Enable the safe return in neura dataflow ir #223
Conversation
| Example: | ||
| neura.return_void %cond : !neura.data<i1, i1> // In dataflow mode | ||
|
|
||
| This is NOT a terminator - the block ends with neura.yield instead. |
There was a problem hiding this comment.
Then where is yield? Can the neura.yield be also shown in above example?
|
|
||
| // Looks for any suitable value in the predecessor block. | ||
| if (op.getNumResults() > 0) { | ||
| trigger_value = op.getResult(0); |
There was a problem hiding this comment.
Elaborate on the "suitable value"? This line seems any op can serve as trigger_value, is this sufficient?
| builder.create<neura::CondBr>( | ||
| cond_br.getLoc(), cond_br.getCondition(), true_args, false_args, | ||
| cond_br.getTrueDest(), cond_br.getFalseDest()); | ||
| cond_br.erase(); |
There was a problem hiding this comment.
I thought we need to or all the cond_br's conditions?
| // Handles cond_preds: adds a block argument for the trigger value, and | ||
| // updates each predecessor's terminator to pass the trigger value. | ||
| BlockArgument trigger_arg = | ||
| ret_block->addArgument(builder.getI1Type(), void_ret_op.getLoc()); |
There was a problem hiding this comment.
triggering value is getI1Type? I thought it is or of all conditions, and needs to be a regular type, i.e., <data, predicate>?
| } | ||
|
|
||
| Value getPrecessedCondition(Value condition, bool is_not_condition, | ||
| Value getProcessedCondition(Value condition, bool is_not_condition, |
There was a problem hiding this comment.
Can you remind me why do we need this getProcessedCondition()?
| llvm::errs() << "[ctrl2data] Converting neura.return operations to " | ||
| "return_void/value...\n"; | ||
|
|
||
| for (neura::ReturnOp return_op : return_ops) { |
There was a problem hiding this comment.
What is the reason that we can't have one pass (e.g., canonicalize_return_pass) to handle the transformation? Because canonicalize_return_pass touches condition, and condition would be transformed in ctrl2data_pass?
There was a problem hiding this comment.
Why remove this? Can't this test be updated/fixed?
| //MAPPING: func.func @kernel | ||
| //MAPPING-SAME: accelerator = "neura", dataflow_mode = "predicate" | ||
| //MAPPING-SAME: mapping_info = {compiled_ii = 11 : i32, mapping_mode = "spatial-temporal", mapping_strategy = "heuristic", rec_mii = 9 : i32, res_mii = 5 : i32, x_tiles = 4 : i32, y_tiles = 4 : i32} | ||
| //MAPPING-SAME: mapping_info = {compiled_ii = 13 : i32, mapping_mode = "spatial-temporal", mapping_strategy = "heuristic", rec_mii = 9 : i32, res_mii = 6 : i32, x_tiles = 4 : i32, y_tiles = 4 : i32} |
There was a problem hiding this comment.
II increased due to more cond ops are added?
2 participants
In this pr:
Introduce two kinds of return ops in neura dataflow ir:
neura.return_void: its operand means the trigger valuesneura.return_value: its operand means the returned valuesThese two kinds of return ops are only enabled in dataflow ir. In conventional neural CDFG ir, we still use the
neura.return.We create a
canonicalize-returnpass to mark the type (return_type = void/value) ofneura.return. AndWe introduce some logic in the
transform-ctrl-to-data-flowpass to transform theneura.returntoneura.return_value/return_voidA
neura.yieldis added at each flattened block's end to act like a terminator for the block. This is just to obey the rules in MLIR block https://mlir.llvm.org/docs/LangRef/#blocks.TODO: Enable the interpreter for these two introduced ops. Since this pr is too large, I will make a separate pr for that.