mlir: Add Enzyme ops removal on structured control flow #2200
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
wsmoses
reviewed
Dec 20, 2024
| return mlir::enzyme::CacheInfo::batchType(mlir::ShapedType::kDynamic); | ||
| } | ||
|
|
||
| mlir::Type mlir::enzyme::CacheInfo::batchType(int64_t dim) { |
There was a problem hiding this comment.
so there is already an Enzyme Autodiff Type interface, which should have a method for batching (and if not that would probably be the right place for this)
wsmoses
approved these changes
Dec 20, 2024
This still requires changes in the tblgenerated derivative files. For example, createForwardModeTangent in MulFOpFwdDerivative could be altered like this:
```
LogicalResult createForwardModeTangent(Operation *op0, OpBuilder &builder, MGradientUtils *gutils) const
{
auto op = cast<arith::MulFOp>(op0);
if (gutils->width != 1) {
auto newop = gutils->getNewFromOriginal(op0);
for (auto res : newop->getResults()) {
res.setType(mlir::RankedTensorType::get({gutils->width}, res.getType()));
}
}
gutils->eraseIfUnused(op);
if (gutils->isConstantInstruction(op))
return success();
mlir::Value res = nullptr;
if (!gutils->isConstantValue(op->getOperand(0)))
{
auto dif = gutils->invertPointerM(op->getOperand(0), builder);
{
mlir::Value itmp = ({
// Computing MulFOp
auto fwdarg_0 = dif;
dif.dump();
// TODO: gutils->makeBatched(...)
auto fwdarg_1 = gutils->getNewFromOriginal(op->getOperand(1));
builder.create<arith::MulFOp>(op.getLoc(), fwdarg_0, fwdarg_1);
});
itmp.dump();
if (!res)
res = itmp;
else
{
auto operandType = cast<AutoDiffTypeInterface>(res.getType());
res = operandType.createAddOp(builder, op.getLoc(), res, itmp);
}
}
}
if (!gutils->isConstantValue(op->getOperand(1)))
{
auto dif = gutils->invertPointerM(op->getOperand(1), builder);
{
mlir::Value itmp = ({
// Computing MulFOp
auto fwdarg_0 = dif;
dif.dump();
auto fwdarg_1 = gutils->getNewFromOriginal(op->getOperand(0));
builder.create<arith::MulFOp>(op.getLoc(), fwdarg_0, fwdarg_1);
});
if (!res)
res = itmp;
else
{
auto operandType = cast<AutoDiffTypeInterface>(res.getType());
res = operandType.createAddOp(builder, op.getLoc(), res, itmp);
}
}
}
assert(res);
gutils->setDiffe(op->getResult(0), res, builder);
return success();
}
```
wsmoses
reviewed
Jan 3, 2025
| if (!gutils->isConstantValue(prev)) | ||
| gutils->addToDiffe(prev, post, builder); | ||
| auto numIters = getConstantNumberOfIterations(forOp); | ||
| Value inductionVariable; // [0, N[ counter |
wsmoses
reviewed
Jan 3, 2025
| @@ -27,7 +27,11 @@ getFunctionTypeForClone(mlir::FunctionType FTy, DerivativeMode mode, | |||
| for (auto &&[Ty, returnPrimal, returnShadow, activity] : llvm::zip( | |||
| FTy.getResults(), returnPrimals, returnShadows, ReturnActivity)) { | |||
| if (returnPrimal) { | |||
| RetTypes.push_back(Ty); | |||
| if (width != 1) { | |||
There was a problem hiding this comment.
this shouldn't be modified since width only applies to the derivative not primal return
wsmoses
reviewed
Jan 3, 2025
| @@ -232,6 +240,11 @@ FunctionOpInterface CloneFunctionWithReturns( | |||
|
|
|||
| { | |||
| auto &blk = NewF.getFunctionBody().front(); | |||
| if (width != 1) { | |||
wsmoses
reviewed
Jan 3, 2025
| assert(width == 1 && "unsupported width != 1"); | ||
| return self; | ||
| Type getShadowType(Type self, int64_t width) const { | ||
| return batchType(self, width); |
There was a problem hiding this comment.
in a separate different PR, it may be worthwhile switching getShadowType and the likes to take an ArrayRef<int64_t> indices to batch on (@jumerckx did something similar when adding batched differentiation broadcast earlier)
wsmoses
approved these changes
Jan 3, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
TODO:
Cache<f32> -> tensor<?xf32>.