Run a Sample HCFFT ? about hcc HOT 15 CLOSED

from hcc.

@psteinb No, we tried it and we were getting error. We installed ROCm on AMD 7860K APU, then we followed instructions in the Bit bucket, but we got some runtime issues . So I was confused with the steps I was doing wrong.

from hcc.

@VishwasRao17 A10-7860K uses a different ISA version (7:0:0) than the default one (8:0:3). Could you try build HCFFT with an extra flag --amdgpu-target=gfx700?

from hcc.

@whchung Sure, I will check that and let you know. Thank you.
By the way, the A10-7860K is well supported by ROCM and HCC right. I mean from Hardware perspective

from hcc.

From compiler perspective HCC is able to produce ISA for A10-7860K. But some OS & runtime-level features may not be present on Kaveri architecture, so we don't claim it's fully supported now. For more information please refer to: https://github.com/RadeonOpenCompute/ROCm#are-you-ready-to-rock .

from hcc.

@whchung Thank you, but in ROCM Kernel Driver we see tested HW platform is found to be
The following is a reference hardware configuration that was used for testing purposes:

APU Config:
•APU: AMD A10-7850K APU
•Motherboard: ASUS A88X-PRO motherboard (ATX form factor)
•Memory: G.SKILL Ripjaws X Series 16GB (2 x 8GB) 240-Pin DDR3 SDRAM DDR3 2133
•OS: Ubuntu 14.04 64-bit edition
•No discrete GPU present in the system

"https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver/tree/roc-1.4.0"

So I had bought this system for development.

from hcc.

@whchung me and @VishwasRao17 has done the ROCm installation and testing of vector_copy example to test if it is installed properly is successful. now we want to compile hcfft_1D_R2C.cpp as shown in https://bitbucket.org/multicoreware/hcfft/src and compilation is successful using following command
/opt/rocm/hcc/bin/clang++ /opt/rocm/hcc/bin/hcc-config --cxxflags --ldflags -lhc_am -L/opt/rocm/hcfft/lib -lhcfft hcfft_1D_R2C.cpp -o hcfft_1D_R2C.out
and when i try to execute it we get following error:
No suitable runtime detected. Fall back to CPU!
Segmentation fault (core dumped)

•APU used AMD A10-7850K APU for above experiment.

from hcc.

@vrajeshlm could you let me know the following information:

• Linux distro version
• g++ --version
• /opt/rocm/hcc/bin/clang++ --version
• /opt/rocm/hip/bin/hipconfig

I would also recommend you open an ticket at HCFFT site, as it seems to be a library issues instead of a compiler issue.

from hcc.

@whchung

• Linux Distro version: Ubuntu 16.04 64 bit (Ubuntu 16.04.4 64 bit)
• g++ --version: 5.4.0
• /opt/rocm/hcc/bin/clang++ --version: HCC clang version 3.5.0 (based on HCC 0.10.16501-81f0a2f-02246a0 LLVM 3.5.0svn) ... Target: x86_64-unknown-linux-gnu ... Thread model: posix

from hcc.

The configuration seem pretty normal. I do recommend you open a ticket at HCFFT site, as to my understanding, HCFFT would invoke extra HCC instances on-the-fly to generate FFT kernels. And perhaps in its implementation it doesn't pass --amdgpu-target=gfx700 to HCC.

from hcc.

@whchung i tried using --amdgpu-target=gfx700 option using following command:
/opt/rocm/hcc/bin/hcc /opt/rocm/hcc/bin/hcc-config --cxxflags --ldflags --amdgpu-target=gfx700 -lhc_am -lhcfft hcfft_1D_R2C.cpp
i get warning: -amdgpu-target argument 'gfx700' is not recognized, using fiji instead
i hope this is fine but still i am getting no suitable runtime detected. fall back to cpu!

from hcc.

@vrajeshlm unfortunately if you see the warning then it's not fine. Fiji is of ISA version 8:0:3, but A10-7850K is based on Kaveri which is of ISA version 7:0:0.

As a first step we need to get rid of this warning message, could you try replace the flag with --amdgpu-target=AMD:AMDGPU:7:0:0 ? That was the old (now obsolete) syntax to specify ISA version for the compiler. I assume you are using stock HCC from ROCm 1.4 installation so that may be why "gfx700" doesn't work for you.

If you are past the compile-time warning, the next step is to fix the runtime error. I may need more information about when the error is raised when you run the executable to help determine whether it's an ROCR/HSA runtime issue, an HCC language runtime issue, or an HCFFT library issue.

To help isolate the issue, please first use this HCC unit test:
https://github.com/RadeonOpenCompute/hcc/blob/roc-1.4.x/tests/Unit/HC/reduction_hc.cpp

compile it with:

/opt/rocm/hcc/bin/clang `/opt/rocm/hcc/bin/hcc-config --cxxflags --ldflags` --amdgpu-target=AMD:AMDGPU:7:0:0 reduction_hc.cpp

and see if it runs. If it does run, then it's more likely an HCFFT issue.

from hcc.

@whchung as per your guidance I tried reduction_hc.cpp and it compiled and runs fine.
following is the output of reduction_hc.cpp program which shows it is detecting amdgpu gfx7000:
Using device : AMD HSA Agent gfx7000
Running kernels...
SUCCESS: sequential_reduction.
SUCCESS: reduction_simple_1.
SUCCESS: reduction_simple_2.
SUCCESS: reduction_tiled_1.
SUCCESS: reduction_tiled_2.
SUCCESS: reduction_tiled_3.
SUCCESS: reduction_tiled_4.
SUCCESS: reduction_cascade.

according to this it looks more like Hcfft issue. we will raise the query on HCFFT website.

thanks for your help. If you suggest us on any other installation related checks, then it will be really good for us.

from hcc.

A10-7860K is not officially supported Processor in ROCm, please refer to the officially supported hardware section at https://rocm.github.io/hardware.html.

I will get the Readme corrected

Best regards

G

from hcc.

@whchung is there any way to declare this amdgpu hcc compiler flag as a variable which will be taken care for any further compilation as I have many dependent files and libraries to compile.

from hcc.