Add support for integer division to TCP #97
Conversation
| Tcp_SignednessAttr:$signedness, | ||
| Tcp_RoundingModeAttr:$rounding_mode |
There was a problem hiding this comment.
Why do we need a "Ceil" rounding mode here? Are they any cases where that is necessary?
There was a problem hiding this comment.
I don't think there's a good use-case for the tcp.diviop itself to support a ceil rounding mode, given that such a thing does not exist for the torch.div op. I added it because there is a natural mapping to the arith.ceildiviop. I can remove it.
However, I think the RoundingMode type itself should have ceil in it for completion though.
| @@ -32,4 +32,22 @@ def Tcp_Signedness : I32EnumAttr<"Signedness", | |||
|
|
|||
| def Tcp_SignednessAttr : EnumAttr<Tcp_Dialect, Tcp_Signedness, "signedness">; | |||
|
|
|||
| // TCP rounding mode | |||
| def Tcp_RoundingMode_Trunc : I32EnumAttrCase<"Trunc", 0>; | |||
| def Tcp_RoundingMode_Floor : I32EnumAttrCase<"Floor", 1>; | |||
There was a problem hiding this comment.
minor nit: fix spacing here and below
| return {b.create<arith::DivSIOp>(loc, payloadArgs[0], payloadArgs[1])}; | ||
| else if (divOp.getRoundingMode() == RoundingMode::Ceil) | ||
| return { | ||
| b.create<arith::CeilDivUIOp>(loc, payloadArgs[0], payloadArgs[1])}; |
There was a problem hiding this comment.
Shouldn't this be arith::CeilDivSIOp instead?
There was a problem hiding this comment.
yikes :( thanks for the typo fix.
| let arguments = (ins | ||
| Tcp_IntTensor:$in1, | ||
| Tcp_IntTensor:$in2, | ||
| Tcp_RoundingModeAttr:$rounding_mode |
There was a problem hiding this comment.
Do we need the rounding mode here? Should we just stick to using trunc mode for divui?
There was a problem hiding this comment.
Given that the arith dialect has straightforward mapping to ceildiv, why not just have it for the sake of expressibility? I do not have a strong opinion on this though. We could just leave it out now and introduce it later if we find a need for it.
As titled, add support for integer division in TCP. This closely mimics the existing capabilities in pytorch and arith.