Skip to content

Fix crash in ConvertParallelToGPU1 pass on grid×block loops with multiple inner parallels #2564

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

Open
xys-syx wants to merge 3 commits into
base: recreateexpr
Choose a base branch
from
+93 −2
Open

Fix crash in ConvertParallelToGPU1 pass on grid×block loops with multiple inner parallels #2564

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b62ef63
modify if guard
xys-syx Jun 18, 2026
bee8cc7
fix for lowerdim
xys-syx Jul 16, 2026
fbc1da0
fix
xys-syx Jul 24, 2026
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
11 changes: 5 additions & 6 deletions src/enzyme_ad/jax/Passes/ConvertParallelToGPU.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1886,13 +1886,12 @@ struct InnerParallelSerialization : public OpRewritePattern<scf::ParallelOp> {
if (!parallelOp->getParentOfType<enzymexla::GPUWrapperOp>()) {
return failure();
}
size_t parallelCount = 0;

size_t enclosingGpuDims = 0;
auto par = parallelOp;
while ((par = par->getParentOfType<scf::ParallelOp>())) {
parallelCount++;
}
// is presently one of the three outer parallel loops;
if (parallelCount < 2)
while ((par = par->getParentOfType<scf::ParallelOp>()))
enclosingGpuDims += par.getNumLoops();
if (enclosingGpuDims < 6)
return failure();
Comment thread
vimarsh6739 marked this conversation as resolved.
Outdated

// For a parallel loop, we essentially need to create an n-dimensional loop
Expand Down
51 changes: 51 additions & 0 deletions test/lit_tests/lowering/convert-parallel-to-gpu1-fused-siblings.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,51 @@
// RUN: enzymexlamlir-opt %s --pass-pipeline="builtin.module(convert-parallel-to-gpu1)" | FileCheck %s

module {
func.func @fused_siblings(%g0: index, %g1: index, %g2: index,
%b0: index, %b1: index, %b2: index,
%arg: memref<8xf64>) {
%r = "enzymexla.gpu_wrapper"(%g0, %g1, %g2, %b0, %b1, %b2) ({
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%c8 = arith.constant 8 : index
%cst = arith.constant 1.0 : f64
scf.parallel (%i0, %i1, %i2, %i3, %i4, %i5) =
(%c0, %c0, %c0, %c0, %c0, %c0) to (%g0, %g1, %g2, %b0, %b1, %b2)
step (%c1, %c1, %c1, %c1, %c1, %c1) {
scf.parallel (%j) = (%c0) to (%c8) step (%c1) {
memref.store %cst, %arg[%j] : memref<8xf64>
scf.reduce
}
scf.parallel (%k) = (%c0) to (%c8) step (%c1) {
memref.store %cst, %arg[%k] : memref<8xf64>
scf.reduce
}
scf.parallel (%l) = (%c0) to (%c8) step (%c1) {
memref.store %cst, %arg[%l] : memref<8xf64>
scf.reduce
}
scf.reduce
}
"enzymexla.polygeist_yield"() : () -> ()
}) : (index, index, index, index, index, index) -> index
return
}
}

// CHECK-LABEL: func.func @fused_siblings(
// CHECK: "enzymexla.alternatives"() ({
// CHECK: "enzymexla.gpu_error"() ({
// CHECK: scf.if %{{.*}} {
// CHECK: gpu.launch blocks(%{{.*}}, %{{.*}}, %{{.*}}) in (%{{.*}} = %{{.*}}, %{{.*}} = %{{.*}}, %{{.*}} = %{{.*}}) threads(%{{.*}}, %{{.*}}, %{{.*}}) in (%{{.*}} = %{{.*}}, %{{.*}} = %{{.*}}, %{{.*}} = %{{.*}}) {
// CHECK: scf.for %{{.*}} = %{{.*}} to %{{.*}} step %{{.*}} {
// CHECK: memref.store %{{.*}}, %{{.*}}[%{{.*}}] : memref<8xf64>
// CHECK: }
// CHECK: scf.for %{{.*}} = %{{.*}} to %{{.*}} step %{{.*}} {
// CHECK: memref.store %{{.*}}, %{{.*}}[%{{.*}}] : memref<8xf64>
// CHECK: }
// CHECK: scf.for %{{.*}} = %{{.*}} to %{{.*}} step %{{.*}} {
// CHECK: memref.store %{{.*}}, %{{.*}}[%{{.*}}] : memref<8xf64>
// CHECK: }
// CHECK: gpu.terminator
// CHECK: "enzymexla.polygeist_yield"() : () -> ()
// CHECK: "enzymexla.polygeist_yield"() : () -> ()