UCX fails to build with Intel compilers in optimized mode
Closed this issue · 8 comments
Recent Intel compilers (I checked starting from 2023.2) automatically enable fast-math mode for optimized builds. This mode allows optimizations for floating-point arithmetic that assume that arguments and results are not NaNs or +-Infs. While most compilers will simply translate isnan and isinf to always false, recent Intel compilers took a different approach resulting in the following errors while building UCX in optimized mode.
.../ucx-master/src/ucs/datastruct/linear_func.h:171:9: error: explicit comparison with NaN in fast floating point mode [-Werror,-Wtautological-constant-compare]
if (isnan(x) || isinf(x)) {
^~~~~~~~
.../ucx-master/src/ucs/datastruct/linear_func.h:171:21: error: explicit comparison with infinity in fast floating point mode [-Werror,-Wtautological-constant-compare]
if (isnan(x) || isinf(x)) {
A quick hack is to force Intel compilers to switch in precise floating point mode for the incriminating function (ucs_linear_func_compose
).
diff --git a/src/ucs/datastruct/linear_func.h b/src/ucs/datastruct/linear_func.h
index 31e168df1..0808f775e 100644
--- a/src/ucs/datastruct/linear_func.h
+++ b/src/ucs/datastruct/linear_func.h
@@ -157,6 +157,7 @@ ucs_linear_func_compose(ucs_linear_func_t outer, ucs_linear_func_t inner)
* double value.
*
*/
+#pragma float_control(precise,on)
static inline ucs_status_t
ucs_linear_func_intersect(ucs_linear_func_t func1, ucs_linear_func_t func2,
double *x_intersect)
Unfortunately, the proposed "solution" only works for the Intel compilers, all other compilers will choke with an unknown-pragma error. I fiddled a little around and it seems the culprit is the extremely restrictive -Werror flag added by configure. Without it I was able to successfully compile in all cases.
Can you please try to partially disable fast-math optimization with -fno-finite-math-only
? Else we could wrap isnan()/isinf() with __FINITE_MATH_ONLY__
being false, but seems that ICX does not use that define, do you find anything related that would be set by ICX?
Adding CFLAGS=-fno-finite-math-only
solves the problem.
Few additional issues:
- ADD_COMPILER_FLAG_IF_SUPPORTED add the flags to all compilers, which requires the C and CXX compilers to be of the same type. I confirm that building with icx/g++ does not work.
- the
-fopenmp
OpenMP flag added by configure is not supported by icx. Instead it asks for-qopenmp
. Again, this is triggered by the use of-Werror
, if I remote it everything goes smoothly.
Could you please try #9893 using build instructions in description? For openmp, how are you configuring UCX?
- Using a slightly older version of the Intel compiler 2021.10.0 (gcc version 11.4.1 compatibility), everything works (including the OpenMP support) without the patch.
configure: C compiler: icc -O3 -g -Wall -Werror -diag-error 10006 -diag-error 10148 -diag-disable 269 -Wno-missing-field-initializers -Wno-unused-parameter -Wno-long-long -Wno-endif-labels -Wno-sign-compare -Wno-multichar -Wno-deprecated-declarations -Winvalid-pch -fno-finite-math-only -Wno-pointer-sign -Werror-implicit-function-declaration -Wshadow -Werror=declaration-after-statement
configure: C++ compiler: icpc -O3 -g -Wall -Werror -diag-error 10006 -diag-error 10148 -diag-disable 269 -Wno-missing-field-initializers -Wno-unused-parameter -Wno-long-long -Wno-endif-labels -Wno-sign-compare -Wno-multichar -Wno-deprecated-declarations -Winvalid-pch -fno-finite-math-only
- Using Intel compilers 2024.0.2, with
**--disable-openmp** --disable-logging --disable-params-check
and usingCC=icx-cc CXX=icx
(the old icc/icpc compilers names were removed) with your patch works.
configure: C compiler: icx-cc -O3 -g -Wall -Werror -diag-error 10006 -diag-error 10148 -diag-disable 269 -fmax-type-align=16 -funwind-tables -Wno-missing-field-initializers -Wno-unused-parameter -Wno-unused-label -Wno-long-long -Wno-endif-labels -Wno-sign-compare -Wno-multichar -Wno-deprecated-declarations -Winvalid-pch -Wno-language-extension-token -fno-finite-math-only -Wno-c99-extensions -Wno-pointer-sign -Werror-implicit-function-declaration -Wno-format-zero-length -Wnested-externs -Wshadow -Werror=declaration-after-statement
configure: C++ compiler: icx -O3 -g -Wall -Werror -diag-error 10006 -diag-error 10148 -diag-disable 269 -fmax-type-align=16 -funwind-tables -Wno-missing-field-initializers -Wno-unused-parameter -Wno-unused-label -Wno-long-long -Wno-endif-labels -Wno-sign-compare -Wno-multichar -Wno-deprecated-declarations -Winvalid-pch -Wno-language-extension-token -fno-finite-math-only -Wno-c99-extensions
- Using Intel compilers 2024.0.2, with
--disable-logging --disable-params-check
and usingCC=icx-cc CXX=icx
(I just removed the--disable-openmp
) with your patch does not work.
configure: C compiler: icx-cc -O3 -g -Wall -Werror -diag-error 10006 -diag-error 10148 -diag-disable 269 -fmax-type-align=16 -funwind-tables -Wno-missing-field-initializers -Wno-unused-parameter -Wno-unused-label -Wno-long-long -Wno-endif-labels -Wno-sign-compare -Wno-multichar -Wno-deprecated-declarations -Winvalid-pch -Wno-language-extension-token -fno-finite-math-only -Wno-c99-extensions -Wno-pointer-sign -Werror-implicit-function-declaration -Wno-format-zero-length -Wnested-externs -Wshadow -Werror=declaration-after-statement
configure: C++ compiler: icx -O3 -g -Wall -Werror -diag-error 10006 -diag-error 10148 -diag-disable 269 -fmax-type-align=16 -funwind-tables -Wno-missing-field-initializers -Wno-unused-parameter -Wno-unused-label -Wno-long-long -Wno-endif-labels -Wno-sign-compare -Wno-multichar -Wno-deprecated-declarations -Winvalid-pch -Wno-language-extension-token -fno-finite-math-only -Wno-c99-extensions
The error message is:
icx: error: icx: Use of '-qopenmp' recommended over '-fopenmp' [-Werror,-Wrecommended-option]
- I gave it what it wants and set
CFLAGS="-Wno-recommended-option"
before configure. With this and Intel compilers 2024.0.2 and your patch everything WORKS.
thanks, updated PR with -Wno-recommended-option
, could you please double check?
I confirm the new patch works as expected.
fix merged