support PTX loading on CUDA 7.5

This is a pretty huge hack right now. It basically lies about the nvcc compiler version if you compile with CUDA 8. However, it works. More investigation is warranted.

Author: unixpickle

Makefile

@@ -5,7 +5,7 @@ kernels32.go: kernels32.cu
 	echo 'package cudavec' >$@
 	echo '' >>$@
 	echo 'var kernels32PTX = `' >>$@
-	cat kernels32.ptx >>$@
+	cat kernels32.ptx | sed -E 's/.version 5\../.version 4.3/' >>$@
 	echo '`' >>$@
 	rm kernels32.ptx
 

kernels32.go

@@ -9,7 +9,7 @@ var kernels32PTX = `
 // Based on LLVM 3.4svn
 //
 
-.version 5.0
+.version 4.3
 .target sm_30
 .address_size 64