Commit 519d3ac7 authored by Cianciosa, Mark's avatar Cianciosa, Mark
Browse files

Fix cuda bugs after update to managed buffers.

parent 65acc3c4

graph_framework/cuda_context.hpp

+2 −6
Original line number Diff line number Diff line
@@ -103,9 +103,6 @@ namespace gpu { 
            for (CUdeviceptr &ptr : buffers) {

                check_error(cuMemFree(ptr), "cuMemFree");

            }

            for (CUdeviceptr &ptr : result_buffers) {

                check_error(cuMemFree(ptr), "cuMemFree");

            }



            check_error(cuStreamDestroy(stream), "cuStreamDestroy");

            check_error(cuDevicePrimaryCtxRelease(device), "cuDevicePrimaryCtxRelease");
@@ -194,7 +191,6 @@ namespace gpu { 
            free(ptx);



            buffers.resize(inputs.size() + outputs.size());

            result_buffers.resize(inputs.size() + outputs.size());



            buffer_element_size = sizeof(typename BACKEND::base);

            buffer_offset = ray_index;
@@ -203,14 +199,14 @@ namespace gpu { 
            for (size_t i = 0, ie = inputs.size(); i < ie; i++) {

                const BACKEND backend = inputs[i]->evaluate();



                check_error(cuMemAlloc(&buffers[i], backend.size()*buffer_element_size), "cuMemAlloc");

                check_error(cuMemAllocManaged(&buffers[i], backend.size()*buffer_element_size, CU_MEM_ATTACH_GLOBAL), "cuMemAllocManaged");

                check_error(cuMemcpyHtoD(buffers[i], &backend[0], backend.size()*buffer_element_size), "cuMemcpyHtoD");

                kernel_arguments.push_back(reinterpret_cast<void *> (&buffers[i]));

            }

            for (size_t i = inputs.size(), ie = buffers.size(), j = 0; i < ie; i++, j++)	{

                const BACKEND backend = outputs[j]->evaluate();



                check_error(cuMemAlloc(&buffers[i], backend.size()*buffer_element_size), "cuMemAlloc");

                check_error(cuMemAllocManaged(&buffers[i], backend.size()*buffer_element_size, CU_MEM_ATTACH_GLOBAL), "cuMemAllocManaged");

                check_error(cuMemcpyHtoD(buffers[i], &backend[0], backend.size()*buffer_element_size), "cuMemcpyHtoD");

                kernel_arguments.push_back(reinterpret_cast<void *> (&buffers[i]));

            }