Undefined identifiers in all_reduce.cu

[apaszke]

Hello!

I'm trying to compile nccl, but I'm getting the following errors:

Compiling src/all_reduce.cu         > build/obj/all_reduce.o   
src/reduce_kernel.h(199): error: identifier "__half22float2" is undefined

src/reduce_kernel.h(203): error: identifier "__float22half2_rn" is undefined

src/reduce_kernel.h(214): error: identifier "__half22float2" is undefined

src/reduce_kernel.h(218): error: identifier "__float22half2_rn" is undefined

src/reduce_kernel.h(229): error: identifier "__half22float2" is undefined

src/reduce_kernel.h(233): error: identifier "__float22half2_rn" is undefined

src/reduce_kernel.h(248): error: identifier "__half22float2" is undefined

src/reduce_kernel.h(252): error: identifier "__float22half2_rn" is undefined

8 errors detected in the compilation of "/tmp/tmpxft_000004bb_00000000-13_all_reduce.compute_52.cpp1.ii".
make: *** [build/obj/all_reduce.o] Error 2

I have a TITAN X and cuda-7.5 installed. I ran make CUDA_HOME=/usr/local/cuda-7.5 test to build the library.

Do you have any idea why does it fail? I've seen that these identifiers are defined in /usr/local/cuda-7.5/includes/cuda_fp16.h, but it's not included in reduce_kernel.h. Also, they are guarded by a check __CUDA_ARCH__ >= 530, but my GPU has capability 5.2. Since TITAN is a Maxwell card, then it should be supported, right?

[nluehr]

Your build line is correct. As for cuda_fp16.h, it gets included from nccl.h (which is in turn included in core.h) whenever CUDART_VERSION >= 7050.

I'm a bit confused about the include guards you reference in reduce_kernel.h. These should select either Maxwell (CUDA_ARCH>= 500) or Kepler (CUDA_ARCH >= 300 && CUDA_ARCH < 500). There shouldn't be any reference to a CUDA_ARCH 530.

[cliffwoolley]

Do you have more than one CUDA Toolkit version installed? Any chance you
have an RC rather than production release version of CUDA 7.5?
On Jan 24, 2016 5:14 PM, "Nathan Luehr" [email protected] wrote:

Your build line is correct. As for cuda_fp16.h, it gets included from
nccl.h (which is in turn included in core.h) whenever CUDART_VERSION >=
7050.

I'm a bit confused about the include guards you reference in
reduce_kernel.h. These should select either Maxwell (CUDA_ARCH>= 500)
or Kepler (CUDA_ARCH >= 300 && CUDA_ARCH < 500). There shouldn't be
any reference to a CUDA_ARCH 530.

—
Reply to this email directly or view it on GitHub
#9 (comment).

[apaszke]

That's wierd. I think I should have a production release. When I print contents of /usr/local/cuda/version.txt and /usr/local/cuda-7.5/version they both say CUDA Version 7.5.7.

This is what I've found in cuda_fp16.h:

#if __CUDA_ARCH__ >= 530 || !defined(__CUDA_ARCH__)                                 
__CUDA_FP16_DECL__ __half2 __heq2(const __half2 a, const __half2 b);                
__CUDA_FP16_DECL__ __half2 __hne2(const __half2 a, const __half2 b);                
__CUDA_FP16_DECL__ __half2 __hle2(const __half2 a, const __half2 b);                
__CUDA_FP16_DECL__ __half2 __hge2(const __half2 a, const __half2 b);                
__CUDA_FP16_DECL__ __half2 __hlt2(const __half2 a, const __half2 b);                
__CUDA_FP16_DECL__ __half2 __hgt2(const __half2 a, const __half2 b);                
__CUDA_FP16_DECL__ __half2 __hequ2(const __half2 a, const __half2 b);               
__CUDA_FP16_DECL__ __half2 __hneu2(const __half2 a, const __half2 b);               
__CUDA_FP16_DECL__ __half2 __hleu2(const __half2 a, const __half2 b);               
__CUDA_FP16_DECL__ __half2 __hgeu2(const __half2 a, const __half2 b);               
__CUDA_FP16_DECL__ __half2 __hltu2(const __half2 a, const __half2 b);               
__CUDA_FP16_DECL__ __half2 __hgtu2(const __half2 a, const __half2 b);               
__CUDA_FP16_DECL__ __half2 __hadd2(const __half2 a, const __half2 b);               
__CUDA_FP16_DECL__ __half2 __hsub2(const __half2 a, const __half2 b);               
__CUDA_FP16_DECL__ __half2 __hmul2(const __half2 a, const __half2 b);               
__CUDA_FP16_DECL__ __half2 __hadd2_sat(const __half2 a, const __half2 b);           
__CUDA_FP16_DECL__ __half2 __hsub2_sat(const __half2 a, const __half2 b);           
__CUDA_FP16_DECL__ __half2 __hmul2_sat(const __half2 a, const __half2 b);           
__CUDA_FP16_DECL__ __half2 __hfma2(const __half2 a, const __half2 b, const __half2 c); 
__CUDA_FP16_DECL__ __half2 __hfma2_sat(const __half2 a, const __half2 b, const __half2 c); 
__CUDA_FP16_DECL__ __half __hadd(const __half a, const __half b);                   
__CUDA_FP16_DECL__ __half __hsub(const __half a, const __half b);                   
__CUDA_FP16_DECL__ __half __hmul(const __half a, const __half b);                   
__CUDA_FP16_DECL__ __half __hadd_sat(const __half a, const __half b);               
__CUDA_FP16_DECL__ __half __hsub_sat(const __half a, const __half b);               
__CUDA_FP16_DECL__ __half __hmul_sat(const __half a, const __half b);               
__CUDA_FP16_DECL__ __half __hfma(const __half a, const __half b, const __half c); 
__CUDA_FP16_DECL__ __half __hfma_sat(const __half a, const __half b, const __half c); 
__CUDA_FP16_DECL__ float __low2float(const __half2 l);                              
__CUDA_FP16_DECL__ float __high2float(const __half2 l);                             
__CUDA_FP16_DECL__ float2 __half22float2(const __half2 l);  

You can see that one of the missing identifiers in on the last line of this snippet. It only appears in a it's forward declaration and it's definition is that file.

[cliffwoolley]

7.5.7 was the release candidate build, actually. The production release
build was numbered 7.5.18.

Please redownload and reinstall CUDA 7.5; it should work with the 7.5.18
build.

Thanks!
On Jan 25, 2016 6:51 AM, "Adam Paszke" [email protected] wrote:

That's wierd. I think I should have a production release. When I print
contents of /usr/local/cuda/version.txt and /usr/local/cuda-7.5/version
they both say CUDA Version 7.5.7.

This is what I've found in cuda_fp16.h:

#if CUDA_ARCH >= 530 || !defined(CUDA_ARCH)
CUDA_FP16_DECL half2 __heq2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hne2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hle2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hge2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hlt2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hgt2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hequ2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hneu2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hleu2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hgeu2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hltu2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hgtu2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hadd2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hsub2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hmul2(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hadd2_sat(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hsub2_sat(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hmul2_sat(const __half2 a, const __half2 b);
__CUDA_FP16_DECL half2 __hfma2(const __half2 a, const __half2 b, const __half2 c);
__CUDA_FP16_DECL half2 __hfma2_sat(const __half2 a, const __half2 b, const __half2 c);
__CUDA_FP16_DECL half __hadd(const __half a, const __half b);
__CUDA_FP16_DECL half __hsub(const __half a, const __half b);
__CUDA_FP16_DECL half __hmul(const __half a, const __half b);
__CUDA_FP16_DECL half __hadd_sat(const __half a, const __half b);
__CUDA_FP16_DECL half __hsub_sat(const __half a, const __half b);
__CUDA_FP16_DECL half __hmul_sat(const __half a, const __half b);
__CUDA_FP16_DECL half __hfma(const __half a, const __half b, const __half c);
__CUDA_FP16_DECL half __hfma_sat(const __half a, const __half b, const __half c);
__CUDA_FP16_DECL float low2float(const __half2 l);
__CUDA_FP16_DECL float high2float(const __half2 l);
__CUDA_FP16_DECL float2 __half22float2(const __half2 l);

You can see that one of the missing identifiers in on the last line of
this snippet. It only appears in a it's forward declaration and it's
definition is that file.

—
Reply to this email directly or view it on GitHub
#9 (comment).

[apaszke]

I'll try it. Thanks for help! 😊

[apaszke]

Yes, that worked. Thanks a lot!