gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[OPTIMIZER] Fix Num in AsyncWaitOp generated by the pipeline pass (#72)

Browse Source
This commit is contained in:
Da Yan
2022-08-23 06:58:10 +08:00
committed by GitHub
parent 10ba51c3bb
commit 92ef552a54
2 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
lib/Dialect/TritonGPU/Transforms/Pipeline.cpp
View File

@@ -337,8 +337,9 @@ scf::ForOp LoopPipeliner::createNewForOp() {
for (Operation &op : forOp.getBody()->without_terminator()) { for (Operation &op : forOp.getBody()->without_terminator()) {
if (!asyncWaitInserted && isDirectUserOfAsyncLoad(op)) { if (!asyncWaitInserted && isDirectUserOfAsyncLoad(op)) {
asyncWaitInserted = true; asyncWaitInserted = true;
assert(numStages >= 2);
builder.create<triton::gpu::AsyncWaitOp>(op.getLoc(), builder.create<triton::gpu::AsyncWaitOp>(op.getLoc(),
loads.size() * (numStages - 1)); loads.size() * (numStages - 2));
} }
Operation *newOp = builder.clone(op, mapping); Operation *newOp = builder.clone(op, mapping);
// update mapping of results // update mapping of results

6
test/TritonGPU/loop-pipeline.mlir
View File

@@ -14,7 +14,7 @@
// CHECK: %[[A1:.*]] = triton_gpu.copy_async // CHECK: %[[A1:.*]] = triton_gpu.copy_async
// CHECK: %[[B1:.*]] = triton_gpu.copy_async // CHECK: %[[B1:.*]] = triton_gpu.copy_async
// CHECK: scf.for {{.*}} iter_args({{.*}}, {{.*}}, {{.*}}, %[[arg_a0:.*]] = %[[A0]], %[[arg_a1:.*]] = %[[A1]], %[[arg_b0:.*]] = %[[B0]], %[[arg_b1:.*]] = %[[B1]], {{.*}}) // CHECK: scf.for {{.*}} iter_args({{.*}}, {{.*}}, {{.*}}, %[[arg_a0:.*]] = %[[A0]], %[[arg_a1:.*]] = %[[A1]], %[[arg_b0:.*]] = %[[B0]], %[[arg_b1:.*]] = %[[B1]], {{.*}})
// CHECK: triton_gpu.async_wait {num = 4 : i32} // CHECK: triton_gpu.async_wait {num = 2 : i32}
// CHECK: tt.dot %[[arg_a0]], %[[arg_b0]], {{.*}} // CHECK: tt.dot %[[arg_a0]], %[[arg_b0]], {{.*}}
// CHECK: %[[NEXT_A:.*]] = triton_gpu.copy_async // CHECK: %[[NEXT_A:.*]] = triton_gpu.copy_async
// CHECK: %[[NEXT_B:.*]] = triton_gpu.copy_async // CHECK: %[[NEXT_B:.*]] = triton_gpu.copy_async
@@ -55,7 +55,7 @@ func @matmul_loop(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f16>, %B
// CHECK: %[[A1:.*]] = triton_gpu.copy_async // CHECK: %[[A1:.*]] = triton_gpu.copy_async
// CHECK: %[[B1:.*]] = triton_gpu.copy_async // CHECK: %[[B1:.*]] = triton_gpu.copy_async
// CHECK: scf.for {{.*}} iter_args({{.*}}, {{.*}}, {{.*}}, %[[arg_a0:.*]] = %[[A0]], %[[arg_a1:.*]] = %[[A1]], %[[arg_b0:.*]] = %[[B0]], %[[arg_b1:.*]] = %[[B1]], {{.*}}) // CHECK: scf.for {{.*}} iter_args({{.*}}, {{.*}}, {{.*}}, %[[arg_a0:.*]] = %[[A0]], %[[arg_a1:.*]] = %[[A1]], %[[arg_b0:.*]] = %[[B0]], %[[arg_b1:.*]] = %[[B1]], {{.*}})
// CHECK: triton_gpu.async_wait {num = 4 : i32} // CHECK: triton_gpu.async_wait {num = 2 : i32}
// CHECK: tt.dot %[[arg_a0]], %[[arg_b0]], {{.*}} // CHECK: tt.dot %[[arg_a0]], %[[arg_b0]], {{.*}}
// CHECK: %[[NEXT_A:.*]] = triton_gpu.copy_async // CHECK: %[[NEXT_A:.*]] = triton_gpu.copy_async
// CHECK: %[[NEXT_B:.*]] = triton_gpu.copy_async // CHECK: %[[NEXT_B:.*]] = triton_gpu.copy_async
@@ -95,7 +95,7 @@ func @matmul_loop_nested(%lb : index, %ub : index, %step : index, %A : !tt.ptr<f
// CHECK: %[[B0:.*]] = triton_gpu.copy_async // CHECK: %[[B0:.*]] = triton_gpu.copy_async
// CHECK: %[[B1:.*]] = triton_gpu.copy_async // CHECK: %[[B1:.*]] = triton_gpu.copy_async
// CHECK: scf.for {{.*}} iter_args({{.*}}, {{.*}}, %[[arg_b0:.*]] = %[[B0]], %[[arg_b1:.*]] = %[[B1]], {{.*}}) // CHECK: scf.for {{.*}} iter_args({{.*}}, {{.*}}, %[[arg_b0:.*]] = %[[B0]], %[[arg_b1:.*]] = %[[B1]], {{.*}})
// CHECK: triton_gpu.async_wait {num = 2 : i32} // CHECK: triton_gpu.async_wait {num = 1 : i32}
// CHECK: tt.dot {{.*}}, %[[arg_b0]], {{.*}} // CHECK: tt.dot {{.*}}, %[[arg_b0]], {{.*}}
// CHECK: %[[NEXT_B:.*]] = triton_gpu.copy_async // CHECK: %[[NEXT_B:.*]] = triton_gpu.copy_async
// CHECK: scf.yield {{.*}}, {{.*}}, %[[arg_b1]], %[[NEXT_B]] // CHECK: scf.yield {{.*}}, {{.*}}, %[[arg_b1]], %[[NEXT_B]]