chelini wrote: > It was suggested to me by @chelini to only have the `reshape` op to handle > all cases and get rid of the `expand_shape` op. We can then implement a > canonicalizer to convert when valid. I'm all for this, however want to make > sure this is the direction we want to go before I start making test changes. > Because there will be a lot. For each test there will be several extra lines > for populating the `shape` operand of `reshape`, like (from the test i added > in this PR) > > ``` > // CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index > // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index > // CHECK-DAG: %[[DIM0:.*]] = tensor.dim %[[ARG1]], %[[C0]] > // CHECK-DAG: %[[DIM1:.*]] = tensor.dim %[[ARG1]], %[[C1]] > // CHECK-DAG: %[[C3:.*]] = arith.constant 3 : index > // CHECK-DAG: %[[C2:.*]] = arith.constant 2 : index > // CHECK-DAG: %[[DIM2:.*]] = tensor.dim %[[ARG1]], %[[C2]] > // CHECK-DAG: %[[DIM3:.*]] = tensor.dim %[[ARG1]], %[[C3]] > // CHECK-NEXT: %[[INIT_SHAPE:.*]] = tensor.empty() : tensor<4xindex> > // CHECK-NEXT: %[[SHAPE0:.*]] = tensor.insert %[[DIM0]] into > %[[INIT_SHAPE]][%[[C0]]] > // CHECK-NEXT: %[[SHAPE1:.*]] = tensor.insert %[[DIM2]] into > %[[SHAPE0]][%[[C1]]] > // CHECK-NEXT: %[[SHAPE2:.*]] = tensor.insert %[[DIM1]] into > %[[SHAPE1]][%[[C2]]] > // CHECK-NEXT: %[[SHAPE3:.*]] = tensor.insert %[[DIM3]] into > %[[SHAPE2]][%[[C3]]] > // CHECK-NEXT: %[[EXPANDED:.*]] = tensor.reshape %[[PADDED]](%[[SHAPE3]]) > ``` > > Since this is a relatively expensive change, I'd like to get opinions before > I do it.
Thank you @srcarroll for pushing on this. Indeed, to generalize the lowering, we would need to emit a reshape operation, and I think it would be better to consistently emit the reshape and then "strength" reduce it to an expanded shape when possible. What folks think here @nicolasvasilache and @hanhanW? Thanks! https://github.com/llvm/llvm-project/pull/76003 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
