Summary:
This patch adds a temporary implementation that uses a struct-based
interface in lieu of varargs support. Once varargs support exists we
will move this implementation to the "real" printf implementation.
Conceptually, this patch has the client copy over its format string and
arguments to the server. The server will then scan the format string
searching for any specifiers that are actually a string. If it is a
string then we will send the pointer back to the server to tell it to
copy it back. This copied value will then replace the pointer when the
final formatting is done.
This will require a built-in extension to the varargs support to get
access to the underlying struct. The varargs used on the GPU will simply
be a struct wrapped in a varargs ABI.
Summary:
The current implementation of RPC tied everything to device IDs and
forced us to do init / shutdown to manage some global state. This turned
out to be a bad idea in situations where we want to track multiple
hetergeneous devices that may report the same device ID in the same
process.
This patch changes the interface to instead create an opaque handle to
the internal device and simply allocates it via `new`. The user will
then take this device and store it to interface with the attached
device. This interface puts the burden of tracking the device identifier
to mapped d evices onto the user, but in return heavily simplifies the
implementation.
Fixes#86546 and removes the macro `LIBC_HAS_BUILTIN`. This was
necessary to support older compilers that did not support
`__has_builtin`. All of the compilers we support already have this
builtin.
See: https://libc.llvm.org/compiler_support.html
All uses now use `__has_builtin` directly
cc @nickdesaulniers
This patch updates the construction of packet headers to replace the
usage of ACQUIRE/RELEASE with SCACQUIRE/SCRELEASE which is now
recommended.
The patch also ensures consistency across kernel dispatches.
Summary:
The libc build has a few utilties that need to be built before we can do
everything in the full build. The one requirement currently is the
`libc-hdrgen` binary. If we are doing a full build runtimes mode we
first add `libc` to the projects list and then only use the `projects`
portion to buld the `libc` portion. We also use utilities for the GPU
build, namely the loader utilities. Previously we would build these
tools on-demand inside of the cross-build, which tool some hacky
workarounds for the dependency finding and target triple. This patch
instead just builds them similarly to libc-hdrgen and then passses them
in. We now either pass it manually it it was built, or just look it up
like we do with the other `clang` tools.
Depends on https://github.com/llvm/llvm-project/pull/84664
Summary:
We previously changed the data layout for the RPC buffer to make it lane
size agnostic. I put off changing the size for the server case to make
the patch smaller. This patch simply reorganizes code by making the lane
size an argument to the port rather than a templtae size. Heavily
simplifies a lot of code, no more `std::variant`.
Summary:
These values were incorrectly flipped, setting the size of the blocks to
the threads and vice-versa. When I originally wrote the thread utilities
it was using COV4 which used an implicit format. Then when I updated I
accidentally flipped them and never noticed because nothing depended on
the size of the threads until I checked it manually.
Summary:
Recent changes added an include path in the float128 type that used the
internal `libc` path to find the macro. This doesn't work once it's
installed because we need to search from the root of the install dir.
This patch adds "include/" to the include path so that our inclusion
of installed headers always match the internal use.
Summary:
This directory is leftover from when we handled both AMDGPU and NVPTX in
the same build and merged them into a pseudo triple. Now the only thing
it contains is the RPC server header. This gets rid of it, but now that
it's in the base install directory we should make it clear that it's an
LLVM libc header.
Summary:
This is a massive patch because it reworks the entire build and
everything that depends on it. This is not split up because various bots
would fail otherwise. I will attempt to describe the necessary changes
here.
This patch completely reworks how the GPU build is built and targeted.
Previously, we used a standard runtimes build and handled both NVPTX and
AMDGPU in a single build via multi-targeting. This added a lot of
divergence in the build system and prevented us from doing various
things like building for the CPU / GPU at the same time, or exporting
the startup libraries or running tests without a full rebuild.
The new appraoch is to handle the GPU builds as strict cross-compiling
runtimes. The first step required
https://github.com/llvm/llvm-project/pull/81557 to allow the `LIBC`
target to build for the GPU without touching the other targets. This
means that the GPU uses all the same handling as the other builds in
`libc`.
The new expected way to build the GPU libc is with
`LLVM_LIBC_RUNTIME_TARGETS=amdgcn-amd-amdhsa;nvptx64-nvidia-cuda`.
The second step was reworking how we generated the embedded GPU library
by moving it into the library install step. Where we previously had one
`libcgpu.a` we now have `libcgpu-amdgpu.a` and `libcgpu-nvptx.a`. This
patch includes the necessary clang / OpenMP changes to make that not
break the bots when this lands.
We unfortunately still require that the NVPTX target has an `internal`
target for tests. This is because the NVPTX target needs to do LTO for
the provided version (The offloading toolchain can handle it) but cannot
use it for the native toolchain which is used for making tests.
This approach is vastly superior in every way, allowing us to treat the
GPU as a standard cross-compiling target. We can now install the GPU
utilities to do things like use the offload tests and other fun things.
Some certain utilities need to be built with
`--target=${LLVM_HOST_TRIPLE}` as well. I think this is a fine
workaround as we
will always assume that the GPU `libc` is a cross-build with a
functioning host.
Depends on https://github.com/llvm/llvm-project/pull/81557
Summary:
The RPC interface needs to handle an entire warp or wavefront at once.
This is currently done by using a compile time constant indicating the
size of the buffer, which right now defaults to some value on the client
(GPU) side. However, there are currently attempts to move the `libc`
library to a single IR build. This is problematic as the size of the
wave fronts changes between ISAs on AMDGPU. The builitin
`__builtin_amdgcn_wavefrontsize()` will return the appropriate value,
but it is only known at runtime now.
In order to support this, this patch restructures the packet. Now
instead of having an array of arrays, we simply have a large array of
buffers and slice it according to the runtime value if we don't know it
ahead of time. This also somewhat has the advantage of making the buffer
contiguous within a page now that the header has been moved out of it.
Summary:
The RPC interface uses several ports to provide parallel access. Right
now we begin the search at the beginning, which heavily contests the
early ports. Using the SMID allows us to stagger the starting index
based off of the cluster identifier that is executing the current warp.
Multiple warps can share an SM, but it will guaruntee that the
contention for the low indices is lower.
This also increases the maximum port size to around 4096, this is
because 512 isn't enough to cover the full hardare parallelism needed to
guarantee this doesdn't deadlock.
Prior to this change, we wouldn't build headers that aren't referenced
by other parts of the libc which would result in a build error during
installation. To address this, we make the header target a dependency of
the libc archive. Additionally, we also redo the install targets, moving
the install targets closer to build targets and simplifying the
hierarchy and generally matching what we do for other runtimes.
Summary:
This pointer has been causing issues. Allocating and reading from coarse
memory on the CPU is not guaranteed and varies depending on the kernel
version and support. Previously we attempted to pin the memory but this
caused unexpected failures. This should be a legal operation and work
around the problem as fine-grained memory should be always legal to
write to by both sides.
Summary:
This may be problematic to pin a stack pointer. Allocate it via the OS
allocator instead as the documentation suggests.
For some reason, if you attempt to free this pointer after the memory
region has been unlocked, it will return an invalid pointer.
Summary:
Previously, we determined that coarse grained memory cannot be used in
the general case. That removed the buffer used to transfer the memory,
however we still had this lookup. Though we do not access the symbol
directly, it still conflicts with the agents apparently. Pin this as
well.
This resolves the problems @lntue was having with the `libc` GPU build.
Summary:
This portion of code handles mapping the RPC client memory over to the
device. HSA copies need to be between two slices of memory that HSA has
allocated. Previously we used coarse-grained memory to act as the host
source. However, the support for this varies depending on the kernel and
version and should not be relied upon. This patch changes that handling
to use the `hsa_amd_memory_lock` API to explicitly pin memory to a
location sufficient for a DMA transfer to the GPU.
Summary:
The puts call appends a newline. With multiple threads, this can be done
out of order such that another thread puts something before we finish
appending the newline. Add a flockfile and funlockfile to ensure that
the whole string is printed before another string can appear.
Summary:
This patch includes the necessary changes to make the `libc` tests
running on AMD GPUs run using the newer code object version. The 'code
object version' is AMD's internal ABI for making kernel calls. The move
from 4 to 5 changed how we handle arguments for builtins such as
obtaining the grid size or setting up the size of the private stack.
Fixes: https://github.com/llvm/llvm-project/issues/72517
If you build with dynamic_hsa, the symbol is known and compilation
succeeds. If you then run with a slightly older libhsa, this argument is
not recognised and an error returned. I'd rather the program runs with a
misleading omp wtime than refuses to run at all.
I think it follows from the HSA spec that a write to the first byte is
deemed significant to the GPU in which case writing to the second short
and reading back from it later would be safe. However, the examples for
this all involve an atomic write to the first 32 bits and it seems a
credible risk that the occasional CI errors abound invalid packets have
as their root cause that the firmware notices the early write to
packet->setup and treats that as a sign that the packet is ready to go.
That was overly-paranoid, however in passing noticed the code in libc is
genuinely invalid. The memset writes a zero to the header byte, changing
it from type_invalid (1) to type_vendor (0), at which point the GPU is
free to read the 64 byte packet and interpret it as a vendor packet,
which is probably why libc CI periodically errors about invalid packets.
Also a drive by change to do the atomic store on a uint32_t
consistently. I'm not sure offhand what __atomic_store_n on a uint16_t*
and an int resolves to, seems better to be unambiguous there.
Summary:
The `fgets` function as implemented is not functional currently when
called with multiple threads. This is because we rely on reapeatedly
polling the character to detect EOF. This doesn't work when there are
multiple threads that may with to poll the characters. this patch pulls
out the logic into a standalone RPC call to handle this in a single
operation such that calling it from multiple threads functions as
expected. It also makes it less slow because we no longer make N RPC
calls for N characters.
Summary:
This function follows closely with the pattern of all the other
functions. That is, making a new opcode and forwarding the call to the
host. However, this also required modifying the test somewhat. It seems
that not all `libc` implementations follow the same error rules as are
tested here, and it is not explicit in the standard, so we simply
disable these EOF checks when targeting the GPU.
Summary:
Currently, we use the RPC server to respond to different ports which
each contain a request from some client thread wishing to do work on the
server. This scan starts at zero and continues until its checked all
ports at which point it resets. If we find an active port, we service it
and then restart the search.
This is bad for two reasons. First, it means that we will always bias
the lower ports. If a thread grabs a high port it will be stuck for a
very long time until all the other work is done. Second, it means that
the `handle_server` function can technically run indefinitely as long as
the client is always pushing new work. Because the OpenMP implementation
uses the user thread to service the kernel, this means that it could be
stalled with another asyncrhonous device's kernels.
This patch addresses this by making the server restart at the next port
over. This means we will always do a full scan of the ports before
quitting.
Summary:
This variable needs a reserved name starting with `__`. It was
mistakenly changed with a mass replace. It happened to work because the
tests still picked up the associated symbol, but it just became a bad
name because it's not reserved anymore.
Summary:
This patch adds the necessary entrypoints to handle the `fseek`,
`fflush`, and `ftell` functions. These are all very straightfoward, we
simply make RPC calls to the associated function on the other end.
Implementing it this way allows us to more or less borrow the state of
the stream from the server as we intentionally maintain no internal
state on the GPU device. However, this does not implement the `errno`
functinality so that must be ignored.
Summary:
Normally, the implementation of `puts` simply writes a second newline
charcter after printing the first string. However, because the GPU does
everything in batches of the SIMT group size, this will end up with very
poor output where you get the strings printed and then 1-64 newline
characters all in a row. Optimizations like to turn `printf` calls into
`puts` so it's a good idea to make this produce the expected output.
The least invasive way I could do this was to add a new opcode. It's a
little bloated, but it avoids an unneccessary and slow send operation to
configure this.
Summary:
Previously, the `fread` operation was wrong in cases when we read less
data than was requested. That is, if we tried to read N bytes while the
file was in EOF, it would still copy N bytes of garbage. This is fixed
by only copying over the sizes we got from locally opening it rather
than just using the provided size.
Additionally, this patch simplifies the interface. The output functions
have special variants for writing to stdout / stderr. This is primarily
an optimization for these common cases so we can avoid sending the
stream as an argument which has a high delay. Because for input, we
already need to start with a `send` to tell the server how much data to
read, it costs us nothing to send the file along with it so this is
redundant. Re-use the file encoding scheme from the other
implementations, the one that stores the stream type in the LSBs of the
FILE pointer.
Summary:
This patch removes the `rpc_reset` function. This was previously used to
initialize the RPC client on the device by setting up the pointers to
communicate with the server. The purpose of this was to make it easier
to initialize the device for testing. However, this prevented us from
enforcing an invariant that the buffers are all read-only from the
client side.
The expected way to initialize the server is now to copy it from the
host runtime. This will allow us to maintain that the RPC client is in
the constant address space on the GPU, potentially through inference,
and improving caching behaviour.
Summary:
This patch implements the `fgets`, `getc`, `fgetc`, and `getchar`
functions on the GPU. Their implementations are straightforward enough.
One thing worth noting is that the implementation of `fgets` will be
extremely slow due to the high latency to read a single char. A faster
solution would be to make a new RPC call to call `fgets` (due to the
special rule that newline or null breaks the stream). But this is left
out because performance isn't the primary concern here.
A recent patch required the implementation to define `LIBC_NAMESPACE`.
For GPU offloading we provide a static library whose internal
implementation relies on the `libc` headers. This is a separate library
that is constructed during the "bootstrap" phase. This patch moves the
definition of the `LIBC_NAMESPACE` CMake variable up so its available
during bootstrapping and adds it to the definition of the RPC server.
Summary:
The HSA API explicitly states that the size is a count of uint32_t's not
a byte count. This was erroneously being used as a simple memcpy,
causing some weird behaviour. Fix this by correctly passing `1` to
initialize a single integer to zero.
This function implements the `abort` function on the GPU. The
implementation here closely mirros the `exit` call where we first
synchornize with the RPC server to make sure it's listening and then we
exit on the GPU.
I was unsure if this should be a simple `__builtin_assert` on the GPU. I
elected to go with an RPC approach to make this a more "true" `abort`
call. That is, it should invoke some signal handlers and exit with the
proper code according to the implemented C library on the server.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D159210
This `MAX_LANE_SIZE` was a hack from the days when we used a single
instance of the server and had some GPU state handle it. Now that we
have everything templated this really shouldn't be used. This patch
removes its use and replaces it with template arguments.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D158633
This patch adds support for `fread` on the GPU via the RPC mechanism.
Here we simply pass the size of the read to the server and then copy it
back to the client via the RPC channel. This should allow us to do the
basic operations on files now. This will obviously be slow for large
sizes due ot the number of RPC calls involved, this could be optimized
further by having a special RPC call that can initiate a memcpy between
the two pointers.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D155121
This patch does the noisy work of removing the test opcodes from the
exported interface to an interface that is only visible in `libc`. The
benefit of this is that we both test the exported RPC registration more
directly, and we do not need to give this interface to users.
I have decided to export any opcode that is not a "core" libc feature as
having its MSB set in the opcode. We can think of these as non-libc
"extensions".
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D154848
Currently we keep an internal buffer of device memory that is used to
indicate ownership of a port. Since we only use this as a single bit we
can simply turn this into a bitfield. I did this manually rather than
having a separate type as we need very special handling of the masks
used to interact with the locks.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D155511
