Can't access global variables about hcc HOT 10 CLOSED

note that the same problem exists when using static local variables.
fails.cpp:

#include <amp.h>
#include <iostream>

int &get_value() restrict(amp,cpu)
{
        static int v = 6;
        return v;
}

int main(void)
{
        ::std::atomic_uint sum;
        get_value() = 8;
        sum = get_value();
        parallel_for_each(concurrency::extent<1>(100),
                          [&](concurrency::index<1> i) restrict (amp)
        {
                sum += get_value();
        });
        ::std::cout << sum << "\n";
        return 0;
}

$ ./test
608

works.cpp:

#include <amp.h>

#include <iostream>

int &get_value() restrict(amp,cpu)
{
        static int v = 6;
        return v;
}

int main(void)
{
        ::std::atomic_uint sum;
        get_value() = 7;
        sum = get_value();
        int &val_l = get_value();
        parallel_for_each(concurrency::extent<1>(100),
                          [&](concurrency::index<1> i) restrict (amp)
        {
                sum += val_l;
        });
        ::std::cout << sum << "\n";
        return 0;
}

$ ./test
707

from hcc.

note program built with the latest hcc gives this error:

$ ./atomic-global 

Backtrace:
0x00007fc48fbe4c91:	Kalmar::HSADevice::BuildOfflineFinalizedProgramImpl(void*, int) + 0x221
0x00007fc48fbd4660:	Kalmar::HSADevice::BuildProgram(void*, void*) + 0x140
0x000000000040fe54:	Kalmar::KalmarBootstrap::KalmarBootstrap() + 0x124
0x000000000040fd09:	__hcc_shared_library_init + 0x29
0x0000000000410f0d:	__libc_csu_init + 0x4d
0x00007fc4906b149b:	__libc_start_main + 0x7b
0x0000000000410e0a:	_start + 0x2a

### HCC STATUS_CHECK Error: HSA_STATUS_ERROR_VARIABLE_UNDEFINED (0x1015) at file:mcwamp_hsa.cpp line:2918
Aborted (core dumped)

from hcc.

We get the same error when a constexpr static data member of a class is odr-used (required definition at namespace scope is provided).

from hcc.

@jvesely regarding your first example, this now outputs "101"

perhaps it has been resolved by now?

from hcc.

Thanks. My machines are offline atm. I'll try to rerun the tests next week.

from hcc.

@dekken my machines are back up, but I can't reproduce your success.
the latest hcc fails to build (configure failure with incorrect clang path), after fixing that issue it complains when compiling for my gpu:

-amdgpu-target argument 'gfx801' is not recognized; using gfx803 instead [-Winvalid-command-line-argument]
objcopy: /tmp/tmp.Xlqbe1bWHS/static-test-6023c5.host.o: Failed to find link section for section 19
objcopy: /tmp/tmp.Xlqbe1bWHS/static-test-6023c5.host.o: Failed to find link section for section 19

it does produce a binary though, but trying to run it results in:
### HCC STATUS_CHECK Error: HSA_STATUS_ERROR_INVALID_ISA_NAME (0x1017) at file:mcwamp_hsa.cpp line:2683

from hcc.

@jvesely The original case works for me with the latest rocm 1.9.

from hcc.

I'm trying to setup rocm 1.9 on my machine but it looks like carrizo (or is it all apus?) is not supported.
I fixed the objcopy complaint so now I just get the warning and "invalid isa name" crash.
So I don't think I'll be able to confirm whether this has been resolved.

from hcc.

Turns out that my setup is no longer supported. Clang code explicitly includes:

bool is_valid(const std::string& gfxip)
     {
         static constexpr std::array<const char*, 4u> valid = {
            { "gfx701", "gfx803", "gfx900", "gfx906" }};
 
         return std::find(valid.cbegin(), valid.cend(), gfxip) != valid.cend();
     }

It'd be nice if AMD was less cryptic about throwing users with older HW overboard. Especially since Carrizo APU was advertised for HSA ("And the upcoming “Carrizo” APU takes HSA support even further, with design features to make it fully compliant with the HSA 1.0 specification.​")[0].

[0] https://www.amd.com/en-us/innovations/software-technologies/server-solution/hsa

from hcc.

After recovering Carrizo suport in hcc, I can confirm that global variable test works as expected.
The static variable test does not, I guess the __CPU__ and __HC__ functions are different with different static namespace.

from hcc.