Skip to content

Commit 03d362d

Browse files
authored
[libc][Docs] Update the GPU RPC documentation (#79069)
Summary: This adds some more concrete information on the RPC interface. Hopefully this is intelligable and provides some useful examples.
1 parent d360963 commit 03d362d

File tree

3 files changed

+296
-7
lines changed

3 files changed

+296
-7
lines changed

libc/docs/gpu/rpc-diagram.svg

Lines changed: 1 addition & 0 deletions
Loading

libc/docs/gpu/rpc.rst

Lines changed: 292 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -11,10 +11,298 @@ Remote Procedure Calls
1111
Remote Procedure Call Implementation
1212
====================================
1313

14-
Certain features from the standard C library, such as allocation or printing,
15-
require support from the operating system. We instead implement a remote
16-
procedure call (RPC) interface to allow submitting work from the GPU to a host
17-
server that forwards it to the host system.
14+
Traditionally, the C library abstracts over several functions that interface
15+
with the platform's operating system through system calls. The GPU however does
16+
not provide an operating system that can handle target dependent operations.
17+
Instead, we implemented remote procedure calls to interface with the host's
18+
operating system while executing on a GPU.
19+
20+
We implemented remote procedure calls using unified virtual memory to create a
21+
shared communicate channel between the two processes. This memory is often
22+
pinned memory that can be accessed asynchronously and atomically by multiple
23+
processes simultaneously. This supports means that we can simply provide mutual
24+
exclusion on a shared better to swap work back and forth between the host system
25+
and the GPU. We can then use this to create a simple client-server protocol
26+
using this shared memory.
27+
28+
This work treats the GPU as a client and the host as a server. The client
29+
initiates a communication while the server listens for them. In order to
30+
communicate between the host and the device, we simply maintain a buffer of
31+
memory and two mailboxes. One mailbox is write-only while the other is
32+
read-only. This exposes three primitive operations: using the buffer, giving
33+
away ownership, and waiting for ownership. This is implemented as a half-duplex
34+
transmission channel between the two sides. We decided to assign ownership of
35+
the buffer to the client when the inbox and outbox bits are equal and to the
36+
server when they are not.
37+
38+
In order to make this transmission channel thread-safe, we abstract ownership of
39+
the given mailbox pair and buffer around a port, effectively acting as a lock
40+
and an index into the allocated buffer slice. The server and device have
41+
independent locks around the given port. In this scheme, the buffer can be used
42+
to communicate intent and data generically with the server. We them simply
43+
provide multiple copies of this protocol and expose them as multiple ports.
44+
45+
If this were simply a standard CPU system, this would be sufficient. However,
46+
GPUs have my unique architectural challenges. First, GPU threads execute in
47+
lock-step with each other in groups typically called warps or wavefronts. We
48+
need to target the smallest unit of independent parallelism, so the RPC
49+
interface needs to handle an entire group of threads at once. This is done by
50+
increasing the size of the buffer and adding a thread mask argument so the
51+
server knows which threads are active when it handles the communication. Second,
52+
GPUs generally have no forward progress guarantees. In order to guarantee we do
53+
not encounter deadlocks while executing it is required that the number of ports
54+
matches the maximum amount of hardware parallelism on the device. It is also
55+
very important that the thread mask remains consistent while interfacing with
56+
the port.
57+
58+
.. image:: ./rpc-diagram.svg
59+
:width: 75%
60+
:align: center
61+
62+
The above diagram outlines the architecture of the RPC interface. For clarity
63+
the following list will explain the operations done by the client and server
64+
respectively when initiating a communication.
65+
66+
First, a communication from the perspective of the client:
67+
68+
* The client searches for an available port and claims the lock.
69+
* The client checks that the port is still available to the current device and
70+
continues if so.
71+
* The client writes its data to the fixed-size packet and toggles its outbox.
72+
* The client waits until its inbox matches its outbox.
73+
* The client reads the data from the fixed-size packet.
74+
* The client closes the port and continues executing.
75+
76+
Now, the same communication from the perspective of the server:
77+
78+
* The server searches for an available port with pending work and claims the
79+
lock.
80+
* The server checks that the port is still available to the current device.
81+
* The server reads the opcode to perform the expected operation, in this
82+
case a receive and then send.
83+
* The server reads the data from the fixed-size packet.
84+
* The server writes its data to the fixed-size packet and toggles its outbox.
85+
* The server closes the port and continues searching for ports that need to be
86+
serviced
87+
88+
This architecture currently requires that the host periodically checks the RPC
89+
server's buffer for ports with pending work. Note that a port can be closed
90+
without waiting for its submitted work to be completed. This allows us to model
91+
asynchronous operations that do not need to wait until the server has completed
92+
them. If an operation requires more data than the fixed size buffer, we simply
93+
send multiple packets back and forth in a streaming fashion.
94+
95+
Server Library
96+
--------------
97+
98+
The RPC server's basic functionality is provided by the LLVM C library. A static
99+
library called ``libllvmlibc_rpc_server.a`` includes handling for the basic
100+
operations, such as printing or exiting. This has a small API that handles
101+
setting up the unified buffer and an interface to check the opcodes.
102+
103+
Some operations are too divergent to provide generic implementations for, such
104+
as allocating device accessible memory. For these cases, we provide a callback
105+
registration scheme to add a custom handler for any given opcode through the
106+
port API. More information can be found in the installed header
107+
``<install>/include/gpu-none-llvm/rpc_server.h``.
108+
109+
Client Example
110+
--------------
111+
112+
The Client API is not currently exported by the LLVM C library. This is
113+
primarily due to being written in C++ and relying on internal data structures.
114+
It uses a simple send and receive interface with a fixed-size packet. The
115+
following example uses the RPC interface to call a function pointer on the
116+
server.
117+
118+
This code first opens a port with the given opcode to facilitate the
119+
communication. It then copies over the argument struct to the server using the
120+
``send_n`` interface to stream arbitrary bytes. The next send operation provides
121+
the server with the function pointer that will be executed. The final receive
122+
operation is a no-op and simply forces the client to wait until the server is
123+
done. It can be omitted if asynchronous execution is desired.
124+
125+
.. code-block:: c++
126+
127+
void rpc_host_call(void *fn, void *data, size_t size) {
128+
rpc::Client::Port port = rpc::client.open<RPC_HOST_CALL>();
129+
port.send_n(data, size);
130+
port.send([=](rpc::Buffer *buffer) {
131+
buffer->data[0] = reinterpret_cast<uintptr_t>(fn);
132+
});
133+
port.recv([](rpc::Buffer *) {});
134+
port.close();
135+
}
136+
137+
Server Example
138+
--------------
139+
140+
This example shows the server-side handling of the previous client example. When
141+
the server is checked, if there are any ports with pending work it will check
142+
the opcode and perform the appropriate action. In this case, the action is to
143+
call a function pointer provided by the client.
144+
145+
In this example, the server simply runs forever in a separate thread for
146+
brevity's sake. Because the client is a GPU potentially handling several threads
147+
at once, the server needs to loop over all the active threads on the GPU. We
148+
abstract this into the ``lane_size`` variable, which is simply the device's warp
149+
or wavefront size. The identifier is simply the threads index into the current
150+
warp or wavefront. We allocate memory to copy the struct data into, and then
151+
call the given function pointer with that copied data. The final send simply
152+
signals completion and uses the implicit thread mask to delete the temporary
153+
data.
154+
155+
.. code-block:: c++
156+
157+
for(;;) {
158+
auto port = server.try_open(index);
159+
if (!port)
160+
return continue;
161+
162+
switch(port->get_opcode()) {
163+
case RPC_HOST_CALL: {
164+
uint64_t sizes[LANE_SIZE];
165+
void *args[LANE_SIZE];
166+
port->recv_n(args, sizes, [&](uint64_t size) { return new char[size]; });
167+
port->recv([&](rpc::Buffer *buffer, uint32_t id) {
168+
reinterpret_cast<void (*)(void *)>(buffer->data[0])(args[id]);
169+
});
170+
port->send([&](rpc::Buffer *, uint32_t id) {
171+
delete[] reinterpret_cast<uint8_t *>(args[id]);
172+
});
173+
break;
174+
}
175+
default:
176+
port->recv([](rpc::Buffer *) {});
177+
break;
178+
}
179+
}
180+
181+
CUDA Server Example
182+
-------------------
183+
184+
The following code shows an example of using the exported RPC interface along
185+
with the C library to manually configure a working server using the CUDA
186+
language. Other runtimes can use the presence of the ``__llvm_libc_rpc_client``
187+
in the GPU executable as an indicator for whether or not the server can be
188+
checked. These details should ideally be handled by the GPU language runtime,
189+
but the following example shows how it can be used by a standard user.
190+
191+
.. code-block:: cuda
192+
193+
#include <cstdio>
194+
#include <cstdlib>
195+
#include <cuda_runtime.h>
196+
197+
#include <gpu-none-llvm/rpc_server.h>
198+
199+
[[noreturn]] void handle_error(cudaError_t err) {
200+
fprintf(stderr, "CUDA error: %s\n", cudaGetErrorString(err));
201+
exit(EXIT_FAILURE);
202+
}
203+
204+
[[noreturn]] void handle_error(rpc_status_t err) {
205+
fprintf(stderr, "RPC error: %d\n", err);
206+
exit(EXIT_FAILURE);
207+
}
208+
209+
// The handle to the RPC client provided by the C library.
210+
extern "C" __device__ void *__llvm_libc_rpc_client;
211+
212+
__global__ void get_client_ptr(void **ptr) { *ptr = __llvm_libc_rpc_client; }
213+
214+
// Obtain the RPC client's handle from the device. The CUDA language cannot look
215+
// up the symbol directly like the driver API, so we launch a kernel to read it.
216+
void *get_rpc_client() {
217+
void *rpc_client = nullptr;
218+
void **rpc_client_d = nullptr;
219+
220+
if (cudaError_t err = cudaMalloc(&rpc_client_d, sizeof(void *)))
221+
handle_error(err);
222+
get_client_ptr<<<1, 1>>>(rpc_client_d);
223+
if (cudaError_t err = cudaDeviceSynchronize())
224+
handle_error(err);
225+
if (cudaError_t err = cudaMemcpy(&rpc_client, rpc_client_d, sizeof(void *),
226+
cudaMemcpyDeviceToHost))
227+
handle_error(err);
228+
return rpc_client;
229+
}
230+
231+
// Routines to allocate mapped memory that both the host and the device can
232+
// access asychonrously to communicate with eachother.
233+
void *alloc_host(size_t size, void *) {
234+
void *sharable_ptr;
235+
if (cudaError_t err = cudaMallocHost(&sharable_ptr, sizeof(void *)))
236+
handle_error(err);
237+
return sharable_ptr;
238+
};
239+
240+
void free_host(void *ptr, void *) {
241+
if (cudaError_t err = cudaFreeHost(ptr))
242+
handle_error(err);
243+
}
244+
245+
// The device-side overload of the standard C function to call.
246+
extern "C" __device__ int puts(const char *);
247+
248+
// Calls the C library function from the GPU C library.
249+
__global__ void hello() { puts("Hello world!"); }
250+
251+
int main() {
252+
int device = 0;
253+
// Initialize the RPC server to run on a single device.
254+
if (rpc_status_t err = rpc_init(/*num_device=*/1))
255+
handle_error(err);
256+
257+
// Initialize the RPC server to run on the given device.
258+
if (rpc_status_t err =
259+
rpc_server_init(device, RPC_MAXIMUM_PORT_COUNT,
260+
/*warp_size=*/32, alloc_host, /*data=*/nullptr))
261+
handle_error(err);
262+
263+
// Initialize the RPC client by copying the buffer to the device's handle.
264+
void *rpc_client = get_rpc_client();
265+
if (cudaError_t err =
266+
cudaMemcpy(rpc_client, rpc_get_client_buffer(device),
267+
rpc_get_client_size(), cudaMemcpyHostToDevice))
268+
handle_error(err);
269+
270+
cudaStream_t stream;
271+
if (cudaError_t err = cudaStreamCreate(&stream))
272+
handle_error(err);
273+
274+
// Execute the kernel.
275+
hello<<<1, 1, 0, stream>>>();
276+
277+
// While the kernel is executing, check the RPC server for work to do.
278+
while (cudaStreamQuery(stream) == cudaErrorNotReady)
279+
if (rpc_status_t err = rpc_handle_server(device))
280+
handle_error(err);
281+
282+
// Shut down the server running on the given device.
283+
if (rpc_status_t err =
284+
rpc_server_shutdown(device, free_host, /*data=*/nullptr))
285+
handle_error(err);
286+
287+
// Shut down the entire RPC server interface.
288+
if (rpc_status_t err = rpc_shutdown())
289+
handle_error(err);
290+
291+
return EXIT_SUCCESS;
292+
}
293+
294+
The above code must be compiled in CUDA's relocatable device code mode and with
295+
the advanced offloading driver to link in the library. Currently this can be
296+
done with the following invocation. Using LTO avoids the overhead normally
297+
associated with relocatable device code linking.
298+
299+
.. code-block:: sh
300+
301+
$> clang++ -x cuda rpc.cpp --offload-arch=native -fgpu-rdc -lcudart -lcgpu \
302+
-I<install-path>include -L<install-path>/lib -lllvmlibc_rpc_server \
303+
-O3 -foffload-lto -o hello
304+
$> ./hello
305+
Hello world!
18306
19307
Extensions
20308
----------

libc/docs/gpu/testing.rst

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -18,9 +18,9 @@ Testing Infrastructure
1818
======================
1919

2020
The testing support in LLVM's libc implementation for GPUs is designed to mimic
21-
the standard unit tests as much as possible. We use the `remote procedure call
22-
<libc_gpu_rpc>`_ support to provide the necessary utilities like printing from
23-
the GPU. Execution is performed by emitting a ``_start`` kernel from the GPU
21+
the standard unit tests as much as possible. We use the :ref:`libc_gpu_rpc`
22+
support to provide the necessary utilities like printing from the GPU. Execution
23+
is performed by emitting a ``_start`` kernel from the GPU
2424
that is then called by an external loader utility. This is an example of how
2525
this can be done manually:
2626

0 commit comments

Comments
 (0)