Skip to content

cuda: Fix up base addrs on all gpu channels to translate host address… #447

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
muraj wants to merge 1 commit into
base: main
Choose a base branch
from
Open

cuda: Fix up base addrs on all gpu channels to translate host address… #447

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
76 changes: 68 additions & 8 deletions src/realm/cuda/cuda_internal.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -737,8 +737,21 @@ namespace Realm {
stream = select_stream(out_gpu, in_gpu, channel->get_gpu(), out_mapping,
in_port->mem, out_port->mem);
assert(stream != NULL);

AutoGPUContext agc(stream->get_gpu());

if(!stream->get_gpu()->info->host_gpu_same_va) {
CUdeviceptr in_device = static_cast<CUdeviceptr>(in_base);
CUdeviceptr out_device = static_cast<CUdeviceptr>(out_base);
if(CUDA_SUCCESS == CUDA_DRIVER_FNPTR(cuMemHostGetDevicePointer)(
&in_device, reinterpret_cast<void *>(in_base), 0)) {
in_base = static_cast<uintptr_t>(in_device);
}
if(CUDA_SUCCESS == CUDA_DRIVER_FNPTR(cuMemHostGetDevicePointer)(
&out_device, reinterpret_cast<void *>(out_base), 0)) {
out_base = static_cast<uintptr_t>(out_device);
}
}

size_t copy_info_total = 0;
size_t min_align = 16; // Hope for the highest type alignment we can get, 16 bytes
copy_infos.num_rects = 0;
Expand Down Expand Up @@ -1119,6 +1132,22 @@ namespace Realm {
}
}

auto stream = select_stream(out_gpu, in_gpu, channel->get_gpu(), out_mapping,
in_port->mem, out_port->mem);
AutoGPUContext agc(stream->get_gpu());
if(!stream->get_gpu()->info->host_gpu_same_va) {
CUdeviceptr in_device = static_cast<CUdeviceptr>(in_base);
CUdeviceptr out_device = static_cast<CUdeviceptr>(out_base);
if(CUDA_SUCCESS == CUDA_DRIVER_FNPTR(cuMemHostGetDevicePointer)(
&in_device, reinterpret_cast<void *>(in_base), 0)) {
in_base = static_cast<uintptr_t>(in_device);
}
if(CUDA_SUCCESS == CUDA_DRIVER_FNPTR(cuMemHostGetDevicePointer)(
&out_device, reinterpret_cast<void *>(out_base), 0)) {
out_base = static_cast<uintptr_t>(out_device);
}
}

TransferIterator::AddressInfo addr_info;
memset(&addr_info, 0, sizeof(TransferIterator::AddressInfo));

Expand Down Expand Up @@ -1187,10 +1216,6 @@ namespace Realm {
else
write_ind_bytes = (max_bytes / field_size) * addr_size;

auto stream = select_stream(out_gpu, in_gpu, channel->get_gpu(), out_mapping,
in_port->mem, out_port->mem);
AutoGPUContext agc(stream->get_gpu());

// We can't do gather-scatter yet.
assert(!(out_port->indirect_port_idx >= 0 && in_port->indirect_port_idx >= 0));
assert(!in_nonaffine && !out_nonaffine);
Expand Down Expand Up @@ -2023,6 +2048,14 @@ namespace Realm {

AutoGPUContext agc(channel->gpu);
GPUStream *stream = channel->gpu->get_next_d2d_stream();
if(!stream->get_gpu()->info->host_gpu_same_va) {
CUdeviceptr out_device = static_cast<CUdeviceptr>(out_base);
if(CUDA_SUCCESS == CUDA_DRIVER_FNPTR(cuMemHostGetDevicePointer)(
&out_device, reinterpret_cast<void *>(out_base), 0)) {
out_base = static_cast<uintptr_t>(out_device);
}
}

Realm::Cuda::AffineFillInfo<2, size_t> fill_info{};
size_t total_info_bytes = 0;
for(size_t offset = 0;
Expand Down Expand Up @@ -2573,7 +2606,6 @@ namespace Realm {
const size_t out_span_start, const size_t in_elem_size,
const size_t out_elem_size, const size_t elems, const bool has_transpose)
{
AutoGPUContext agc(channel->gpu);
size_t threads_per_block = 256;
size_t blocks_per_grid = 1 + ((elems - 1) / threads_per_block);
const void *host_proxy = nullptr;
Expand Down Expand Up @@ -2686,6 +2718,7 @@ namespace Realm {
};
KernelArgs *args = 0; // allocate on demand
size_t args_size = sizeof(KernelArgs) + redop->sizeof_userdata;
AutoGPUContext agc(channel->gpu);
while(true) {
size_t min_xfer_size = 4096; // TODO: make controllable
const InstanceLayoutPieceBase *in_nonaffine = nullptr;
Expand Down Expand Up @@ -2772,6 +2805,21 @@ namespace Realm {
in_base = reinterpret_cast<uintptr_t>(in_port->mem->get_direct_ptr(0, 0));
}

if(!channel->gpu->info->host_gpu_same_va) {
CUdeviceptr in_device = static_cast<CUdeviceptr>(in_base);
CUdeviceptr out_device = static_cast<CUdeviceptr>(out_base);
if(CUDA_SUCCESS ==
CUDA_DRIVER_FNPTR(cuMemHostGetDevicePointer)(
&in_device, reinterpret_cast<void *>(in_base), 0)) {
in_base = static_cast<uintptr_t>(in_device);
}
if(CUDA_SUCCESS ==
CUDA_DRIVER_FNPTR(cuMemHostGetDevicePointer)(
&out_device, reinterpret_cast<void *>(out_base), 0)) {
out_base = static_cast<uintptr_t>(out_device);
}
}

while(total_elems < max_elems) {
AddressListCursor &in_alc = in_port->addrcursor;
AddressListCursor &out_alc = out_port->addrcursor;
Expand Down Expand Up @@ -2831,8 +2879,6 @@ namespace Realm {
static_cast<size_t>(CUDA_MAX_BLOCKS_PER_GRID));

{
AutoGPUContext agc(channel->gpu);

if(kernel != 0) {
// Use params array to pass kernel arguments (pointers to each
// parameter) instead of CU_LAUNCH_PARAM_BUFFER_POINTER (packed
Expand Down Expand Up @@ -3191,6 +3237,20 @@ namespace Realm {
}
uintptr_t out_base =
reinterpret_cast<uintptr_t>(out_port->mem->get_direct_ptr(0, 0));

if(!stream->get_gpu()->info->host_gpu_same_va) {
CUdeviceptr in_device = static_cast<CUdeviceptr>(in_base);
CUdeviceptr out_device = static_cast<CUdeviceptr>(out_base);
if(CUDA_SUCCESS == CUDA_DRIVER_FNPTR(cuMemHostGetDevicePointer)(
&in_device, reinterpret_cast<void *>(in_base), 0)) {
in_base = static_cast<uintptr_t>(in_device);
}
if(CUDA_SUCCESS == CUDA_DRIVER_FNPTR(cuMemHostGetDevicePointer)(
&out_device, reinterpret_cast<void *>(out_base), 0)) {
out_base = static_cast<uintptr_t>(out_device);
}
}

size_t bytes_left = max_bytes;
bool has_transpose = false;
size_t bytes_to_copy = 0;
Expand Down
Loading