Add threshold for ops using openmp macro #5584
Conversation
|
@pytorchbot test this please |
aten/src/TH/generic/THTensorMath.c
Outdated
| #endif | ||
|
|
||
| #ifdef _OPENMP | ||
| LAB_IMPLEMENT_BASIC_FUNCTION(log,TH_MATH_NAME(log),HYPER_TH_OMP_OVERHEAD_THRESHOLD) |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
|
CC @cpuhrsch |
|
Thanks @ezyang. I'm not sure what the philosophy here is, but CC @colesbury should know more. |
|
@ezyang @colesbury Could you spare some time to review this PR? I want to optimize nn module based on these macros. |
ezyang
left a comment
ezyang
left a comment
There was a problem hiding this comment.
I'm OK with merging this but I don't regularly interact with this code.
|
@pytorchbot retest this please |
|
@pytorchbot retest this please |
|
@fmassa @zdevito @vedanuj , some changes in this PR is related to #5010. A parameter named VECTORIZE is added to determine whether to use THVector_(NAME) implementation or not. This can avoid defining a new macro just for sigmoid. |
|
@pytorchbot test this please |
| } else {\ | ||
| PRAGMA(simd) \ | ||
| PRAGMA( omp parallel for if (SIZE > TH_OMP_OVERHEAD_THRESHOLD_OMP) ) \ | ||
| PRAGMA( omp parallel for if (SIZE > OMP_THRESHOLD * 10) ) \ |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
|
|
||
| #define HYPER_TH_OMP_OVERHEAD_THRESHOLD 2000 | ||
| #define ORDIN_TH_OMP_OVERHEAD_THRESHOLD 20000 | ||
| #define UNCERTAIN_TH_OMP_OVERHEAD_THRESHOLD 50000 |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
Modify macro LAB_IMPLEMENT_VECTORIZED_FUNCTION to enable optional parameters This reverts commit 8ef783a. Conflicts: aten/src/TH/generic/THTensorMath.c
|
@pytorchbot test this please |
zy97140
requested review from
colesbury,
gchanan,
soumith and
zdevito
as code owners
|
@ezyang Could you spare some time to review this PR? I want to optimize nn module based on these macros. |
* add threshold for ops using omp macro * modify interface for ops using omp macro * modify some thresholds * implement C macros with optional parameters to avoid duplicating definitions for all pointwise operations * add a parameter of LAB_IMPLEMENT_BASIC_FUNCTION for vectorizing * modify the comment * Revert "add a parameter of LAB_IMPLEMENT_BASIC_FUNCTION for vectorizing" Modify macro LAB_IMPLEMENT_VECTORIZED_FUNCTION to enable optional parameters This reverts commit 8ef783a. Conflicts: aten/src/TH/generic/THTensorMath.c * fix build error on windows * retrigger the test
* add threshold for ops using omp macro * modify interface for ops using omp macro * modify some thresholds * implement C macros with optional parameters to avoid duplicating definitions for all pointwise operations * add a parameter of LAB_IMPLEMENT_BASIC_FUNCTION for vectorizing * modify the comment * Revert "add a parameter of LAB_IMPLEMENT_BASIC_FUNCTION for vectorizing" Modify macro LAB_IMPLEMENT_VECTORIZED_FUNCTION to enable optional parameters This reverts commit 8ef783a. Conflicts: aten/src/TH/generic/THTensorMath.c * fix build error on windows * retrigger the test
6 participants
As described in issue #4188 and benchmark, optimal OpenMP threshold is dependent on operation type, CPU type and tensors' contiguity.
This PR add a parameter named OMP_THRESHOLD for macros TH_TENSOR_APPLYX_OMP to control the OpenMP threshold.