This is an archive of the discontinued LLVM Phabricator instance.

Differential D156009

[mlir][rocdl] Change the translation of `GridDim*Op` to `__ockl_get_num_groups`
ClosedPublic

Authored by fmorac on Jul 21 2023, 6:25 PM.

Details

Reviewers
ftynse
dcaballe
krzysz00
nicolasvasilache
Commits
rG4538347fb2da: [mlir][rocdl] Change the translation of `GridDim*Op` to `__ockl_get_num_groups`
Summary

Currently, ROCDL::GridDim*Op is being translated to __ockl_get_global_size, however
to match the meaning of gpu.grid_dim it should instead be translated to
__ockl_get_num_groups. This change would also make it agree with the meaning
of gridDimx.* in HIP, see:
https://github.com/ROCm-Developer-Tools/hipamd/blob/develop/include/hip/amd_detail/amd_hip_runtime.h#L257

Difference between the functions:

__ockl_get_global_size =  blockDim * numBlocks
__ockl_get_num_groups = numBlocks

Diff Detail

Event Timeline

fmorac created this revision.Jul 21 2023, 6:25 PM
Herald added a reviewer: ftynse. · View Herald TranscriptJul 21 2023, 6:25 PM
Herald added a reviewer: dcaballe. · View Herald Transcript
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: gysit, Dinistro, bviyer and 24 others. · View Herald Transcript
fmorac edited the summary of this revision. (Show Details)Jul 21 2023, 6:27 PM
fmorac added a reviewer: krzysz00.
fmorac edited the summary of this revision. (Show Details)
fmorac published this revision for review.Jul 21 2023, 6:32 PM
Herald added a reviewer: nicolasvasilache. · View Herald TranscriptJul 21 2023, 6:32 PM
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
krzysz00 added a comment.Jul 21 2023, 7:12 PM

Looks good to me, thanks for catching this!

krzysz00 accepted this revision.Jul 21 2023, 7:12 PM
This revision is now accepted and ready to land.Jul 21 2023, 7:12 PM
Harbormaster completed remote builds in B247365: Diff 543130.Jul 21 2023, 11:39 PM
Closed by commit rG4538347fb2da: [mlir][rocdl] Change the translation of `GridDim*Op` to `__ockl_get_num_groups` (authored by fmorac). · Explain WhyJul 22 2023, 4:33 AM
This revision was automatically updated to reflect the committed changes.
fmorac added a commit: rG4538347fb2da: [mlir][rocdl] Change the translation of `GridDim*Op` to `__ockl_get_num_groups`.

Revision Contents

PathSize
mlir/
include/
mlir/
Dialect/
LLVMIR/
6 lines
test/
Target/
LLVMIR/
6 lines

Diff 543169

mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td

Show First 20 Lines • Show All 95 Lines • ▼ Show 20 Lines
def ROCDL_BlockDimYOp : ROCDL_DeviceFunctionOp<"workgroup.dim.y", def ROCDL_BlockDimYOp : ROCDL_DeviceFunctionOp<"workgroup.dim.y",
"__ockl_get_local_size", 1>; "__ockl_get_local_size", 1>;
def ROCDL_BlockDimZOp : ROCDL_DeviceFunctionOp<"workgroup.dim.z", def ROCDL_BlockDimZOp : ROCDL_DeviceFunctionOp<"workgroup.dim.z",
"__ockl_get_local_size", 2>; "__ockl_get_local_size", 2>;
def ROCDL_GridDimXOp : ROCDL_DeviceFunctionOp<"grid.dim.x", def ROCDL_GridDimXOp : ROCDL_DeviceFunctionOp<"grid.dim.x",
"__ockl_get_global_size", 0>; "__ockl_get_num_groups", 0>;
def ROCDL_GridDimYOp : ROCDL_DeviceFunctionOp<"grid.dim.y", def ROCDL_GridDimYOp : ROCDL_DeviceFunctionOp<"grid.dim.y",
"__ockl_get_global_size", 1>; "__ockl_get_num_groups", 1>;
def ROCDL_GridDimZOp : ROCDL_DeviceFunctionOp<"grid.dim.z", def ROCDL_GridDimZOp : ROCDL_DeviceFunctionOp<"grid.dim.z",
"__ockl_get_global_size", 2>; "__ockl_get_num_groups", 2>;
//===----------------------------------------------------------------------===// //===----------------------------------------------------------------------===//
// Synchronization primitives // Synchronization primitives
def ROCDL_BarrierOp : ROCDL_Op<"barrier"> { def ROCDL_BarrierOp : ROCDL_Op<"barrier"> {
string llvmBuilder = [{ string llvmBuilder = [{
llvm::LLVMContext &llvmContext = builder.getContext(); llvm::LLVMContext &llvmContext = builder.getContext();
builder.CreateFence(llvm::AtomicOrdering::Release, builder.CreateFence(llvm::AtomicOrdering::Release,
▲ Show 20 Lines • Show All 253 Lines • Show Last 20 Lines

mlir/test/Target/LLVMIR/rocdl.mlir

Show All 14 Linesllvm.func @rocdl_special_regs() -> i32 {
// CHECK: call i32 @llvm.amdgcn.workgroup.id.z() // CHECK: call i32 @llvm.amdgcn.workgroup.id.z()
%6 = rocdl.workgroup.id.z : i32 %6 = rocdl.workgroup.id.z : i32
// CHECK: call i64 @__ockl_get_local_size(i32 0) // CHECK: call i64 @__ockl_get_local_size(i32 0)
%7 = rocdl.workgroup.dim.x : i64 %7 = rocdl.workgroup.dim.x : i64
// CHECK: call i64 @__ockl_get_local_size(i32 1) // CHECK: call i64 @__ockl_get_local_size(i32 1)
%8 = rocdl.workgroup.dim.y : i64 %8 = rocdl.workgroup.dim.y : i64
// CHECK: call i64 @__ockl_get_local_size(i32 2) // CHECK: call i64 @__ockl_get_local_size(i32 2)
%9 = rocdl.workgroup.dim.z : i64 %9 = rocdl.workgroup.dim.z : i64
// CHECK: call i64 @__ockl_get_global_size(i32 0) // CHECK: call i64 @__ockl_get_num_groups(i32 0)
%10 = rocdl.grid.dim.x : i64 %10 = rocdl.grid.dim.x : i64
// CHECK: call i64 @__ockl_get_global_size(i32 1) // CHECK: call i64 @__ockl_get_num_groups(i32 1)
%11 = rocdl.grid.dim.y : i64 %11 = rocdl.grid.dim.y : i64
// CHECK: call i64 @__ockl_get_global_size(i32 2) // CHECK: call i64 @__ockl_get_num_groups(i32 2)
%12 = rocdl.grid.dim.z : i64 %12 = rocdl.grid.dim.z : i64
// CHECK: call i32 @llvm.amdgcn.workitem.id.x(),{{.*}} !range ![[$RANGE:[0-9]+]] // CHECK: call i32 @llvm.amdgcn.workitem.id.x(),{{.*}} !range ![[$RANGE:[0-9]+]]
%13 = rocdl.workitem.id.x {range = array<i32: 0, 64>} : i32 %13 = rocdl.workitem.id.x {range = array<i32: 0, 64>} : i32
llvm.return %1 : i32 llvm.return %1 : i32
} }
▲ Show 20 Lines • Show All 338 Lines • Show Last 20 Lines