René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"array.device is sometimes returning raw pointers","articleBody":"I am mixing CUDA and arrayfire code extensively. I am using CUDA surface objects, so I need to do a copy from the arrayfire device memory to the cuda array memory. Currently, I'm using cudaMemcpy3D(), and I am assuming that the arrayfire data is in packed form, in H x W x C order (height varies first.) I'm using float32 elements here.\n\nHere's the code I use:\n\n```\n// release any previous content, make a cuda array, copy from arrayfire array\nvoid CudaArray::to_surface(const af::array \u0026in) \n{\n    if (in.dims()[3] != 1)\n        throw \"CudaArray: attempt to allocate a 4D CUDA surface\";\n\n    // free up old CUDA array and allocate new one\n    free();\n    dims_ = in.dims();\n    allocate();\n\n    void *d_srcptr = in.device\u003cfloat\u003e();    // does an implicit eval()\n    af::sync();     // ensure all arrayfire work is complete before we trigger CUDA code\n\n    cudaMemcpy3DParms params = { 0 };\n    params.srcPtr = make_cudaPitchedPtr(d_srcptr, dims_[0] * sizeof(float), dims_[0], dims_[1]);    // dims[0] is what varies fastest\n    params.dstArray = cuda_array;\n    params.kind = cudaMemcpyDeviceToDevice;\n    params.extent = make_cudaExtent(dims_[0], dims_[1], dims_[2]);\n\n    checkCuda(cudaMemcpy3D(\u0026params));\n    in.unlock();                        // done with d_srcptr\n}\n```\n\nI use the following to get the data out of the surface back to an array:\n\n```\n// make a new arrayfire array, copy data from surface to the array\naf::array CudaArray::to_array() \n{\n    // ensure CUDA is done before we start copying\n    checkCuda(cudaDeviceSynchronize());\n    checkCuda(cudaGetLastError());\n\n    af::array out(dims_);\n\n    void *d_dstptr = out.device\u003cfloat\u003e();\n\n    cudaMemcpy3DParms params = { 0 };\n    params.srcArray = cuda_array;\n    params.dstPtr = make_cudaPitchedPtr(d_dstptr, dims_[0] * sizeof(float), dims_[0], dims_[1]);\n    params.kind = cudaMemcpyDeviceToDevice;\n    params.extent = make_cudaExtent(dims_[0], dims_[1], dims_[2]);\n\n    checkCuda(cudaMemcpy3D(\u0026params));\n    out.unlock();                       // done with d_dstptr\n\n    return out;\n}\n```\n\nEverything works fine as long as to_surface() is invoked only on a fresh array, i.e., an array created by `array foo = loadImage(\"file.png\", true); CudaArray baz(foo);   // this works just fine`\n\nHowever, if I do any kind of arrayfire processing on the array, e.g.,\n\n```\narray subwindow(const array \u0026in)\n{\n    int h = in.dims()[0], w = in.dims()[1];\n    array out = in(seq(h/4, h/4+h/2-1), seq(w/4, w/4+w/2-1), span);\n    return out;\n}\n\nint main()\n{\n    af::setDevice(0); .. make a window ..\n\n    array in = loadImage(\"baz.jpg\", true);\n    in = subwindow(in);          // remove this line and the code works\n    CudaArray baz(in);\n    win.image(in.to_array());  // shows what appears to be result of a bad stride \n}\n```\n\nSo the question is, after the call to subwindow, is the memory layout of the array in still tightly packed with floats with no padding and a pitch in bytes equal to the height \\* 4?\n","author":{"url":"https://github.com/AEHousman","@type":"Person","name":"AEHousman"},"datePublished":"2016-06-06T00:07:41.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":11},"url":"https://github.com/1450/arrayfire/issues/1450"}