Add consistent NaN/Inf handling in sumsqv. #668

devinamatthews · 2022-09-23T16:14:21Z

If there is a NaN (either real or imaginary), then return a sum of NaN and unit scale.
Else, if there is an Inf (either real or imaginary), then return a sum of +Inf and unit scale.
Otherwise as normal.

- If there is a NaN (either real or imaginary), then return a sum of NaN and unit scale. - Else, if there is an Inf (either real or imaginary), then return a sum of +Inf and unit scale. - Otherwise as normal.

abs_chi1_r was reused for imaginary component.

devinamatthews · 2022-09-23T17:19:02Z

@fgvanzee after CI should be ready t merge. I tested on what should be a representative enough set of test cases to give the desired behavior.

fgvanzee · 2022-09-23T18:22:54Z

Thanks for acting on this so quickly, @devinamatthews.

EleniVlachop · 2022-09-28T14:38:00Z

frame/util/bli_util_unb_var1.c

+		{ \
+			sumsq_r = abs_chi1_r; \
+			scale_r = one_r; \
+			break; \


Assuming that in a vector x we have x[0] = inf and x[1] = NaN, then we break out of the for loop and return inf in sumsq_r. From the description in the commit message, I expected to return NaN. Could you please clarify which one is the correct return value?

This was fixed in 92b9af4.

In that case, since we do not break in case of NaN, then for x[0] = inf , x[1] = NaN, x[2]= inf, won't we overwrite sumsq_r with inf again?

There is a continue in the most recent version (here). Once a NaN is encountered this will trigger every iteration. That particular line could be a break but it doesn't make a difference to the result.

I see. Thank you very much for the clarification.

Details: - Changed sumsqv implementation as follows: - If there is a NaN (either real or imaginary), then return a sum of NaN and unit scale. - Else, if there is an Inf (either real or imaginary), then return a sum of +Inf and unit scale. - Otherwise behave as normal. (cherry picked from commit b861c71) AMD-Internal: [SWLCSG-1900] Change-Id: Ic7ba9cad1fbaf11823b9ba96e72a4ddd973db5b6

- (cherry picked from commit c803b03) Fix auto-detection of firestorm (Apple M1). - (cherry picked from commit 2dd692b) Added Discord documentation (#677) Details: - Added a docs/Discord.md markdown document that walks the reader through creating a Discord account, obtaining the invite link, and using the link to join the BLIS Discord server. - Updated README.md to reference the new Discord.md document in multiple places, including via the official Discord logo (used with explicit permission from representatives at Discord Inc.). - (cherry picked from commit 88105db) Shuffled checked properties in bli_l3_check.c. (#676) Details: - Added certain checks for matrix structure to the level-3 operations' _check() functions, and slightly reorganized existing checks. - (cherry picked from commit 23f5b8d) CREDITS file update. Details: - This attribution was intended to go in PR #647. - (cherry picked from commit 9453e0f) Reinstate sanity check in bli_pool_finalize. (#671) Details: - Added a reinit argument to bli_pool_finalize(). This bool will signal whether or not the function is being called from bli_pool_reinit(). If it is not being called from _reinit(), we can safely check to confirm that .top_index == 0 (i.e., all blocks have been checked in). But if it *is* being called from _reinit(), then that check will be skipped since one of the predicted use cases for bli_pool_reinit() anticipates that some blocks are (probably) checked out when the pool_t is reinitialized. - Updated existing invocations of bli_pool_finalize() to pass in either FALSE (from bli_apool_free_block() or bli_pba_finalize_pools()) or TRUE (from bli_pool_reinit()) for the new reinit argument. - (cherry picked from commit 76a23bd) Fix some bugs in bli_pool.c (#670) Details: - Add a check for premature pool exhaustion when checking in blocks via bli_pool_checkin_block(). This detects "double-free" and other bad conditions that don't necessarily result in a segfault. - Make sure to copy all block pointers when growing the pool size. Previously, checked-out block pointers (which are guaranteed to be set to NULL) were not being copied, leading to the presence of uninitialized data. - (cherry picked from commit 63470b4) Add AddressSanitizer (-fsanitize=address) option. (#669) Details: - Added support for AddressSanitizer (ASan), a compiler-integrated memory error detector. The option (disabled by default) enables compiling and linking with the -fsanitize=address flag supported by clang, gcc, and probably others. This flag is employed during compilation of all BLIS source files *except* for optimized kernels, which are exempted because ASan usually requires an extra register, which violates the constraints for many gemm microkernels. - Minor whitespace, comment, ordering, and configure help text updates. - (cherry picked from commit 42d0e66) Add consistent NaN/Inf handling in sumsqv. (#668) Details: - Changed sumsqv implementation as follows: - If there is a NaN (either real or imaginary), then return a sum of NaN and unit scale. - Else, if there is an Inf (either real or imaginary), then return a sum of +Inf and unit scale. - Otherwise behave as normal. - (cherry picked from commit b861c71) Parameterized test/3 drivers via command line args. (#667) Details: - Rewrote the drivers in test/3, the Makefile, and the runme.sh script so that most of the important parameters, including parameter combo, datatype, storage combo, induced method, problem size range, dimension bindings, number of repeats, and alpha/beta values can be passed in via command line arguments. (Previously, most of these parameters were hard-coded into the driver source, except a few that were hard-coded into the Makefile.) If no argument is given for any particular option, it will be assigned a sane default. Either way, the values employed at runtime will be printed to stdout before the performance data in a section that is commented out with '%' characters (which is used by matlab and octave for comments), unless the -q option is given, in which case the driver will proceed quietly and output only performance data. Each driver also provides extensive help via the -h option, with the help text tailored for the operation in question (e.g. gemm, hemm, herk, etc.). In this help text, the driver reminds the user which implementation it was linked to (e.g. blis, openblas, vendor, eigen). Thanks to Jeff Diamond for suggesting this CLI-based reimagining of the test/3 drivers. - In the test/3 drivers: converted cpp macro string constants, as well as two string literals (for the opname and pc_str) used in each test driver, to global (or static) const char* strings, and replaced the use of strncpy() for storing the results of the command line argument parsing with pointer copies from the corresponding strings in argv. This works because the argv array is guaranteed by the C99 standard to persist throughout the life of the program. This new approach uses less storage and executes faster. Thanks to Minh Quan Ho for recommending this change. - Renamed the IMP_STR cpp macro that gets defined on the command line, via the test/3/Makefile, to IMPL_STR. - Updated runme.sh to set the problem size ranges for single-threaded and multithreaded execution independently from one another, as well as on a per-system basis. - Added a 'quiet' variable to runme.sh that can easily toggle quiet mode for the test drivers' output. - Very minor typecast fix in call to bli_getopt() in bli_utils.c. - In bli_getopt(), changed the nextchar variable from being a local static variable to a field of the getopt_t state struct. (Not sure why it was ever declared static to begin with.) - Other minor changes to bli_getopt() to accommodate the rewritten test drivers' command line parsing needs. - (cherry picked from commit ee81efc)

- (cherry picked from commit c803b03) Fix auto-detection of firestorm (Apple M1). - (cherry picked from commit 2dd692b) Added Discord documentation (#677) Details: - Added a docs/Discord.md markdown document that walks the reader through creating a Discord account, obtaining the invite link, and using the link to join the BLIS Discord server. - Updated README.md to reference the new Discord.md document in multiple places, including via the official Discord logo (used with explicit permission from representatives at Discord Inc.). - (cherry picked from commit 88105db) Shuffled checked properties in bli_l3_check.c. (#676) Details: - Added certain checks for matrix structure to the level-3 operations' _check() functions, and slightly reorganized existing checks. - (cherry picked from commit 23f5b8d) CREDITS file update. Details: - This attribution was intended to go in PR #647. - (cherry picked from commit 9453e0f) Reinstate sanity check in bli_pool_finalize. (#671) Details: - Added a reinit argument to bli_pool_finalize(). This bool will signal whether or not the function is being called from bli_pool_reinit(). If it is not being called from _reinit(), we can safely check to confirm that .top_index == 0 (i.e., all blocks have been checked in). But if it *is* being called from _reinit(), then that check will be skipped since one of the predicted use cases for bli_pool_reinit() anticipates that some blocks are (probably) checked out when the pool_t is reinitialized. - Updated existing invocations of bli_pool_finalize() to pass in either FALSE (from bli_apool_free_block() or bli_pba_finalize_pools()) or TRUE (from bli_pool_reinit()) for the new reinit argument. - (cherry picked from commit 76a23bd) Fix some bugs in bli_pool.c (#670) Details: - Add a check for premature pool exhaustion when checking in blocks via bli_pool_checkin_block(). This detects "double-free" and other bad conditions that don't necessarily result in a segfault. - Make sure to copy all block pointers when growing the pool size. Previously, checked-out block pointers (which are guaranteed to be set to NULL) were not being copied, leading to the presence of uninitialized data. - (cherry picked from commit 63470b4) Add AddressSanitizer (-fsanitize=address) option. (#669) Details: - Added support for AddressSanitizer (ASan), a compiler-integrated memory error detector. The option (disabled by default) enables compiling and linking with the -fsanitize=address flag supported by clang, gcc, and probably others. This flag is employed during compilation of all BLIS source files *except* for optimized kernels, which are exempted because ASan usually requires an extra register, which violates the constraints for many gemm microkernels. - Minor whitespace, comment, ordering, and configure help text updates. - (cherry picked from commit 42d0e66) Add consistent NaN/Inf handling in sumsqv. (#668) Details: - Changed sumsqv implementation as follows: - If there is a NaN (either real or imaginary), then return a sum of NaN and unit scale. - Else, if there is an Inf (either real or imaginary), then return a sum of +Inf and unit scale. - Otherwise behave as normal. - (cherry picked from commit b861c71) Parameterized test/3 drivers via command line args. (#667) Details: - Rewrote the drivers in test/3, the Makefile, and the runme.sh script so that most of the important parameters, including parameter combo, datatype, storage combo, induced method, problem size range, dimension bindings, number of repeats, and alpha/beta values can be passed in via command line arguments. (Previously, most of these parameters were hard-coded into the driver source, except a few that were hard-coded into the Makefile.) If no argument is given for any particular option, it will be assigned a sane default. Either way, the values employed at runtime will be printed to stdout before the performance data in a section that is commented out with '%' characters (which is used by matlab and octave for comments), unless the -q option is given, in which case the driver will proceed quietly and output only performance data. Each driver also provides extensive help via the -h option, with the help text tailored for the operation in question (e.g. gemm, hemm, herk, etc.). In this help text, the driver reminds the user which implementation it was linked to (e.g. blis, openblas, vendor, eigen). Thanks to Jeff Diamond for suggesting this CLI-based reimagining of the test/3 drivers. - In the test/3 drivers: converted cpp macro string constants, as well as two string literals (for the opname and pc_str) used in each test driver, to global (or static) const char* strings, and replaced the use of strncpy() for storing the results of the command line argument parsing with pointer copies from the corresponding strings in argv. This works because the argv array is guaranteed by the C99 standard to persist throughout the life of the program. This new approach uses less storage and executes faster. Thanks to Minh Quan Ho for recommending this change. - Renamed the IMP_STR cpp macro that gets defined on the command line, via the test/3/Makefile, to IMPL_STR. - Updated runme.sh to set the problem size ranges for single-threaded and multithreaded execution independently from one another, as well as on a per-system basis. - Added a 'quiet' variable to runme.sh that can easily toggle quiet mode for the test drivers' output. - Very minor typecast fix in call to bli_getopt() in bli_utils.c. - In bli_getopt(), changed the nextchar variable from being a local static variable to a field of the getopt_t state struct. (Not sure why it was ever declared static to begin with.) - Other minor changes to bli_getopt() to accommodate the rewritten test drivers' command line parsing needs. - (cherry picked from commit ee81efc) Allow test/3 drivers to use default ind_t method. (#804) Details: - Previously, the standalone performance drivers in test/3 were written under the assumption that the user would want to explicitly test either native execution *or* 1m. But because the accompanying runme.sh script defaults to passing "native" in for the -i command line option (which explicitly sets the induced method type), running the script without modification causes the test drivers to use slow reference microkernels on systems where native complex-domain microkernels are not registered -- which will yield poor performance for complex-domain level-3 operations. Furthermore, even if a user was aware of this, the test drivers did not support any single value for the -i option that would test BLIS using the library's default behavior -- that is, using 1m on systems where it is needed and native execution on systems that have native microkernels implemented and registered. - This commit addresses the aforementioned issue by supporting a new value for the -i option: "auto". The "auto" value causes the driver to avoid explicitly setting the induced method altogether, leaving BLIS's default behavior in place. This "auto" option is also now the default setting within the runme.sh script. Thanks to Leick Robinson for finding and reporting this issue. - Also added support for "nat" as a shorthand for "native", which the help text already (erroneously) claimed was supported. - (cherry picked from commit fd1a7e3) Use "-i auto" by default in test/3 drivers. Details: - Request default induced method behavior of BLIS via "-i auto" when running the standalone performance drivers in test/3 via the runme.sh script present in that directory. (Previously, the runme.sh script would use "-i native" by default.) This change was originally intended for fd1a7e3. - (cherry picked from commit cad5149)

devinamatthews added 3 commits September 23, 2022 11:14

Add consistent NaN/Inf handling in sumsqv.

18b0abb

- If there is a NaN (either real or imaginary), then return a sum of NaN and unit scale. - Else, if there is an Inf (either real or imaginary), then return a sum of +Inf and unit scale. - Otherwise as normal.

Added abs_chi1_i

68cff8f

abs_chi1_r was reused for imaginary component.

Inf before NaN wasn't correct.

92b9af4

Whitespace changes.

97aeea5

fgvanzee merged commit b861c71 into master Sep 23, 2022

devinamatthews deleted the nrm2-inf-nan branch September 23, 2022 19:45

EleniVlachop reviewed Sep 28, 2022

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add consistent NaN/Inf handling in sumsqv. #668

Add consistent NaN/Inf handling in sumsqv. #668

devinamatthews commented Sep 23, 2022

devinamatthews commented Sep 23, 2022

fgvanzee commented Sep 23, 2022

EleniVlachop Sep 28, 2022

devinamatthews Sep 28, 2022

EleniVlachop Sep 28, 2022

devinamatthews Sep 28, 2022

EleniVlachop Sep 28, 2022

Add consistent NaN/Inf handling in sumsqv. #668

Add consistent NaN/Inf handling in sumsqv. #668

Conversation

devinamatthews commented Sep 23, 2022

devinamatthews commented Sep 23, 2022

fgvanzee commented Sep 23, 2022

EleniVlachop Sep 28, 2022

Choose a reason for hiding this comment

devinamatthews Sep 28, 2022

Choose a reason for hiding this comment

EleniVlachop Sep 28, 2022

Choose a reason for hiding this comment

devinamatthews Sep 28, 2022

Choose a reason for hiding this comment

EleniVlachop Sep 28, 2022

Choose a reason for hiding this comment