CUDA 9 and sm_20

[xilman]

The latest CUDA compiler no longer supports the cc2.x architecture. This has caused a number of problems and has now hit msieve's Makefile.

Anyone want to decide what to do and update msieve appropriately?

I'd do it but there are doubtless folks here more skilled than I am. For a start, I have no capability to build under Windoze.

FWIW, GMP-ECM has hit the same issue. I'm responsible for the CUDA build there but we explicitly don't support native Windoze builds and so the problem can be safely ignored for the time being.

Added in edit: I really don't want to do it. An attempt to kludge my way through led me in to a twisty little maze of nvcc options all different.

[xilman]

It gets worse:

pcl@anubis ~/nums/msieve-code/trunk $ ./msieve -i W852.n -g 0 -t 6 -np
integrator failed nan inf

Oh well, lets see whether CADO-NFS still works.

Nope it didn't.

I can see the New Year's resolution is to get the old tool kits working again. Ho hum.

[kruoli]

This message occured with "recent" CADO-NFS 2.3.0 with CUDA. For reference, since this was the only exact search result for that exact message I got on Google, I got it working by removing every temporary or intermediate file in the executable's directory. Before that, it crashed with the exact same message.

[kruoli]

Originally Posted by kruoli

... CADO-NFS 2.3.0 ...

Sleep more, think more...
I have no idea why I was thinking about CADO, this is of course totally wrong. Sorry for the misinformation. I was using msieve, SVN 998, and had the problem.

[chris2be8]

The last time I had to install msieve on a system with a CUDA capable GPU I had to update the Makefiles as follows:

chris@sirius:~/factordb$ diff -u /home/chris/msieve.1030/trunk/Makefile /home/chris/msieve.1030.cuda/trunk/Makefile 
--- /home/chris/msieve.1030/trunk/Makefile	2019-12-07 19:15:56.039498635 +0000
+++ /home/chris/msieve.1030.cuda/trunk/Makefile	2019-12-07 20:01:18.819251935 +0000
@@ -196,8 +196,8 @@
 
 #---------------------------------- GPU file lists -------------------------
 
+#	stage1_core_sm20.ptx 
 GPU_OBJS += \
-	stage1_core_sm20.ptx \
 	stage1_core_sm30.ptx \
 	stage1_core_sm35.ptx \
 	stage1_core_sm50.ptx \
@@ -334,8 +334,8 @@
 
 # GPU build rules
 
-stage1_core_sm20.ptx: $(NFS_GPU_HDR)
-	$(NVCC) -arch sm_20 -ptx -o $@ $<
+#stage1_core_sm20.ptx: $(NFS_GPU_HDR)
+#	$(NVCC) -arch sm_20 -ptx -o $@ $<
 
 stage1_core_sm30.ptx: $(NFS_GPU_HDR)
 	$(NVCC) -arch sm_30 -ptx -o $@ $<
@@ -347,4 +347,4 @@
 	$(NVCC) -arch sm_50 -ptx -o $@ $<
 
 cub/built:
-	cd cub && make WIN=$(WIN) WIN64=$(WIN64) sm=200,300,350,520 && cd ..
+	cd cub && make WIN=$(WIN) WIN64=$(WIN64) sm=300,350,520 && cd ..

chris@sirius:~/factordb$ diff -u /home/chris/msieve.1030/trunk/cub/Makefile /home/chris/msieve.1030.cuda/trunk/cub/Makefile 
--- /home/chris/msieve.1030/trunk/cub/Makefile	2019-12-07 19:16:01.763534223 +0000
+++ /home/chris/msieve.1030.cuda/trunk/cub/Makefile	2019-12-09 21:40:42.328453882 +0000
@@ -16,7 +16,7 @@
 	NVCC = "$(shell which nvcc)"
 	CUDA_ROOT = $(shell dirname $(NVCC))/../
 	EXT = so
-	NVCCFLAGS += -Xptxas -v -Xcudafe -\# -shared -Xptxas -abi=no \
+	NVCCFLAGS += -Xptxas -v -Xcudafe -\# -shared \
 			-Xcompiler -fPIC -Xcompiler -fvisibility=hidden
 endif
 
@@ -43,14 +43,6 @@
     SM_TARGETS 	+= -gencode=arch=compute_30,code=\"sm_30,compute_30\"
     SM_DEF 		+= -DSM300
 endif
-ifeq (210, $(findstring 210, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_20,code=\"sm_21,compute_20\"
-    SM_DEF 		+= -DSM210
-endif
-ifeq (200, $(findstring 200, $(SM_ARCH)))
-    SM_TARGETS 	+= -gencode=arch=compute_20,code=\"sm_20,compute_20\"
-    SM_DEF 		+= -DSM200
-endif
 
 rwildcard=$(foreach d,$(wildcard $1*),$(call rwildcard,$d/,$2) $(filter $(subst *,%,$2),$d))

After that it worked for me.

Chris

[Brian Gladman]

Originally Posted by xilman

The latest CUDA compiler no longer supports the cc2.x architecture. This has caused a number of problems and has now hit msieve's Makefile.

Anyone want to decide what to do and update msieve appropriately?

I'd do it but there are doubtless folks here more skilled than I am. For a start, I have no capability to build under Windoze.

FWIW, GMP-ECM has hit the same issue. I'm responsible for the CUDA build there but we explicitly don't support native Windoze builds and so the problem can be safely ignored for the time being.

Added in edit: I really don't want to do it. An attempt to kludge my way through led me in to a twisty little maze of nvcc options all different.

GMP-ECM explicitly does support native Windows builds and has done so for many years (I maintain this part of the build which uses Microsoft Visual Studio). I also fully support the CUDA build of MSIEVE on Windows.

[xilman]

Originally Posted by Brian Gladman

GMP-ECM explicitly does support native Windows builds and has done so for many years (I maintain this part of the build which uses Microsoft Visual Studio). I also fully support the CUDA build of MSIEVE on Windows.

Excellent!

Do you have fully static builds available? As I can't get anything to work natively it might be worth trying to run a Windows executable under WINE.

[Brian Gladman]

Originally Posted by xilman

Excellent!

Do you have fully static builds available? As I can't get anything to work natively it might be worth trying to run a Windows executable under WINE.

HI Paul,

The builds in the official repositories for GMP-ECM and MSIEVE are both static. If you don't know the repository locations I can let you have them (I am not sure about posting their locations here).

[firejuggler]

I would love a ptx compatible with a compute ability 7.5....

[Happy5214]

Originally Posted by firejuggler

I would love a ptx compatible with a compute ability 7.5....

Seconded. Attempting to build msieve and GMP-ECM on my brand-new Kubuntu Focus (RTX 2060) was very frustrating, and I still don't think I did it right.

[Brian Gladman]

Originally Posted by Happy5214

Seconded. Attempting to build msieve and GMP-ECM on my brand-new Kubuntu Focus (RTX 2060) was very frustrating, and I still don't think I did it right.

Anyone who is having problems in building Windows x64 CUDA versions of MSIEVE or GMP-ECM using Visual Studio 2019 is welcome to report their issues here and I will do what I can to help.