Change offset, len syntax in dma_bd
#1137
Conversation
Co-authored-by: Ronan Keryell <ronan.keryell@amd.com>
makslevental
changed the title
offset, len syntax in dma_bd
Coverage ReportCreated: 2024-03-19 14:30Click here for information about interpreting this report.
Generated by llvm-cov -- llvm version 14.0.0 |
There was a problem hiding this comment.
I think GitHub does not let me add comments to lines you didn't change, but line 833 needs to be changed to use the new syntax for dims:
aie.dma_bd(%buf : memref<128xi32>, 0, 128, [<8, 16>, <2, 1>, <8, 2>])
There was a problem hiding this comment.
This part of the PR, which is only about moving offset and len out into the attr-dict, hasn't been decided on yet (and I haven't rebased either after the first part landed). But I'll try to remember that if it gets green-lit.
| @@ -835,13 +863,21 @@ def AIE_DMABDOp: AIE_Op<"dma_bd", []> { | |||
| let hasVerifier = 1; | |||
|
|
|||
| let assemblyFormat = [{ | |||
| `(` $buffer `:` type($buffer) (`,` $offset^)? (`,` $len^)? (`,` $dimensions^)? `)` attr-dict | |||
| `(` $buffer `:` type($buffer) (`,` `dims` `=` $dims^)? `)` attr-dict | |||
There was a problem hiding this comment.
Do we still need dims in assemblyFormat? You also add "dims" as an attribute on line 894 below, does that make it part of attr-dict?
There was a problem hiding this comment.
All attributes are always in the attribute dictionary (on the C++ side) but the reason I put dims here is because if you put them into attr-dict (i.e., tell the parser to parse them from there) then you have to annotate each size and stride with their type (size: i32, etc) where here you don't need to. You can do some extra work to actually get i32s automatically "inferred" even in the attr-dict (and I tried) but I just don't want to do all of that.
| } else { | ||
| stride = dims.value()[i].getStride(); | ||
| size = static_cast<uint16_t>(dims.value()[i].getSize() * | ||
| elementWidthIn32bWords); |
There was a problem hiding this comment.
It appears to me from this line we are making the strides be multiples of the memref data type, where previously it was always in 4-byte increments. (Correct me if wrong.)
I'm assuming you all discussed this and agreed to make this change? The reason I bring this up is because we had a lively discussion about this back when I had my original PR for this feature, see here.
I suggested:
So maybe, then, I should issue an error if multi-dim features are used on < i32 arrays. For > i32 arrays, we could have the dimensions be in units of elemental size. So for an i32 array, one unit of stride is 4 bytes, for an i64 arrays, one unit of stride is 8 bytes. This would be consistent with the "length" argument (but I believe the "offset" argument to the dmaBdOp is in bytes ...)
@fifield argued to not do this:
In my opinion the wrap and step should expose the hardware registers directly without trying to do any conversion (e.g. no bytes -> words conversion). This is the cleanest and doesn't require casting to use multi-dimensional BDs in the general case (non-i32 data).
Just want to make sure you are aware of these previous discussions.
There was a problem hiding this comment.
This part of the PR already landed #1111.
It appears to me from this line we are making the strides be multiples of the memref data type, where previously it was always in 4-byte increments. (Correct me if wrong.)
You got it.
I had my original PR for this feature, see #547 (comment).
I wasn't aware such a discussion occurred (would've been nice to know - guess I should've checked...).
Regarding prior opinions/arguments/claims, specifically
It doesn't matter if it's intuitive. That's what the hardware does and at this level of aie dialect we should expose the hardware to the user.
I don't speak for everyone I guess but I think one of the cardinal sins of software is exposing unintuitive interfaces when you have the power to do otherwise. I mean like what's even the point of building an "abstraction" if you're just going to forward arguments/values untouched? Might as well not build it at all and tell people to just use the lower-level interface 🤷.
Too bad I wasn't around when you put up your PR - I would've advocated on your behalf to have it as it is here.
There was a problem hiding this comment.
Okay I will leave it to you whether this warrants further discussion. I think Jeff's argument was that our MLIR dialect is not supposed to be a high level abstraction but instead give easy low level access to the hardware (i.e. allow use of data layout transformations that aren't multiples of the word size without the need for casts). I personally think your change is good though, especially since you also made the offset more coherent.
Why does it show up as a diff here if the PR is already merged though?
There was a problem hiding this comment.
Why does it show up as a diff here if the PR is already merged though?
Because I haven't rebased (so its "base" is some previous commit in main before the merge).
There was a problem hiding this comment.
I think Jeff's argument was that our MLIR dialect is not supposed to be a high level abstraction but instead give easy low level access to the hardware (i.e. allow use of data layout transformations that aren't multiples of the word size without the need for casts).
Yes that's right. I was mostly arguing against casts iirc, and not against intuitive abstractions but for dialect constructs representing the hardware, even if unintuitive. The intuitive interfaces then legalize to the low level (unintuitive) dialect, which is less intended for humans to produce/consume.
|
Joe pointed me to this, I hope it's cool that I left three quick comments. Looks good |
Picked from #1111