Skip to content

[InsertGPUAllocs] Use gpu.memcpy for opencl instead of memref.copy #851

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
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[InsertGPUAllocs] Use gpu.memcpy for opencl instead of memref.copy #851

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
15 changes: 10 additions & 5 deletions lib/Transforms/InsertGPUAllocs.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -360,8 +360,10 @@ class InsertGPUAllocsPass final
auto newAlloc = builder.create<mlir::memref::AllocOp>(
loc, alloc.getType(), alloc.getDynamicSizes(),
alloc.getSymbolOperands());
builder.create<mlir::memref::CopyOp>(loc, allocResult,
newAlloc.getResult());
builder.create<mlir::gpu::MemcpyOp>(
loc, /*asyncToken*/ static_cast<mlir::Type>(nullptr),
/*asyncDependencies*/ std::nullopt, newAlloc.getResult(),
allocResult);
use.set(newAlloc.getResult());
}
}
Expand Down Expand Up @@ -401,8 +403,9 @@ class InsertGPUAllocsPass final
/*symbolOperands*/ std::nullopt, hostShared);
auto allocResult = gpuAlloc.getResult(0);
if (access.hostWrite && access.deviceRead) {
auto copy =
builder.create<mlir::memref::CopyOp>(loc, op, allocResult);
auto copy = builder.create<mlir::gpu::MemcpyOp>(
loc, /*asyncToken*/ static_cast<mlir::Type>(nullptr),
/*asyncDependencies*/ std::nullopt, allocResult, op);
filter.insert(copy);
}

Expand All @@ -421,7 +424,9 @@ class InsertGPUAllocsPass final
op.replaceAllUsesExcept(allocResult, filter);
builder.setInsertionPoint(term);
if (access.hostRead && access.deviceWrite) {
builder.create<mlir::memref::CopyOp>(loc, allocResult, op);
builder.create<mlir::gpu::MemcpyOp>(
loc, /*asyncToken*/ static_cast<mlir::Type>(nullptr),
/*asyncDependencies*/ std::nullopt, op, allocResult);
}
builder.create<mlir::gpu::DeallocOp>(loc, std::nullopt, allocResult);
}
Expand Down
4 changes: 2 additions & 2 deletions test/Transforms/InsertGpuAllocs/add-gpu-alloc.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,9 +7,9 @@ func.func @addt(%arg0: memref<2x5xf32>, %arg1: memref<2x5xf32>) -> memref<2x5xf3
%c1 = arith.constant 1 : index
%c5 = arith.constant 5 : index
// OPENCL: %[[MEMREF0:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
// OPENCL: memref.copy %arg1, %[[MEMREF0]] : memref<2x5xf32> to memref<2x5xf32>
// OPENCL: gpu.memcpy %[[MEMREF0]], %arg1 : memref<2x5xf32>, memref<2x5xf32>
// OPENCL: %[[MEMREF1:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
// OPENCL: memref.copy %arg0, %[[MEMREF1]] : memref<2x5xf32> to memref<2x5xf32>
// OPENCL: gpu.memcpy %[[MEMREF1]], %arg0 : memref<2x5xf32>, memref<2x5xf32>
// VULKAN: %[[MEMREF0:.*]] = memref.alloc() : memref<2x5xf32>
// VULKAN: memref.copy %arg1, %[[MEMREF0]] : memref<2x5xf32> to memref<2x5xf32>
// VULKAN: %[[MEMREF1:.*]] = memref.alloc() : memref<2x5xf32>
Expand Down
4 changes: 2 additions & 2 deletions test/Transforms/InsertGpuAllocs/memref-get-global.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,10 +17,10 @@ func.func @addt(%arg0: memref<2x5xf32>, %arg1: memref<2x5xf32>) -> memref<2x5xf3

// OPENCL: [[VAR0:%.*]] = memref.get_global @__constant_2x5xf32 : memref<2x5xf32>
// OPENCL: %[[MEMREF0:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
// OPENCL: memref.copy [[VAR0]], %[[MEMREF0]] : memref<2x5xf32> to memref<2x5xf32>
// OPENCL: gpu.memcpy %[[MEMREF0]], [[VAR0]] : memref<2x5xf32>, memref<2x5xf32>
// OPENCL: [[VAR1:%.*]] = memref.get_global @__constant_2x5xf32_0 : memref<2x5xf32>
// OPENCL: %[[MEMREF1:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
// OPENCL: memref.copy [[VAR1]], %[[MEMREF1]] : memref<2x5xf32> to memref<2x5xf32>
// OPENCL: gpu.memcpy %[[MEMREF1]], [[VAR1]] : memref<2x5xf32>, memref<2x5xf32>
// OPENCL: %[[MEMREF2:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
// VULKAN: [[VAR0:%.*]] = memref.get_global @__constant_2x5xf32 : memref<2x5xf32>
// VULKAN: %[[MEMREF0:.*]] = memref.alloc() : memref<2x5xf32>
Expand Down
2 changes: 1 addition & 1 deletion test/Transforms/InsertGpuAllocs/memref-returned-from-call.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ func.func @main() {
// OPENCL: func.func @main()
%0 = func.call @alloc_buffer() : () -> memref<8xf32>
// OPENCL: %[[MEMREF:.*]] = gpu.alloc host_shared () : memref<8xf32>
// OPENCL: memref.copy %0, %[[MEMREF]] : memref<8xf32> to memref<8xf32>
// OPENCL: gpu.memcpy %[[MEMREF]], %0 : memref<8xf32>, memref<8xf32>
%1 = memref.alloc() : memref<8xf32>
%2 = memref.alloc() : memref<8xf32>
gpu.launch blocks(%arg0, %arg1, %arg2) in (%arg6 = %c8, %arg7 = %c1, %arg8 = %c1) threads(%arg3, %arg4, %arg5) in (%arg9 = %c1, %arg10 = %c1, %arg11 = %c1) {
Expand Down