[PASS] add a pass for the specific hardware accelarator when it is not binded #1999
Conversation
|
Thanks for contribution. Could you add unittest for this? |
| bounds = tvm.schedule.InferBound(s) | ||
| stmt = tvm.schedule.ScheduleOps(s, bounds) | ||
| stmt = tvm.ir_pass.DeviceMark(stmt) | ||
| print stmt |
There was a problem hiding this comment.
try to use assert instead of print, you can find some examples in other tests
|
add the unit case @yzhliu |
include/tvm/ir_pass.h
Outdated
| * \param stmt The stmt to be trasnformed | ||
| * \return Transformed stmt. | ||
| */ | ||
| Stmt DeviceMark(Stmt stmt); |
There was a problem hiding this comment.
as per our current convention, this should be Mark Device
src/pass/split_host_device.cc
Outdated
| op->attr_key == attr::pipeline_exec_scope) { | ||
| op->attr_key == attr::pipeline_exec_scope || | ||
| (op->attr_key == attr::pragma_scope_prefix && | ||
| op->value.as<StringImm>()->value == "device")) { |
There was a problem hiding this comment.
in our current convention, we use attr=pragma_xyz to indicate a special pragma, so if we want to do this, we should use pragma_device in the key and ignore the value(can be just constant 0)
|
Actually, it is a feature for some hardware accelerator which can't be binded. The pass is not essential for GPU or CPU. Using a pass will be flexible, user can add it in their own custom pass list, just like coproc_sync. And there will be no side-effect for current solution. @tqchen |
src/pass/detect_device.cc
Outdated
| namespace ir { | ||
|
|
||
| class DetectDevice : public IRMutator { | ||
| public: |
There was a problem hiding this comment.
IRMutator is not necessary, if we want to support this one, we can just inline the logic to MarkDevice
src/pass/detect_device.cc
Outdated
| } | ||
| }; | ||
|
|
||
| Stmt MarkDevice(Stmt stmt) { |
There was a problem hiding this comment.
MarkDevice->DecorateDeviceScope, which reflects what we really want
include/tvm/ir_pass.h
Outdated
| Stmt LowerStorageAccessInfo(Stmt stmt); | ||
|
|
||
| /*! | ||
| * \brief insert the mark of device for the hardware accelarator when |
There was a problem hiding this comment.
decorate the stmt with a device scope, this is helpful for hardware accelerator without thread blocks
|
@tqchen Changed per the request |
|
Thanks @kun-zh @ZihengJiang , this is merged |
4 participants
Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers.
For accelerators, host device splitting is different from gpu, as discussed here,
https://discuss.tvm.ai/t/split-we-want-to-split-the-host-and-device-code-flexibly/932
this is a solution we are using for DaVinci core, and can use for other accelerators.
@tqchen @yzhliu @tmoreau89 @ZihengJiang please review, thanks!