为什么我的二维数组复制参数被驱动程序 API 拒绝？

Question

I'm trying to use the CUDA Driver API to copy data into a 2D array, in the program listed below, but am getting an "invalid value" error when I pass my copy parameters. What value in them is wrong?

#include <cuda.h>

#include <iostream>
#include <iomanip>
#include <numeric>
#include <limits>
#include <cstring>

[[noreturn]] void die_(const std::string& message) {
    std::cerr << message << "\n";
    exit(EXIT_FAILURE);
}

void die_if_error(CUresult status, const std::string& extra_message) {
    if (status != CUDA_SUCCESS) {
        const char* error_string;
        cuGetErrorString(status, &error_string);
        die_(extra_message + ": " + error_string);
    }
}

template <typename T = void>
T* as_pointer(CUdeviceptr address) noexcept { return reinterpret_cast<T*>(address); }

CUdeviceptr as_address(void* ptr) noexcept { return reinterpret_cast<CUdeviceptr>(ptr); }

int main() {
    CUresult status;
    int device_id = 0;
    status = cuInit(0);
    die_if_error(status, "Initializing the CUDA driver");
    CUcontext pctx;
    status = cuDevicePrimaryCtxRetain(&pctx, device_id);
    die_if_error(status, "Obtaining the primary device context");
    cuCtxSetCurrent(pctx);
    struct { unsigned width, height; } dims = { 3, 3 };
    std::cout << "Creating a " << dims.width << " x " << dims.height << " CUDA array" << std::endl;
    CUarray arr_handle;
    {
        CUDA_ARRAY_DESCRIPTOR array_descriptor;
        array_descriptor.Width = dims.width;
        array_descriptor.Height = dims.height;
        array_descriptor.Format = CU_AD_FORMAT_FLOAT;
        array_descriptor.NumChannels = 1;
        status = cuArrayCreate(&arr_handle, &array_descriptor);
        die_if_error(status, "Failed creating a 2D CUDA array");
    }
    auto arr_size = dims.width * dims.height;
    CUdeviceptr dptr;
    status = cuMemAllocManaged(&dptr, arr_size, CU_MEM_ATTACH_GLOBAL);
    die_if_error(status, "Failed allocating managed memory");
    float* ptr_in = as_pointer<float>(dptr);
    std::iota(ptr_in, ptr_in + arr_size, 0);
    CUmemorytype ptr_in_memory_type;
    status = cuPointerGetAttribute(&ptr_in_memory_type, CU_POINTER_ATTRIBUTE_MEMORY_TYPE, as_address(ptr_in));
    if (not (ptr_in_memory_type == CU_MEMORYTYPE_UNIFIED or ptr_in_memory_type == CU_MEMORYTYPE_DEVICE)) {
        die_("Unexpected memory type for ptr_in");
    }
    std::cout << "The memory type of ptr_in is " << (ptr_in_memory_type == CU_MEMORYTYPE_DEVICE ? "DEVICE" : "UNIFIED") << std::endl;
    std::cout << "Will copy from ptr_in into a 2D CUDA array" << std::endl;

    CUDA_MEMCPY2D cp;
    {
        // Source

        cp.srcXInBytes = 0; cp.srcY = 0; // No offset
        cp.srcMemoryType = ptr_in_memory_type;
        cp.srcDevice = as_address(ptr_in);
        // no extra source pitch
        cp.srcPitch = dims.width * sizeof(float);

        // Destination

        cp.dstXInBytes = 0; cp.dstY = 0; // No destination offset
        cp.dstMemoryType = CU_MEMORYTYPE_ARRAY;
        cp.dstArray = arr_handle;

        cp.WidthInBytes = dims.width * sizeof(float);
        cp.Height = dims.height;
    }
    status = cuMemcpy2D(&cp);
    die_if_error(status, "cuMemcpy2D failed");
    cuMemFree(as_address(ptr_in));
}

Full output of this program:

Creating a 3 x 3 CUDA array
The memory type of ptr_in is DEVICE
Will copy from ptr_in into a 2D CUDA array
cuMemcpy2D failed: invalid argument

Additional information:

• CUDA toolkit version: 11.4
• NVIDIA driver version: 470.57.02
• OS distribution: Devuan Chimaera GNU/Linux
• GPU: GeForce 1050 TI Boost (Compute Capability 6.1)
• Host architecture: amd64

Answer 1

The error is here:

auto arr_size = dims.width * dims.height;
CUdeviceptr dptr;
status = cuMemAllocManaged(&dptr, arr_size, CU_MEM_ATTACH_GLOBAL);
                                  ^^^^^^^^

That should be arr_size*sizeof(float)

cuMemAllocManaged() , like malloc() takes a size argument in bytes. This size needs to be consistent with (greater than or equal to) your implied size of transfer in the cuMemcpy2D call.

Answer 2

tl;dr: "invalid value" can be a pointer without sufficient allocated memory

_{(@RobertCrovella noticed the error, but I want to emphasize a point:)}

We are used to APIs not being able to scrutinize pointers too much, accepting them on faith, then possibly failing with invalid access errors (segmentation fault on the host side, invalid memory access on the device side etc.)

However, CUDA (in particular, the CUDA driver) scrutinizes pointers more. You already know this to be the case, seeing how it can tell you what memory type a pointer points to.

Well, it seems cuMemCpy2D() also checks the amount of memory allocated at ptr_in - and figures out that it's not enough to suffice for filling the area, ie it would copy from unallocated memory. That's why it returns the "invalid value" error. So the error code is valid, albeit rather vague.

Specifically, and as @RobertCrovella points out, you did not allocate enough memory for 3x3 floats - your arr_size is in elements, ie 9, while you need to allocate 9 floats, ie 36 bytes. You lucked out writing to it, probably because of CUDA's memory allocation quantum, or memory page granularity etc.