligned both during NVCC's CPU and GPU code build passes - and not accidentally assume even if a type may seem to naturally align to it's equivalent in the other pass that it will be ok, as otherwise bad hard to understand things may happen.
So for example the _ align _(16) is necessary in code like this:
struct FTransform : thrust::unary_function<A, B>
{
#ifdef __CUDA_ARCH__
struct __align__(16) Context { float v[4]; } context;
#else
__m128 context;
#endif
FTransform(Context context) : context(context){}
__device__ __host__ B operator()(const A& a) const; // function makes use of context
};
I would like to create a list of function pointers dynamically on the CPU (with some sort of push_back() method called from main()) and copy it to a GPU __constant__ or __device__ array, without needing to resort to static __device__ function pointers. I believe this question is related to my problem; however, my goal is to create the __host__ function pointer array iteratively and then copy it to the __constant__ function pointer array instead of initialising the latter on declaration.
A working code example with static function pointers (as seen here or here) would be:
common.h:
#ifndef COMMON_H
#define COMMON_H
#include <stdio.h>
#include <iostream>
#define num_functions 3
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const 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);
}
}
// fptr_t: Pointer to void function that takes two integer lvalues
typedef void (*fptr_t)(int&, int&);
// some examples of void(int&, int&) functions...
__device__ void Add(int &a, int &b) {printf("Add... %i + %i = %i\n", a, b, a+b);}
__device__ void Subtract(int &a, int &b) {printf("Subtract... %i - %i = %i\n", a, b, a-b);}
__device__ void Multiply(int &a, int &b) {printf("Multiply... %i * %i = %i\n", a, b, a*b);}
// List of function pointers in device memory
__constant__ fptr_t constant_fList[num_functions];
// Kernel called from main(): choose the function to apply whose index is equal to thread ID
__global__ void kernel(int a, int b) {
fptr_t f;
if (threadIdx.x < num_functions) {
f = constant_fList[threadIdx.x];
f(a,b);
}
}
#endif
main.cu:
#include "common.h"
// Static device function pointers
__device__ fptr_t p_Add = Add;
__device__ fptr_t p_Sub = Subtract;
__device__ fptr_t p_Mul = Multiply;
// Load function list to constant memory
void loadList_staticpointers() {
fptr_t h_fList[num_functions];
gpuErrchk( cudaMemcpyFromSymbol(&h_fList[0], p_Add, sizeof(fptr_t)) );
gpuErrchk( cudaMemcpyFromSymbol(&h_fList[1], p_Sub, sizeof(fptr_t)) );
gpuErrchk( cudaMemcpyFromSymbol(&h_fList[2], p_Mul, sizeof(fptr_t)) );
gpuErrchk( cudaMemcpyToSymbol(constant_fList, h_fList, num_functions * sizeof(fptr_t)) );
}
int main() {
loadList_staticpointers();
int a = 12, b = 15;
kernel<<<1,3>>>(a, b);
gpuErrchk(cudaGetLastError());
gpuErrchk(cudaDeviceSynchronize());
return 0;
}
Specs: GeForce GTX 670, compiled for -arch=sm_30, CUDA 6.5, Ubuntu 14.04
I wish to avoid the use of static device function pointers, as appending each function would require code maintenance on the user side - declaration of a new static pointer like p_Add or p_Mul, manipulation of void loadList_functionpointers(), etc. To make it clear, I am trying something like the following (crashing) code:
main_wrong.cu:
#include "common.h"
#include <vector>
// Global variable: list of function pointers in host memory
std::vector<fptr_t> vec_fList;
// Add function to functions list
void addFunc(fptr_t f) {vec_fList.push_back(f);}
// Upload the functions in the std::vector<fptr_t> to GPU memory
// Copies CPU-side pointers to constant_fList, therefore crashes on kernel call
void UploadVector() {
fptr_t* h_vpointer = vec_fList.data();
gpuErrchk( cudaMemcpyToSymbol(constant_fList, h_vpointer, vec_fList.size() * sizeof(fptr_t)) );
}
int main() {
addFunc(Add);
addFunc(Subtract);
addFunc(Multiply);
int a = 12, b = 15;
UploadVector();
kernel<<<1,3>>>(a, b); // Wrong to call a host-side function pointer from a kernel
gpuErrchk(cudaGetLastError());
gpuErrchk(cudaDeviceSynchronize());
return 0;
}
My understanding is that function pointers pointing to host addresses are copied to the GPU and are unusable by the kernel, which needs pointers pointing to GPU addresses when the function f(a,b) is called. Populating a host-side array with device-side pointers would work for me with raw data (see this question) but not with function pointers. Trivial attempts with Unified Memory have failed as well... so far, I have only found static device-side pointers to work. Is there no other way to copy a dynamically created CPU array of function pointers onto the GPU?
If you can use C++11 (supported since CUDA 7), you could use the following to auto-generate the function table:
template <fptr_t... Functions>
__global__ void kernel(int a, int b)
{
constexpr auto num_f = sizeof...(Functions);
constexpr fptr_t table[] = { Functions... };
if (threadIdx.x < num_f)
{
fptr_t f = table[threadIdx.x];
f(a,b);
}
}
You would then call this kernel using
kernel<Add, Subtract, Multiply><<<1,3>>>(a, b);
Inspired by m.s.'s answer, I chose to pass the function pointer as a template parameter -this was in fact the key to solve my problem- and discovered that filling a __device__ array of function pointers dev_fList from the main() function iteratively without the help of static function pointers is indeed possible, plus C++11 compatibility is not even needed!
Here is a working example on a __device__ array in global memory. I have not tried its constant memory counterpart yet, but once a global memory array has been satisfactorily created, my guess is that a cudaMemcpyToSymbol(..., cudaMemcpyDeviceToDevice) should do the trick.
A kernel kernel() creates a GPU address for function pointer dev_f and copies the function f that was passed as a template argument. Since this is an iterative process from the CPU, only one thread (thread 0) is involved in this kernel, which is launched with configuration <<<1,1>>>. The static variable count_f takes care of indexing in dev_fList.
common.h:
#ifndef COMMON_H
#define COMMON_H
#include <stdio.h>
#include <iostream>
#define num_functions 3
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const 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);
}
}
// fptr_t: Pointer to void function that takes two integer lvalues
typedef void (*fptr_t)(int&, int&);
// some examples of void(int&, int&) functions...
__device__ void Add(int &a, int &b) {printf("Add... %i + %i = %i\n", a, b, a+b);}
__device__ void Subtract(int &a, int &b) {printf("Subtract... %i - %i = %i\n", a, b, a-b);}
__device__ void Multiply(int &a, int &b) {printf("Multiply... %i * %i = %i\n", a, b, a*b);}
// List of function pointers in device memory
// Note that, in my example, it resides in global memory space, not constant memory
__device__ fptr_t dev_fList[num_functions];
#endif
main.cu:
#include "common.h"
// Index in dev_fList[] == number of times addFunc<>() was launched
static int count_f = 0;
// Kernel that copies function f to the GPU
template<fptr_t f>
__global__ void kernel(int a, int b, int idx) {
fptr_t dev_f = f; // Create device function pointer
dev_fList[idx] = dev_f; // Populate the GPU array of function pointers
dev_fList[idx](a,b); // Make sure that the array was populated correctly
}
// Add function to functions list
template<fptr_t f>
void addFunc(const int &a, const int &b) {
if (count_f >= num_functions) {
std::cout << "Error: not enough memory statically allocated on device!\n";
exit(EXIT_FAILURE);
}
kernel<f><<<1,1>>>(a,b,count_f);
gpuErrchk(cudaGetLastError());
gpuErrchk(cudaDeviceSynchronize());
count_f++;
}
int main() {
int a = 12, b = 15;
addFunc<Add>(a,b);
addFunc<Subtract>(a,b);
addFunc<Multiply>(a,b);
return 0;
}
Edit: Added copy of the array of function pointers to constant memory
For what it's worth, here is how to copy our dev_fList array to constant memory:
In common.h:
__constant__ fptr_t cst_fList[num_functions];
__global__ void cst_test(int a, int b, int idx) {
if (threadIdx.x < idx) cst_fList[threadIdx.x](a,b);
}
In main.cu main() function, after all desired functions have been added:
fptr_t *temp;
gpuErrchk( cudaMemcpyFromSymbol((void**)&temp, dev_fList[0], count_f * sizeof(fptr_t)) );
gpuErrchk( cudaMemcpyToSymbol(cst_fList[0], &temp, count_f * sizeof(fptr_t)) );
cst_test<<<1,count_f>>>(a,b, count_f);
gpuErrchk(cudaGetLastError());
gpuErrchk(cudaDeviceSynchronize());
It may look ugly as I understand that memory is transferred to the host via temp and then back to the device; more elegant suggestions are welcome.
It is impossible to use dynamically created CUDA device function pointers (at least not without crash or UB). The template based solutions work at compile time (not dynamic). The CUDA device function pointer approaches you see everywhere need device symbols in global space. This means that for every function a device function pointer must be already declared. This also means you cannot use normal C function pointers as reference, which are e.g. set at runtime. In comprehension, using CUDA device function pointers is questionable. Template based approaches look user-friendly, but are per definition not dynamic.
Example showing structure with function pointers:
This example shows a structure having some function pointers. In normal C++ code, you can set and change the device function pointers while the program is running (dynamically). With CUDA this example below is impossible, because the function pointers in the struct are no valid device symbols. This means they cannot be used with "cudaMemcpyFromSymbol". To circumvent this, either the original function (target of the function pointers) or global cuda device function pointers must be created. Both is not dynamic.
This is dynamic assignment:
typedef float (*pDistanceFu) (float, float);
typedef float (*pDecayFu) (float, float, float);
// In C++ you can set and reset the function pointer during run time whenever you want ..
struct DistFunction {
/*__host__ __device__*/ pDistanceFu distance; // uncomment for NVCC ..
/*__host__ __device__*/ pDecayFu rad_decay;
/*__host__ __device__*/ pDecayFu lrate_decay;
};
// you can do what you want ..
DistFunction foo, bar;
foo.distance = bar.distance;
// ..
This is how it should be with CUDA, but it will fail, because there is no valid device symbol :(
pDistanceFu hDistance;
pDecayFu hRadDay;
pDecayFu hLRateDecay;
void DeviceAssign(DistFunction &dist) {
cudaMemcpyFromSymbol(&hDistance, dist.distance, sizeof(pDistanceFu) );
cudaMemcpyFromSymbol(&hRadDay, dist.rad_decay, sizeof(pDecayFu) );
cudaMemcpyFromSymbol(&hLRateDecay, dist.lrate_decay, sizeof(pDecayFu) );
dist.distance = hDistance;
dist.rad_decay = hRadDay;
dist.lrate_decay = hLRateDecay;
}
Here is the classical way, but you notice, it is not dynamic anymore because the device symbol must refer to the function reference not a pointer which may chnage during run-time..
// .. and this would work
#ifdef __CUDACC__
__host__ __device__
#endif
inline float fcn_rad_decay (float sigma0, float T, float lambda) {
return std::floor(sigma0*exp(-T/lambda) + 0.5f);
}
__device__ pDistanceFu pFoo= fcn_rad_decay; // pointer must target a reference, no host pointer possible
void DeviceAssign2(DistFunction &dist) {
cudaMemcpyFromSymbol(&hLRateDecay, &fcn_rad_decay, sizeof(pDecayFu) );
// the same:
// cudaMemcpyFromSymbol(&hLRateDecay, pFoo, sizeof(pDecayFu) );
// ..
dist.lrate_decay = hLRateDecay;
// ..
}
I would like to create a list of function pointers dynamically on the CPU (with some sort of push_back() method called from main()) and copy it to a GPU __constant__ or __device__ array, without needing to resort to static __device__ function pointers. I believe this question is related to my problem; however, my goal is to create the __host__ function pointer array iteratively and then copy it to the __constant__ function pointer array instead of initialising the latter on declaration.
A working code example with static function pointers (as seen here or here) would be:
common.h:
#ifndef COMMON_H
#define COMMON_H
#include <stdio.h>
#include <iostream>
#define num_functions 3
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const 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);
}
}
// fptr_t: Pointer to void function that takes two integer lvalues
typedef void (*fptr_t)(int&, int&);
// some examples of void(int&, int&) functions...
__device__ void Add(int &a, int &b) {printf("Add... %i + %i = %i\n", a, b, a+b);}
__device__ void Subtract(int &a, int &b) {printf("Subtract... %i - %i = %i\n", a, b, a-b);}
__device__ void Multiply(int &a, int &b) {printf("Multiply... %i * %i = %i\n", a, b, a*b);}
// List of function pointers in device memory
__constant__ fptr_t constant_fList[num_functions];
// Kernel called from main(): choose the function to apply whose index is equal to thread ID
__global__ void kernel(int a, int b) {
fptr_t f;
if (threadIdx.x < num_functions) {
f = constant_fList[threadIdx.x];
f(a,b);
}
}
#endif
main.cu:
#include "common.h"
// Static device function pointers
__device__ fptr_t p_Add = Add;
__device__ fptr_t p_Sub = Subtract;
__device__ fptr_t p_Mul = Multiply;
// Load function list to constant memory
void loadList_staticpointers() {
fptr_t h_fList[num_functions];
gpuErrchk( cudaMemcpyFromSymbol(&h_fList[0], p_Add, sizeof(fptr_t)) );
gpuErrchk( cudaMemcpyFromSymbol(&h_fList[1], p_Sub, sizeof(fptr_t)) );
gpuErrchk( cudaMemcpyFromSymbol(&h_fList[2], p_Mul, sizeof(fptr_t)) );
gpuErrchk( cudaMemcpyToSymbol(constant_fList, h_fList, num_functions * sizeof(fptr_t)) );
}
int main() {
loadList_staticpointers();
int a = 12, b = 15;
kernel<<<1,3>>>(a, b);
gpuErrchk(cudaGetLastError());
gpuErrchk(cudaDeviceSynchronize());
return 0;
}
Specs: GeForce GTX 670, compiled for -arch=sm_30, CUDA 6.5, Ubuntu 14.04
I wish to avoid the use of static device function pointers, as appending each function would require code maintenance on the user side - declaration of a new static pointer like p_Add or p_Mul, manipulation of void loadList_functionpointers(), etc. To make it clear, I am trying something like the following (crashing) code:
main_wrong.cu:
#include "common.h"
#include <vector>
// Global variable: list of function pointers in host memory
std::vector<fptr_t> vec_fList;
// Add function to functions list
void addFunc(fptr_t f) {vec_fList.push_back(f);}
// Upload the functions in the std::vector<fptr_t> to GPU memory
// Copies CPU-side pointers to constant_fList, therefore crashes on kernel call
void UploadVector() {
fptr_t* h_vpointer = vec_fList.data();
gpuErrchk( cudaMemcpyToSymbol(constant_fList, h_vpointer, vec_fList.size() * sizeof(fptr_t)) );
}
int main() {
addFunc(Add);
addFunc(Subtract);
addFunc(Multiply);
int a = 12, b = 15;
UploadVector();
kernel<<<1,3>>>(a, b); // Wrong to call a host-side function pointer from a kernel
gpuErrchk(cudaGetLastError());
gpuErrchk(cudaDeviceSynchronize());
return 0;
}
My understanding is that function pointers pointing to host addresses are copied to the GPU and are unusable by the kernel, which needs pointers pointing to GPU addresses when the function f(a,b) is called. Populating a host-side array with device-side pointers would work for me with raw data (see this question) but not with function pointers. Trivial attempts with Unified Memory have failed as well... so far, I have only found static device-side pointers to work. Is there no other way to copy a dynamically created CPU array of function pointers onto the GPU?
If you can use C++11 (supported since CUDA 7), you could use the following to auto-generate the function table:
template <fptr_t... Functions>
__global__ void kernel(int a, int b)
{
constexpr auto num_f = sizeof...(Functions);
constexpr fptr_t table[] = { Functions... };
if (threadIdx.x < num_f)
{
fptr_t f = table[threadIdx.x];
f(a,b);
}
}
You would then call this kernel using
kernel<Add, Subtract, Multiply><<<1,3>>>(a, b);
Inspired by m.s.'s answer, I chose to pass the function pointer as a template parameter -this was in fact the key to solve my problem- and discovered that filling a __device__ array of function pointers dev_fList from the main() function iteratively without the help of static function pointers is indeed possible, plus C++11 compatibility is not even needed!
Here is a working example on a __device__ array in global memory. I have not tried its constant memory counterpart yet, but once a global memory array has been satisfactorily created, my guess is that a cudaMemcpyToSymbol(..., cudaMemcpyDeviceToDevice) should do the trick.
A kernel kernel() creates a GPU address for function pointer dev_f and copies the function f that was passed as a template argument. Since this is an iterative process from the CPU, only one thread (thread 0) is involved in this kernel, which is launched with configuration <<<1,1>>>. The static variable count_f takes care of indexing in dev_fList.
common.h:
#ifndef COMMON_H
#define COMMON_H
#include <stdio.h>
#include <iostream>
#define num_functions 3
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const 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);
}
}
// fptr_t: Pointer to void function that takes two integer lvalues
typedef void (*fptr_t)(int&, int&);
// some examples of void(int&, int&) functions...
__device__ void Add(int &a, int &b) {printf("Add... %i + %i = %i\n", a, b, a+b);}
__device__ void Subtract(int &a, int &b) {printf("Subtract... %i - %i = %i\n", a, b, a-b);}
__device__ void Multiply(int &a, int &b) {printf("Multiply... %i * %i = %i\n", a, b, a*b);}
// List of function pointers in device memory
// Note that, in my example, it resides in global memory space, not constant memory
__device__ fptr_t dev_fList[num_functions];
#endif
main.cu:
#include "common.h"
// Index in dev_fList[] == number of times addFunc<>() was launched
static int count_f = 0;
// Kernel that copies function f to the GPU
template<fptr_t f>
__global__ void kernel(int a, int b, int idx) {
fptr_t dev_f = f; // Create device function pointer
dev_fList[idx] = dev_f; // Populate the GPU array of function pointers
dev_fList[idx](a,b); // Make sure that the array was populated correctly
}
// Add function to functions list
template<fptr_t f>
void addFunc(const int &a, const int &b) {
if (count_f >= num_functions) {
std::cout << "Error: not enough memory statically allocated on device!\n";
exit(EXIT_FAILURE);
}
kernel<f><<<1,1>>>(a,b,count_f);
gpuErrchk(cudaGetLastError());
gpuErrchk(cudaDeviceSynchronize());
count_f++;
}
int main() {
int a = 12, b = 15;
addFunc<Add>(a,b);
addFunc<Subtract>(a,b);
addFunc<Multiply>(a,b);
return 0;
}
Edit: Added copy of the array of function pointers to constant memory
For what it's worth, here is how to copy our dev_fList array to constant memory:
In common.h:
__constant__ fptr_t cst_fList[num_functions];
__global__ void cst_test(int a, int b, int idx) {
if (threadIdx.x < idx) cst_fList[threadIdx.x](a,b);
}
In main.cu main() function, after all desired functions have been added:
fptr_t *temp;
gpuErrchk( cudaMemcpyFromSymbol((void**)&temp, dev_fList[0], count_f * sizeof(fptr_t)) );
gpuErrchk( cudaMemcpyToSymbol(cst_fList[0], &temp, count_f * sizeof(fptr_t)) );
cst_test<<<1,count_f>>>(a,b, count_f);
gpuErrchk(cudaGetLastError());
gpuErrchk(cudaDeviceSynchronize());
It may look ugly as I understand that memory is transferred to the host via temp and then back to the device; more elegant suggestions are welcome.
It is impossible to use dynamically created CUDA device function pointers (at least not without crash or UB). The template based solutions work at compile time (not dynamic). The CUDA device function pointer approaches you see everywhere need device symbols in global space. This means that for every function a device function pointer must be already declared. This also means you cannot use normal C function pointers as reference, which are e.g. set at runtime. In comprehension, using CUDA device function pointers is questionable. Template based approaches look user-friendly, but are per definition not dynamic.
Example showing structure with function pointers:
This example shows a structure having some function pointers. In normal C++ code, you can set and change the device function pointers while the program is running (dynamically). With CUDA this example below is impossible, because the function pointers in the struct are no valid device symbols. This means they cannot be used with "cudaMemcpyFromSymbol". To circumvent this, either the original function (target of the function pointers) or global cuda device function pointers must be created. Both is not dynamic.
This is dynamic assignment:
typedef float (*pDistanceFu) (float, float);
typedef float (*pDecayFu) (float, float, float);
// In C++ you can set and reset the function pointer during run time whenever you want ..
struct DistFunction {
/*__host__ __device__*/ pDistanceFu distance; // uncomment for NVCC ..
/*__host__ __device__*/ pDecayFu rad_decay;
/*__host__ __device__*/ pDecayFu lrate_decay;
};
// you can do what you want ..
DistFunction foo, bar;
foo.distance = bar.distance;
// ..
This is how it should be with CUDA, but it will fail, because there is no valid device symbol :(
pDistanceFu hDistance;
pDecayFu hRadDay;
pDecayFu hLRateDecay;
void DeviceAssign(DistFunction &dist) {
cudaMemcpyFromSymbol(&hDistance, dist.distance, sizeof(pDistanceFu) );
cudaMemcpyFromSymbol(&hRadDay, dist.rad_decay, sizeof(pDecayFu) );
cudaMemcpyFromSymbol(&hLRateDecay, dist.lrate_decay, sizeof(pDecayFu) );
dist.distance = hDistance;
dist.rad_decay = hRadDay;
dist.lrate_decay = hLRateDecay;
}
Here is the classical way, but you notice, it is not dynamic anymore because the device symbol must refer to the function reference not a pointer which may chnage during run-time..
// .. and this would work
#ifdef __CUDACC__
__host__ __device__
#endif
inline float fcn_rad_decay (float sigma0, float T, float lambda) {
return std::floor(sigma0*exp(-T/lambda) + 0.5f);
}
__device__ pDistanceFu pFoo= fcn_rad_decay; // pointer must target a reference, no host pointer possible
void DeviceAssign2(DistFunction &dist) {
cudaMemcpyFromSymbol(&hLRateDecay, &fcn_rad_decay, sizeof(pDecayFu) );
// the same:
// cudaMemcpyFromSymbol(&hLRateDecay, pFoo, sizeof(pDecayFu) );
// ..
dist.lrate_decay = hLRateDecay;
// ..
}
I don't understand why am I getting the error dynamic initialization is not supported for __device__, __constant__, __shared__ variables when compiling my code.
My code looks like
wrapper.cu
#include "../Params.hpp"
__constant__ Params cparams;
void wrapperFunction(uint& a)
{
Params ab;
a = 20;
}
Params.hpp
#include "Utils.hpp"
typedef struct Params
{
vectorTypef a;
} Params;
Utils.hpp
#include "Vec2.hpp"
typedef unsigned int uint;
typedef Vec2<float> vectorTypef;
Vec2.hpp
template <typename T>
class Vec2
{
public:
Vec2(){ x = 0.0; y = 0.0;}
T x, y;
};
Building with cmake with the command
CUDA_ADD_EXECUTABLE(test main.cpp cudasrc/wrapper.cu
Your Params struct is used in the __constant__ memory definition of cparams.
Your Params struct contains an element a of type vectorTypef which is a typedef for the Vec2 class for float. This class has a default constructor, that is assigning elements ultimately of the Params struct. This method of assigning data to a __constant__ region is not legal either in device code or host code.
In device code it's not legal to modify a __constant__ value at all. In host code (which is what is in view here), __constant__ values should be assigned using the appropriate API, i.e. cudaMemcpyToSymbol. I would recommend that you assign these in your host code explicitly, rather than via a constructor.
So, one possible approach to fix this would be to change your default constructor to an empty one:
public:
__host__ __device__ Vec2(){ }; // change this line
T x, y;
(you could also just delete the empty default constructor line)
And, in wrapper.cu (perhaps in wrapperFunction), initialize your Params __constant__ struct:
Params hparams;
hparams.a.x = 0.0;
hparams.a.y = 0.0;
cudaMemcpyToSymbol(cparams, &hparams, sizeof(Params));
I got the same problems as you and I found two ways to solve it.
define your struct in C-type, like this:
typedef struct {} ClassName;
define both constructor and destructor as __device__ type, like this:
struct ClassName{
public:
__device__ ClassName(){...}
__device__ ~ClassName(){...}
};
I have found some strange runtime behaviour while experimenting with function pointers in CUDA.
Goal
My goal is to make my function pointers choose which function to apply to two objects according to an internal property of the latter.
In short, I want to emulate C++ templates with a CUDA kernel - without actually using template arguments or switch clauses, but function pointers and class/struct members instead.
Approach
Define my custom objects struct customObj with one property (int type) that will emulate the arguments of a template.
Define a bunch of dummy functions (Sum(), Subtract(), etc) to choose from.
Keep the list of functions to apply (functionsList) and respective type members to look up (first_types, second_types) in __constant__ memory, such that function functionsList[i](obj1,obj2) is applied to objects with obj1.type == first_types[i] and obj2.type == second_types[i].
Working code
The following code has been compiled for Linux x86_64 with CUDA 5.0, on a GPU with compute capability 3.0 (GeForce GTX 670), and works.
#include <stdio.h>
#include <iostream>
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
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);
}
}
struct customObj
{
int type;
double d;
// Constructors
__device__ __host__ customObj() {}
__device__ __host__ customObj(const int& _type, const double& _d) : type(_type), d(_d) {}
};
typedef void (*function_t)(customObj&, customObj&);
// Define a bunch of functions
__host__ __device__ void Sum(customObj& obj1, customObj& obj2) {printf("Sum chosen! d1 + d2 = %f\n", obj1.d + obj2.d);}
__host__ __device__ void Subtract(customObj& obj1, customObj& obj2) {printf("Subtract chosen! d1 - d2 = %f\n", obj1.d - obj2.d);}
__host__ __device__ void Multiply(customObj& obj1, customObj& obj2) {printf("Multiply chosen! d1 * d2 = %f\n", obj1.d * obj2.d);}
#define ARRAYLENGTH 3
__constant__ int first_type[ARRAYLENGTH] = {1, 2, 3};
__constant__ int second_type[ARRAYLENGTH] = {1, 1, 2};
__constant__ function_t functionsList[ARRAYLENGTH] = {Sum, Sum, Subtract};
// Kernel to loop through functions list
__global__ void choosefunction(customObj obj1, customObj obj2) {
int i = 0;
function_t f = NULL;
do {
if ((obj1.type == first_type[i]) && (obj2.type == second_type[i])) {
f = functionsList[i];
break;
}
i++;
} while (i < ARRAYLENGTH);
if (f == NULL) printf("No possible interaction!\n");
else f(obj1,obj2);
}
int main() {
customObj obj1(1, 5.2), obj2(1, 2.6);
choosefunction<<<1,1>>>(obj1, obj2);
gpuErrchk(cudaPeekAtLastError());
gpuErrchk(cudaDeviceSynchronize());
return 0;
}
The issue
The problem that I have found is that, as soon as I replace the datatype of member int type and related variables and functions (__constant__ int first_types[...] and so on)... the code compiles but stops working!
If I change the datatype from int to char or int8_t, the memory checker throws error 4 on my call to cudaDeviceSynchronize().
If I change the datatype to unsigned short int, I get a hardware stack overflow.
So, is anybody having similar issues when working with __constant__ memory? I really have no clue about what is going on. As far as I know, char and int8_t are built-in types of 1 byte length, while the size of int is 4 bytes, so maybe it is about data alignment, but I'm just guessing here. Besides, CUDA is supposed to support function pointers on the GPU since compute capability 2.0. Are there any special constraints for function pointers in __constant__ memory that I'm missing?
I was able to reproduce the problem (error 4, unspecified launch failure) on CUDA 5.0 on 64bit RHEL 5.5, but not on CUDA 6.0.
Please update/upgrade to CUDA 6.