Add controlcode to transaction lowering. #840
Conversation
jtuyls
requested review from
makslevental,
MaheshRavishankar,
nirvedhmeshram,
yzhang93 and
Abhishek-Varma
as code owners
jtuyls
changed the title
jtuyls
force-pushed
the
controlcode-to-transaction
branch
4 times, most recently
from
a8c6a57
to
32bac9b
Compare
jtuyls
changed the title
| @@ -6,15 +6,14 @@ | |||
|
|
|||
| #include "iree-amd-aie/IR/AMDAIEOps.h" | |||
|
|
|||
| #include <numeric> | |||
There was a problem hiding this comment.
This is used for std::accumulate.
| @@ -159,6 +159,10 @@ AMDAIEDeviceModel::AMDAIEDeviceModel( | |||
| TRY_XAIE_API_FATAL_ERROR(XAie_TurnEccOff, &devInst); | |||
| } | |||
|
|
|||
| uint8_t AMDAIEDeviceModel::getAddressGranularityInBits() const { | |||
There was a problem hiding this comment.
bruh lol - this is the goofiest name for MinStrideBitWidth.
There was a problem hiding this comment.
Yeah, I just named this in consistency with aie-rt/mlir-aie I believe, but yeah, it's goofy. Your suggestion is way better.. Updated it.
| @@ -222,6 +222,10 @@ struct AMDAIEDeviceModel { | |||
| /// aie-rt for whatever reason. Make sure the parameters can't be retrieved in | |||
| /// another way before adding new fields to this struct. | |||
| struct AMDAIEDeviceConfig { | |||
| /// Set default addressing granularity to 32 bits as this is the value for | |||
| /// all current architecture versions. | |||
| uint8_t addressGranularityInBits{32}; | |||
There was a problem hiding this comment.
yea so this is unncessary today right? is the idea to plan ahead? i don't remember - is it shorter in aie4 or no?
There was a problem hiding this comment.
Yeah, it's mainly to avoid hardcoding this value everywhere else, or hide it inside the getter. I am trying to put all configs that can't be retrieved from aie-rt in this place.
| let description = [{ | ||
| This NPU controller operation patches the address inside the buffer | ||
| descriptor with provided ID on the specified column. This enables codegen to | ||
| provide an argument index and offset at compile time, which is then |
There was a problem hiding this comment.
sorry I haven't been paying attention: arg index and offset to what? hal.buffer bindings or something?
There was a problem hiding this comment.
Arg index is indeed the HAL buffer bindings and this way the firmware knows the index of the buffer to use, so it can retrieve the address. The offset that is provided will be added to that address.
| `(` | ||
| $tile `,` | ||
| $value `,` | ||
| `port_type` `=` $port_type `,` |
There was a problem hiding this comment.
Ideally this would be strm_sw_port_type to match StrmSwPortType in aie-rt? Is that too long you think?
There was a problem hiding this comment.
I think it's a bit long yeah, and also quite low level. Most people reading strm_sw_port_type will need to look up what it means anyway.
| let extraClassDeclaration = [{ | ||
| /// Return the number of leading operands before the `offsets`, `sizes` and | ||
| /// and `strides` operands. | ||
| static unsigned getOffsetSizeAndStrideStartOperandIndex() { return 2; } |
There was a problem hiding this comment.
This needs to be implemented for the OffsetSizeAndStrideOpInterface.
| ArrayRef<uint32_t> instructions = | ||
| transactionBuilder.finalizeAndReturnInstructions(); | ||
| workgroupOp->setAttr( | ||
| "npu_instructions", |
There was a problem hiding this comment.
can you do me a favor - can you make this a real attr on that op? so that we're not doing string look-ups/matches/etc.
| amdaie.workgroup { | ||
| amdaie.controlcode { | ||
| amdaie.npu.address_patch {arg_idx = 0 : ui32, bd_id = 0 : ui32, col = 0 : ui32, offset = 0 : ui32} | ||
| amdaie.end |
There was a problem hiding this comment.
i know you're not defining amdaie.controlcode in this PR but does this op really need a terminator? it doesn't have any control flow right? just "straight line"?
There was a problem hiding this comment.
Tbh, I created this to replace the func.func as used in aie.device and didn't think much about whether it needs a terminator.
it doesn't have any control flow right?
No, it doesn't at this time, an isn't supported in the FW, but maybe in the (far) future?
|
After you land this I think I'll take a break next week and port all this stuff to use the actual txn stuff in aie-rt. |
jtuyls
force-pushed
the
controlcode-to-transaction
branch
from
32bac9b
to
4a7318b
Compare
Ports the
AMDAIEDmaToNpuPasswhich operates on AIE dialect toAMDAIEdialect.NOTE: The
AMDAIEDmaToNpuPassis kept around for the AIR flow.