Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[AMDAIEInsertIntoCores] minor simplification #1118

Merged
merged 3 commits into from
Feb 20, 2025
Merged

[AMDAIEInsertIntoCores] minor simplification #1118

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4740287
they are nested in linalg ops anyway, so will get hoisted into cores …
newling Feb 19, 2025
20b6c61
remove tests
newling Feb 19, 2025
5fc1cb6
Merge branch 'main' into code_simplification
newling Feb 20, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 3 additions & 4 deletions compiler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/AMDAIEInsertCores.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,15 +32,14 @@ namespace {

/// Utility which returns 'true' is the operation needs to be inserted with an
/// `amdaie.core` op.
/// Some ops are surrrounded by scf.for loop nests. Place the entire
/// Some ops are surrounded by scf.for loop nests. Place the entire
/// loop nest inside the amdaie.core op here. Currently look for a
/// subset of ops which we know should be in the core.
/// TODO(newling) improve this design.
static bool isCoreComputeOp(Operation *op) {
return isa<linalg::LinalgOp, vector::ContractionOp,
memref::ExtractStridedMetadataOp, func::CallOp, arith::ExtFOp,
arith::TruncFOp, arith::TruncIOp, vector::TransferReadOp,
vector::TransferWriteOp>(op);
memref::ExtractStridedMetadataOp, func::CallOp,
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not necessarily, see matmul + truncf:

// input ${M}x${K}x${TYPE1}
// input ${K}x${N}x${TYPE1}

func.func @matmul_truncf(%arg0: tensor<${M}x${K}x${TYPE1}>, %arg1: tensor<${K}x${N}x${TYPE1}>) -> tensor<${M}x${N}x${TYPE1}>
{
  %cst = arith.constant ${ZERO} : ${TYPE2}
  %0 = tensor.empty() : tensor<${M}x${N}x${TYPE2}>
  %1 = linalg.fill ins(%cst : ${TYPE2}) outs(%0 : tensor<${M}x${N}x${TYPE2}>) -> tensor<${M}x${N}x${TYPE2}>
  %2 = linalg.matmul ins(%arg0, %arg1 : tensor<${M}x${K}x${TYPE1}>, tensor<${K}x${N}x${TYPE1}>)
    outs(%1: tensor<${M}x${N}x${TYPE2}>) -> tensor<${M}x${N}x${TYPE2}>
  %3 = arith.truncf %2 : tensor<${M}x${N}x${TYPE2}> to tensor<${M}x${N}x${TYPE1}>
  return %3: tensor<${M}x${N}x${TYPE1}>
}

Not sure how this PR is passing though...

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In one of the very early passes (convert-elementwise-to-linalg) It gets converted into a linalg.generic.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I see, thanks, that's great! Then we can assume all computational ops will be linalg ops, which makes these checks a lot less fragile and improves maintainability.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes!

vector::TransferReadOp, vector::TransferWriteOp>(op);
}

/// Utility to map the parallel mapping attributes to the corresponding
Expand Down
50 changes: 0 additions & 50 deletions compiler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/test/insert_cores.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -300,53 +300,3 @@ module {
return
}
}

// -----

// CHECK-LABEL: @insert_truncf_within_core
// CHECK: scf.forall
// CHECK: amdaie.tile
// CHECK: amdaie.core
// CHECK: vector.transfer_read
// CHECK: arith.truncf
// CHECK: vector.transfer_write
// CHECK: amdaie.end
module {
func.func @insert_truncf_within_core(%arg0: memref<10x10xf32, 2 : i32>, %arg1: memref<10x10xbf16, 2 : i32>) {
%cst = arith.constant 0.000000e+00 : f32
%c1 = arith.constant 1 : index
%c3 = arith.constant 3 : index
%c0 = arith.constant 0 : index
scf.forall (%arg3, %arg4) in (2, 2) {
%read = vector.transfer_read %arg0[%c0, %c1], %cst {in_bounds = [true, true]} : memref<10x10xf32, 2 : i32>, vector<1x1xf32>
%truncf = arith.truncf %read : vector<1x1xf32> to vector<1x1xbf16>
vector.transfer_write %truncf, %arg1[%c0, %c1] {in_bounds = [true, true]} : vector<1x1xbf16>, memref<10x10xbf16, 2 : i32>
} {mapping = [#gpu.thread<y>, #gpu.thread<x>]}
return
}
}

// -----

// CHECK-LABEL: @insert_trunci_within_core
// CHECK: scf.forall
// CHECK: amdaie.tile
// CHECK: amdaie.core
// CHECK: vector.transfer_read
// CHECK: arith.trunci
// CHECK: vector.transfer_write
// CHECK: amdaie.end
module {
func.func @insert_trunci_within_core(%arg0: memref<10x10xi32, 2 : i32>, %arg1: memref<10x10xi8, 2 : i32>) {
%cst = arith.constant 0 : i32
%c1 = arith.constant 1 : index
%c3 = arith.constant 3 : index
%c0 = arith.constant 0 : index
scf.forall (%arg3, %arg4) in (2, 2) {
%read = vector.transfer_read %arg0[%c0, %c1], %cst {in_bounds = [true, true]} : memref<10x10xi32, 2 : i32>, vector<1x1xi32>
%trunci = arith.trunci %read : vector<1x1xi32> to vector<1x1xi8>
vector.transfer_write %trunci, %arg1[%c0, %c1] {in_bounds = [true, true]} : vector<1x1xi8>, memref<10x10xi8, 2 : i32>
} {mapping = [#gpu.thread<y>, #gpu.thread<x>]}
return
}
}
Loading