https://llvm.org/bugs/show_bug.cgi?id=28919
Bug ID: 28919
Summary: Use of undeclared identifier 'surf2Dwrite'
Product: clang
Version: trunk
Hardware: Macintosh
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: CUDA
Assignee: unassignedclangb...@nondot.org
Reporter: llvm.org_thaek...@pmoreau.org
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Created attachment 16923
--> https://llvm.org/bugs/attachment.cgi?id=16923&action=edit
File containing the CUDA kernel
Hello,
I wanted to give clang a go on compiling my CUDA kernels. I was able to get the
example at bit.ly/llvm-cuda to compile and run properly, but when I tried the
attached code, I got the error message "error: use of undeclared identifier
'surf2Dwrite'".
Versions used:
* CUDA: 7.5
* clang: 3.8 and 63240447ac6e ("[analyzer] Try to fix coverity CID 1360469.")
* llvm: 3.8 and 04876e5fe54e ("[X86][XOP] Add support for combining target
shuffles to VPERMIL2PD/VPERMIL2PS")
The compilation command as well as the output look as follow (with both
previously mentioned versions of clang+LLVM):
```
% clang++ -std=c++11 --cuda-path=/opt/cuda --cuda-gpu-arch=sm_30 -c surface.cu
surface.cu:15:3: error: use of undeclared identifier 'surf2Dwrite'
surf2Dwrite(make_float4(0.0f, 0.0f, 0.0f, 1.0f), result_surface, index.x *
sizeof(float4), index.y);
^
1 error generated
```
If I include directly `surface_indirect_functions.h` which defines
`surf2Dwrite`, then the program does compile (well, sort of depending on the
clang version, but that looks like a different bug, so different bug report).
Playing with the CUDA headers, I added some `#error "foo"` inside
`device_functions.h` which is included by `cuda_runtime.h`, and in turns
includes `surface_indirect_functions.h`: so one right after the `#define
__DEVICE_FUNCTIONS_H__`, and another one (with a different message) in an else
clause to the `#ifdef __DEVICE_FUNCTIONS_H__`.
To resume, this is how the modified `device_functions.h` looks like:
```
#ifndef __DEVICE_FUNCTIONS_H__
#define __DEVICE_FUNCTIONS_H__
#error "foo"
// [snip] untouched content
#else
#error "bar"
#endif // __DEVICE_FUNCTIONS_H__
```
And here is the corresponding compilation output with clang:
```
% clang++ -std=c++11 --cuda-path=/opt/cuda --cuda-gpu-arch=sm_30 -c surface.cu
In file included from <built-in>:1:
In file included from
/home/pmoreau/privat/userspace/llvm/bin/../lib/clang/4.0.0/include/__clang_cuda_runtime_wrapper.h:98:
In file included from /opt/cuda/include/cuda_runtime.h:115:
/opt/cuda/include/device_functions.h:4247:2: error: "bar"
#error "bar"
^
In file included from <built-in>:1:
In file included from
/home/pmoreau/privat/userspace/llvm/bin/../lib/clang/4.0.0/include/__clang_cuda_runtime_wrapper.h:156:
In file included from /opt/cuda/include/math_functions.hpp:1065:
/opt/cuda/include/device_functions.h:4247:2: error: "bar"
#error "bar"
^
2 errors generated.
```
Apparently `__DEVICE_FUNCTIONS_H__` is getting defined by another file since we
do not get an `error: "foo"`.
Thank you in advance!
Pierre
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
