ROCm / hipBLASLt

hipBLASLt is a library that provides general matrix-matrix operations with a flexible API and extends functionalities beyond a traditional BLAS library
https://rocm.docs.amd.com/projects/hipBLASLt/en/latest/index.html
MIT License
63 stars 88 forks source link

Add TENSILE_ENABLE_MARKER and HIPBLASLT_ENABLE_MARKER for C API #1099

Closed KKyang closed 2 months ago

KKyang commented 2 months ago

TENSILE_ENABLE_MARKER default is off.

HIPBLASLT_ENABLE_MARKER default is on.

TENSILE_ENABLE_MARKER enables markers inside Tensile lib.

HIPBLASLT_ENABLE_MARKER enables markers for hipBLASLt API calls.

KKyang commented 2 months ago

[----------] Global test environment tear-down [==========] 48206 tests from 13 test suites ran. (3365716 ms total) [ PASSED ] 48206 tests. hipBLASLt version: 1000

KKyang commented 2 months ago

===================================== 75 passed, 31 skipped, 48 warnings in 8677.11s (2:24:37) ====================================== py3: exit 0 (8677.37 seconds) /data0/yangwen/hipBLASLt/tensilelite> py.test -v --basetemp=/tmp/.tflite/py3/tmp --junit-xml=/data0/yangwen/hipBLASLt/tensilelite/python_tests.xml --junit-prefix=py3 --color=yes -n 4 --prebuilt-client=/tmp/.tflite/py3/client/0_Build/client/tensile_client Tensile/Tests -m common pid=1089288 py3: OK (8820.89=setup[12.54]+cmd[0.71,130.26,8677.37] seconds) congratulations :) (8820.93 seconds)

KKyang commented 2 months ago

Any other library enable roctx by default?

CK?

KKyang commented 2 months ago

Comfirmed able to compile after the change

jichangjichang commented 2 months ago

Any other library enable roctx by default?

CK?

CK is not ROCM release library

KKyang commented 2 months ago

Any other library enable roctx by default?

CK?

CK is not ROCM release library

SWDEV-433409

KKyang commented 2 months ago

it is also better to have build option for disabling/enabling the roctracer related implementation.

Default on or off?

KKyang commented 2 months ago

Updated with options to turn on/off. Compile successfully.

KKyang commented 2 months ago

Change marker name "unload code object" to "UnloadCodeObjectFiles."

KKyang commented 2 months ago

Update doc and README.md

KKyang commented 2 months ago

could we use inline for markerStart() and makerStop()

Yes, but clang does not guaranteed the function will be inlined.