[mlir][Vector] Add narrow type emulation pattern for vector.maskedload #68443
Conversation
| // (3xi4) from the memref, we need to set the second half of the last i8 | ||
| // that was effectively loaded (i.e. the second i8) to 0. |
There was a problem hiding this comment.
Trying to understand if this is really needed. Given that we are only emulating the vector loads/stores and bitcasting them back to the i4 type (see vector load case), it should be ok to just do an arith.select using the original mask, the loaded i4 vector and the original passthru value, right? If that is the case, all the code below would be really simplified!
There was a problem hiding this comment.
I'm not sure I see how to do this. Do you mean having a loop of arith.select based on the index?
There was a problem hiding this comment.
I'm probably the one missing something :) I meant:
%i8ld = vector.maskedload %0[%linear_index], %new_mask, %new_pass_thru ...
%i4ld = vector.bitcast %i8ld : vector<3xi8> to vector<6xi4>
%final_ld = arith.select %mask, %i4ld, %pass_thru : vector<6xi1>, vector<6xi4>
Would this work? The generated asm may not be ideal but it should definitely reduce the complexity of this lowering.
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
As for vector.load and vector.transfer_read ops, vector.maskedload op with narrow bitwidth (e.g., i4) can now be converted to supported wider bitwidth (e.g., i8 or i32).
… mask is an extraction
…vector.maskedload arith.select should be used instead of a series of manual mask manipulating ops (arith.and/or/extsi)
… emulation pattern
|
Thanks for the review! |
| func.func @fcst_maskedload(%A: memref<?xi4>, %passthru: vector<6xi4>) -> vector<6xi4> { | ||
| %c0 = arith.constant 0: index | ||
| %mask = vector.constant_mask [3] : vector<6xi1> | ||
| %1 = vector.maskedload %A[%c0], %mask, %passthru : | ||
| memref<?xi4>, vector<6xi1>, vector<6xi4> into vector<6xi4> | ||
| return %1 : vector<6xi4> | ||
| } |
There was a problem hiding this comment.
I just noticed that this is not tested because the patterns are not populated..
llvm-project/mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp
Lines 162 to 165 in a0bd636
To test it, we need to set up a type converter and populate those patterns.
No description provided.