如何像C ++ const / constexpr一样定义CUDA设备常量？

Question

In a .cu file I've tried the following in the global scope (ie not in a function):

__device__ static const double cdInf = HUGE_VAL / 4;

And got nvcc error:

error : dynamic initialization is not supported for __device__, __constant__ and __shared__ variables.

How to define a C++ const/constexpr on the device, if that's possible?

NOTE1: #define is out of question not only for aesthetic reasons, but also because in practice the expression is more complex and involves an internal data type, not just double. So calling the constructor each time in each CUDA thread would be too expensive.

NOTE2: I doubt the performance of __constant__ because it's not a compile-time constant, but rather like a variable written with cudaMemcpyToSymbol .

Answer 1

Use a constexpr __device__ function:

#include <stdio.h>
__device__ constexpr double cdInf() { return HUGE_VAL / 4; }
__global__ void print_cdinf() { printf("in kernel, cdInf() is %lf\n", cdInf()); }
int main() { print_cdinf<<<1, 1>>>(); return 0; }

The PTX should be something like:

.visible .entry print_cdinf()(

)
{
        .reg .b64       %SP;
        .reg .b64       %SPL;
        .reg .b32       %r<2>;
        .reg .b64       %rd<7>;


        mov.u64         %rd6, __local_depot0;
        cvta.local.u64  %SP, %rd6;
        add.u64         %rd1, %SP, 0;
        cvta.to.local.u64       %rd2, %rd1;
        mov.u64         %rd3, 9218868437227405312;
        st.local.u64    [%rd2], %rd3;
        mov.u64         %rd4, $str;
        cvta.global.u64         %rd5, %rd4;
        // Callseq Start 0
        {
        .reg .b32 temp_param_reg;
        // <end>}
        .param .b64 param0;
        st.param.b64    [param0+0], %rd5;
        .param .b64 param1;
        st.param.b64    [param1+0], %rd1;
        .param .b32 retval0;
        call.uni (retval0), 
        vprintf, 
        (
        param0, 
        param1
        );
        ld.param.b32    %r1, [retval0+0];

        //{
        }// Callseq End 0
        ret;
}

With no code for the constexpr function. You could also use a constexpr __host__ function, but that's experimental in CUDA 7: use the nvcc command-line options seems to be --expt-relaxed-constexpr and see here for more details (thanks @harrism).

Answer 2

To make the code you have shown compile and work as expected, you need to initialize the variable at runtime, not compile time. To do this, add a host side call to cudaMemcpyToSymbol , something like:

__device__ double cdInf;

// ...

double val = HUGE_VAL / 4
cudaMemcpyToSymbol(cdInf, &val, sizeof(double));

However, for a single value, passing it as a kernel argument would seem far more sensible. The compiler will automagically store the argument in constant memory on all supported architectures, and there is a "free" constant cache broadcast mechanism which should make the cost of accessing the value at runtime negligible.

Answer 3

To initialize it you have to use cudaMemcpyToSymbol . It is not a compile time constant but stored in the constant memory of the device and has some advantages over global memory. From the CUDA blogspot:

For all threads of a half warp, reading from the constant cache is as fast as reading from a register as long as all threads read the same address. Accesses to different addresses by threads within a half warp are serialized, so cost scales linearly with the number of different addresses read by all threads within a half warp.

You do not need to use const , and you cannot use it. It is not a c++ constant since you need to modify it through cudaMemcpyToSymbol . So it is not a "real" constant at least from the c++ point of view. But it behaves like a constant inside the device kernels because you can modify it only through cudaMemcpyToSymbol which is callable only from host.