<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<p>Hi LLVM-Developers and Lang,</p>
<p>I solved the relocation problem and another problem, so I have a
working cuda-runtime-code-interpreter [1].</p>
<p>The solution of the relocation problem is, that I change die
reloc mode from dynamic relocation to PIC (position independent
code).</p>
<p>The second problem was, that I got an <span
class="m_-8219667011232476936enum-member-name-def">cudaErrorInvalidDeviceFunctio<wbr>n
error (error code 8), if I want to run a kernel. After some
research, I found out, that the kernel-code will be registered
(__cudaRegisterFatBinary(...) ) in a global constructor, which
is generated by the cuda backend (</span><span
class="m_-8219667011232476936enum-member-name-def">lib/CodeGen/CGCUDANV.cpp).
The ctor should start before the main function but I called the
main directly. So the error was caused by running directly the
main function and skipped the global cuda ctor and dtor. So I
wrote a fix, which runs the ctor and dtor before and after the
main and all works fine.</span></p>
<p><span class="m_-8219667011232476936enum-member-name-def">Cheers,<br>
Simeon<br>
</span></p>
<br>
<div class="moz-cite-prefix">Am 14.11.2017 um 22:15 schrieb Simeon
Ehrig:<br>
</div>
<blockquote type="cite"
cite="mid:ff92bab2-a5b4-ba07-e3ac-e2fe93c4e3e9@tu-dresden.de">
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<p>Hi Lang,</p>
<p>thank You very much. I've used Your code and the creating of
the object file works. I think the problem is after creating the
object file. When I link the object file with ld I get an
executable, which is working right.</p>
<p>After changing the clang and llvm libraries from the package
control version (.deb) to a own compiled version with debug
options, I get an assert() fault.<br>
In <br>
void RuntimeDyldELF::resolveX86_64Relocation() at the case
ELF::R_X86_64_PC32 <br>
this will throw an assert. You can find the code in the file
llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp . I
don't know exactly, what this function do but after first
research, I think it has something to do with the linking. Maybe
You know more about the function?</p>
<p>Your code also helps me to understand more, how the interpreter
library works. I have also some new ideas, how I could find the
concrete problem and solve it.</p>
Cheers,<br>
Simeon<br>
<br>
<div class="moz-cite-prefix">Am 09.11.2017 um 00:59 schrieb Lang
Hames:<br>
</div>
<blockquote type="cite"
cite="mid:CALLttgqHaBTinwVqC3PGf01KVa8m-0iEaAoSz7SA00+TKWfkDw@mail.gmail.com">
<div dir="ltr">Hi Simon,
<div><br>
</div>
<div>I think the best thing would be to add an
ObjectTransformLayer between your CompileLayer and
LinkingLayer so that you can capture the object files as
they're generated. Then you can inspect the object files
being generated by the compiler to see what might be wrong
with them.</div>
<div><br>
</div>
<div>Something like this:</div>
<div><br>
</div>
<div>
<div><font face="monospace, monospace">class KaleidoscopeJIT
{</font></div>
<div><font face="monospace, monospace">private:</font></div>
<div><font face="monospace, monospace"><br>
</font></div>
<div><font face="monospace, monospace"> using ObjectPtr =
std::shared_ptr<object::OwningBinary<object::ObjectFile>>;</font></div>
<div><font face="monospace, monospace"><br>
</font></div>
<div><font face="monospace, monospace"> static ObjectPtr
dumpObject(ObjectPtr Obj) {</font></div>
<div><font face="monospace, monospace">
SmallVector<char, 256> UniqueObjFileName;</font></div>
<div><font face="monospace, monospace">
sys::fs::createUniqueFile("jit-object-%%%.o",
UniqueObjFileName);</font></div>
<div><font face="monospace, monospace"> std::error_code
EC;</font></div>
<div><font face="monospace, monospace"> raw_fd_ostream
ObjFileStream(UniqueObjFileName.data(), EC,
sys::fs::F_RW);</font></div>
<div><font face="monospace, monospace">
ObjFileStream.write(Obj->getBinary()->getData().data(),
</font></div>
<div><font face="monospace, monospace">
Obj->getBinary()->getData().size());
</font></div>
<div><font face="monospace, monospace"> return Obj;</font></div>
<div><font face="monospace, monospace"> }</font></div>
<div><font face="monospace, monospace"><br>
</font></div>
<div><font face="monospace, monospace">
std::unique_ptr<TargetMachine> TM;</font></div>
<div><font face="monospace, monospace"> const DataLayout
DL;</font></div>
<div><font face="monospace, monospace">
RTDyldObjectLinkingLayer ObjectLayer;</font></div>
<div><font face="monospace, monospace">
ObjectTransformLayer<decltype(ObjectLayer),</font></div>
<div><font face="monospace, monospace">
decltype(&KaleidoscopeJIT::dumpObject)>
DumpObjectsLayer;</font></div>
<div><font face="monospace, monospace">
IRCompileLayer<decltype(DumpObjectsLayer),
SimpleCompiler> CompileLayer;</font></div>
<div><font face="monospace, monospace"><br>
</font></div>
<div><font face="monospace, monospace">public:</font></div>
<div><font face="monospace, monospace"> using ModuleHandle
= decltype(CompileLayer)::ModuleHandleT;</font></div>
<div><font face="monospace, monospace"><br>
</font></div>
<div><font face="monospace, monospace"> KaleidoscopeJIT()</font></div>
<div><font face="monospace, monospace"> :
TM(EngineBuilder().selectTarget()),
DL(TM->createDataLayout()),</font></div>
<div><font face="monospace, monospace">
ObjectLayer([]() { return
std::make_shared<SectionMemoryManager>(); }),</font></div>
<div><font face="monospace, monospace">
DumpObjectsLayer(ObjectLayer,
&KaleidoscopeJIT::dumpObject),</font></div>
<div><font face="monospace, monospace">
CompileLayer(DumpObjectsLayer, SimpleCompiler(*TM)) {</font></div>
<div><font face="monospace, monospace">
llvm::sys::DynamicLibrary::LoadLibraryPermanently(nullptr);</font></div>
<div><font face="monospace, monospace"> }</font></div>
</div>
<div><br>
</div>
<div>Hope this helps!</div>
<div><br>
</div>
<div>Cheers,</div>
<div>Lang.</div>
<div><br>
</div>
</div>
<div class="gmail_extra"><br>
<div class="gmail_quote">On Wed, Sep 27, 2017 at 10:32 AM,
Simeon Ehrig via llvm-dev <span dir="ltr"><<a
href="mailto:llvm-dev@lists.llvm.org" target="_blank"
moz-do-not-send="true">llvm-dev@lists.llvm.org</a>></span>
wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0
.8ex;border-left:1px #ccc solid;padding-left:1ex">
<div text="#000000" bgcolor="#FFFFFF">
<p>Dear LLVM-Developers and Vinod Grover,<br>
<br>
we are trying to extend the cling C++ interpreter (<a
class="m_-8219667011232476936moz-txt-link-freetext"
href="https://github.com/root-project/cling"
target="_blank" moz-do-not-send="true">https://github.com/root-<wbr>project/cling</a>)
with CUDA functionality for Nvidia GPUs.<br>
<br>
I already developed a prototype based on OrcJIT and am
seeking for feedback. I am currently a stuck with a
runtime issue, on which my interpreter prototype fails
to execute kernels with a CUDA runtime error.<br>
</p>
<p><br>
=== How to use the prototype<br>
<br>
This application interprets cuda runtime code. The
program needs the whole cuda-program (.cu-file) and
its pre-compiled device code (as fatbin) as an input:<br>
<br>
command: cuda-interpreter [source].cu
[kernels].fatbin<br>
<br>
I also implemented an alternative mode, which is
generating an object file. The object file can be
linked (ld) to an exectuable. This mode is just
implemented to check if the LLVM module generation
works as expected. Activate it by changing the define
INTERPRET from 1 to 0 .<br>
<br>
=== Implementation<br>
<br>
The prototype is based on the clang example in<br>
<br>
<a class="m_-8219667011232476936moz-txt-link-freetext"
href="https://github.com/llvm-mirror/clang/tree/master/examples/clang-interpreter"
target="_blank" moz-do-not-send="true">https://github.com/llvm-<wbr>mirror/clang/tree/master/<wbr>examples/clang-interpreter</a><br>
<br>
I also pushed the source code to github with the
install instructions and examples:<br>
<a
class="m_-8219667011232476936moz-txt-link-freetext"
href="https://github.com/SimeonEhrig/CUDA-Runtime-Interpreter"
target="_blank" moz-do-not-send="true">https://github.com/<wbr>SimeonEhrig/CUDA-Runtime-<wbr>Interpreter</a><br>
<br>
The device code generation can be performed with
either clang's CUDA frontend or NVCC to ptx.<br>
<br>
Here is the workflow in five stages:<br>
</p>
<ol>
<li>generating ptx device code (a kind of nvidia
assembler)</li>
<li>translate ptx to sass (machine code of ptx)</li>
<li>generate a fatbinray (a kind of wrapper for the
device code)</li>
<li>generate host code object file (use fatbinary as
input)</li>
<li>link to executable</li>
</ol>
<p>(The exact commands are stored in the commands.txt in
the github repo)<br>
<br>
The interpreter replaces the 4th and 5th step. It
interprets the host code with pre-compiled device code
as fatbinary. The fatbinary (Step 1 to 3) will be
generated with the clang compiler and the nvidia tools
ptxas and fatbinary.<br>
<br>
=== Test Cases and Issues<br>
<br>
You find the test sources on GitHub in the directory
"example_prog".<br>
<br>
Run the tests with cuda-interpeter and the two
arguments as above:<br>
<br>
[1] path to the source code in "example_prog"<br>
- note: even for host-only code, use the
file-ending .cu<br>
<br>
[2] path to the runtime .fatbin<br>
- note: needs the file ending .fatbin<br>
- a fatbin file is necessary, but if the program
doesn't need a kernel, the content of the file will
ignore</p>
Note: As a prototype, the input is just static and
barely checked yet.<br>
<br>
1. <a href="http://hello.cu" target="_blank"
moz-do-not-send="true">hello.cu</a>: simple c++ hello
world program with cmath library call sqrt() -> works
without problems<br>
<br>
2. <a href="http://pthread_test.cu" target="_blank"
moz-do-not-send="true">pthread_test.cu</a>: c++
program, which starts a second thread -> works
without problems<br>
<br>
3. <a href="http://fat_memory.cu" target="_blank"
moz-do-not-send="true">fat_memory.cu</a>: use cuda
library and allocate about 191 MB of VRAM. After the
allocation, the program waits for 3 seconds, so you can
check the memory usage with the nvidia-smi -> works
without problems<br>
<br>
4. <a href="http://runtime.cu" target="_blank"
moz-do-not-send="true">runtime.cu</a>: combine cuda
library with a simple cuda kernel -> Generating an
object file, which can be linked (see 5th call in
commands above -> ld ...) to a working executable.<br>
<br>
The last example has the following issues: Running the
executable works fine. Interpreting the code instead
does not work. The Cuda Runtime returns the error 8 (<span
class="m_-8219667011232476936enum-member-name-def">cudaErrorInvalidDeviceFunctio<wbr>n</span>)
, the kernel failed.<br>
<br>
Do you have any idea how to proceed?<br>
<br>
<br>
Best regards,<br>
Simeon Ehrig </div>
<br>
______________________________<wbr>_________________<br>
LLVM Developers mailing list<br>
<a href="mailto:llvm-dev@lists.llvm.org"
moz-do-not-send="true">llvm-dev@lists.llvm.org</a><br>
<a
href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev"
rel="noreferrer" target="_blank" moz-do-not-send="true">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-dev</a><br>
<br>
</blockquote>
</div>
<br>
</div>
</blockquote>
<br>
</blockquote>
<br>
</body>
</html>