I am trying to allocate several SVM buffers and pass them to an OpenCL kernel using the following method. The kernel is run on Intel HD Graphics 530 and NVIDIA GTX 950M. I get different results on these GPUs, and I am not sure which behavior is correct (maybe, both?)

0. Initialize OpenCL.
1. Allocate 3 buffers (data0, data1, pointers) with 3 clSVMAlloc() calls.
2. Check the results for NULL (no failure here).
3. Map all the buffers using blocking clEnqueueSVMMap() calls check for CL_SUCCESS (no failure here).
4. Fill data0 and data1 with data, and pointers with pointers to data0 and data1, so that pointers[0] == data0 and pointers[1] == data1.
5. Unmap all the buffers using clEnqueueSVMUnmap() check for CL_SUCCESS (no failure here).
6. Set the 1st kernel argument using clSetKernelArgSVMPointer(..., data0) check for CL_SUCCESS (no failure here).
7. Set the 2nd kernel argument using clSetKernelArgSVMPointer(..., pointers) check for CL_SUCCESS (no failure here).
8. Run the kernel and test results.

On NVIDIA card, the data is available through both pointers[0] and pointers[1]. This is the expected behavior.

On Intel chip, memory behind pointers[0] is available (because data0 has been set as kernel argument), but memory behind pointers[1] is all zeroes. If I set data1 as kernel argument, then it becomes available through pointers[1] and memory behind pointers[0] becomes zeroes.

The question is: is Intel approach a bug or a feature? I have not found any related information in the OpenCL 3.0 specification.

CodePudding user response：

Intel works correctly and adheres to the specification. NVIDIA is more fool-proof, however.

To fix the problem in question, one needs to not forget to pass an array of SVM buffers to a kernel:

ret = clSetKernelExecInfo(kernel, CL_KERNEL_EXEC_INFO_SVM_PTRS, sizeof(svm_buffers), svm_buffers);

where svm_buffers is void*[] array of pointers to buffers allocated using clSVMAlloc() calls.

The OpenCL specification states the following for clSetKernelExecInfo with parameter CL_KERNEL_EXEC_INFO_SVM_PTRS:

SVM pointers must reference locations contained entirely within buffers that are passed to kernel as arguments, or that are passed through the execution information. Non-argument SVM buffers must be specified by passing pointers to those buffers via clSetKernelExecInfo for coarse-grain and fine- grain buffer SVM allocations but not for finegrain system SVM allocations.