Skip to content

Commit 0b42ea0

Browse files
amd-eochoaloamd-eochoalo
committed
Revert "[mlir][amdgpu] Make tdm_descriptor parametric"
This reverts commit 14eb785.
1 parent 3a9dbb9 commit 0b42ea0

File tree

6 files changed

+20
-34
lines changed

6 files changed

+20
-34
lines changed

‎mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td‎

Lines changed: 0 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -202,11 +202,6 @@ def AMDGPU_TDMDescriptorType : AMDGPU_Type<"TDMDescriptor", "tdm_descriptor"> {
202202
used in tensor_load_to_lds or tensor_store_from_lds.
203203
}];
204204

205-
let parameters = (ins "unsigned": $size);
206-
207-
let assemblyFormat = "`<` $size `>`";
208-
let genVerifyDecl = 1;
209-
210205
}
211206

212207
//===----------------------------------------------------------------------===//

‎mlir/lib/Dialect/AMDGPU/IR/AMDGPUDialect.cpp‎

Lines changed: 0 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -781,15 +781,6 @@ LogicalResult MakeDmaBaseOp::verify() {
781781
// MakeDmaDescriptorOp
782782
//===----------------------------------------------------------------------===//
783783

784-
LogicalResult
785-
TDMDescriptorType::verify(function_ref<InFlightDiagnostic()> emitError,
786-
unsigned size) {
787-
if (!llvm::is_contained<unsigned>({2, 4}, size))
788-
return emitError() << "only groups of size 2 or 4 are valid but got "
789-
<< size;
790-
return success();
791-
}
792-
793784
LogicalResult MakeDmaDescriptorOp::verify() {
794785
ArrayRef<int64_t> globalStaticStrides = getGlobalStaticStrides();
795786

‎mlir/test/Conversion/AMDGPUToROCDL/gfx1250.mlir‎

Lines changed: 7 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -239,7 +239,7 @@ func.func @make_dma_base(%idx: index, %mem: memref<8xi32, #gpu_global_addrspace>
239239

240240
// CHECK-LABEL: func @make_dma_descriptor
241241
// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>)
242-
func.func @make_dma_descriptor(%base: !amdgpu.tdm_base<i32>) -> !amdgpu.tdm_descriptor<2> {
242+
func.func @make_dma_descriptor(%base: !amdgpu.tdm_base<i32>) -> !amdgpu.tdm_descriptor {
243243
// CHECK-DAG: %[[DGROUP0:.+]] = builtin.unrealized_conversion_cast %[[BASE]]
244244

245245
// CHECK-DAG: %[[C0:.+]] = llvm.mlir.constant(0 : i32)
@@ -304,9 +304,9 @@ func.func @make_dma_descriptor(%base: !amdgpu.tdm_base<i32>) -> !amdgpu.tdm_desc
304304
// CHECK: %[[DGROUP1_6:.+]] = llvm.insertelement %[[SGPR6]], %[[DGROUP1_5]][%[[C6]] : i32]
305305
// CHECK: %[[DGROUP1:.+]] = llvm.insertelement %[[SGPR7]], %[[DGROUP1_6]][%[[C7]] : i32]
306306

307-
// CHECK: %[[DGROUPS:.+]] = builtin.unrealized_conversion_cast %[[DGROUP0]], %[[DGROUP1]] : vector<4xi32>, vector<8xi32> to !amdgpu.tdm_descriptor<2>
308-
%descriptor = amdgpu.make_dma_descriptor %base globalSize [128, 64] globalStride [64, 1] sharedSize [128, 64] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
309-
func.return %descriptor : !amdgpu.tdm_descriptor<2>
307+
// CHECK: %[[DGROUPS:.+]] = builtin.unrealized_conversion_cast %[[DGROUP0]], %[[DGROUP1]] : vector<4xi32>, vector<8xi32> to !amdgpu.tdm_descriptor
308+
%descriptor = amdgpu.make_dma_descriptor %base globalSize [128, 64] globalStride [64, 1] sharedSize [128, 64] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
309+
func.return %descriptor : !amdgpu.tdm_descriptor
310310
}
311311

312312
// -----
@@ -317,7 +317,7 @@ func.func @make_dma_descriptor(%base: !amdgpu.tdm_base<i32>) -> !amdgpu.tdm_desc
317317

318318
// CHECK-LABEL: func @make_dma_descriptor_atomic_barrier
319319
// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>, %[[BARRIER:.+]]: {{.*}}, %[[IDX:.+]]: index)
320-
func.func @make_dma_descriptor_atomic_barrier(%base: !amdgpu.tdm_base<i32>, %barrier : memref<8xi32, #gpu_lds_addrspace>, %idx: index) -> !amdgpu.tdm_descriptor<2> {
320+
func.func @make_dma_descriptor_atomic_barrier(%base: !amdgpu.tdm_base<i32>, %barrier : memref<8xi32, #gpu_lds_addrspace>, %idx: index) -> !amdgpu.tdm_descriptor {
321321
// CHECK-DAG: %[[INDEX:.+]] = builtin.unrealized_conversion_cast %[[IDX]] : index to i64
322322
// CHECK-DAG: %[[BARRIER_MEMREF_DESC:.+]] = builtin.unrealized_conversion_cast %[[BARRIER]]
323323
// CHECK-DAG: %[[DGROUP0:.+]] = builtin.unrealized_conversion_cast %[[BASE]]
@@ -361,8 +361,8 @@ func.func @make_dma_descriptor_atomic_barrier(%base: !amdgpu.tdm_base<i32>, %bar
361361
globalStride [64, 1]
362362
sharedSize [128, 64]
363363
atomicBarrier(%barrier[%idx] : memref<8xi32, #gpu_lds_addrspace>)
364-
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
365-
func.return %descriptor : !amdgpu.tdm_descriptor<2>
364+
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
365+
func.return %descriptor : !amdgpu.tdm_descriptor
366366
}
367367

368368
// -----

‎mlir/test/Dialect/AMDGPU/amdgpu-make-dma-descriptor-fold.mlir‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -2,7 +2,7 @@
22

33
// CHECK-LABEL: @make_dma_descriptor_fold
44
// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>, %[[IDX:.+]]: index)
5-
func.func @make_dma_descriptor_fold(%base: !amdgpu.tdm_base<i32>, %idx: index) -> !amdgpu.tdm_descriptor<2> {
5+
func.func @make_dma_descriptor_fold(%base: !amdgpu.tdm_base<i32>, %idx: index) -> !amdgpu.tdm_descriptor {
66
%c64 = arith.constant 64 : index
77

88
// CHECK: amdgpu.make_dma_descriptor %[[BASE]]
@@ -14,6 +14,6 @@ func.func @make_dma_descriptor_fold(%base: !amdgpu.tdm_base<i32>, %idx: index) -
1414
// CHECK-SAME: sharedSize [64, 64]
1515
sharedSize [%c64, %c64]
1616
iterate %idx, %idx, %idx
17-
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
18-
func.return %0 : !amdgpu.tdm_descriptor<2>
17+
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
18+
func.return %0 : !amdgpu.tdm_descriptor
1919
}

‎mlir/test/Dialect/AMDGPU/invalid.mlir‎

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -373,7 +373,7 @@ func.func @make_dma_base_invalid_addressspace(%idx: index, %smem : memref<8xi32,
373373

374374
func.func @make_dma_base_invalid_barrier(%base: !amdgpu.tdm_base<i32>, %barrier: memref<8xi32>, %idx: index) {
375375
// expected-error@+1 {{'amdgpu.make_dma_descriptor' op atomic barrier address must be in LDS.}}
376-
amdgpu.make_dma_descriptor %base globalSize [64, 64] globalStride [64, 1] sharedSize [64, 64] atomicBarrier(%barrier[%idx] : memref<8xi32>) : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
376+
amdgpu.make_dma_descriptor %base globalSize [64, 64] globalStride [64, 1] sharedSize [64, 64] atomicBarrier(%barrier[%idx] : memref<8xi32>) : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
377377
}
378378

379379
// -----
@@ -382,7 +382,7 @@ func.func @make_dma_base_invalid_barrier(%base: !amdgpu.tdm_base<i32>, %barrier:
382382
// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>)
383383
func.func @make_dma_descriptor_invalid_empty_strides(%base: !amdgpu.tdm_base<i32>) {
384384
// expected-error@+1 {{'amdgpu.make_dma_descriptor' op strides must not be empty.}}
385-
amdgpu.make_dma_descriptor %base globalSize [0, 1] globalStride [] sharedSize [1, 0] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
385+
amdgpu.make_dma_descriptor %base globalSize [0, 1] globalStride [] sharedSize [1, 0] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
386386
func.return
387387
}
388388

@@ -392,7 +392,7 @@ func.func @make_dma_descriptor_invalid_empty_strides(%base: !amdgpu.tdm_base<i32
392392
// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>)
393393
func.func @make_dma_descriptor_invalid_innermost_stride(%base: !amdgpu.tdm_base<i32>) {
394394
// expected-error@+1 {{'amdgpu.make_dma_descriptor' op strides for the innermost dimension must be 1.}}
395-
amdgpu.make_dma_descriptor %base globalSize [2, 2] globalStride [1, 2] sharedSize [1, 0] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
395+
amdgpu.make_dma_descriptor %base globalSize [2, 2] globalStride [1, 2] sharedSize [1, 0] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
396396
func.return
397397
}
398398

@@ -402,7 +402,7 @@ func.func @make_dma_descriptor_invalid_innermost_stride(%base: !amdgpu.tdm_base<
402402
// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>)
403403
func.func @make_dma_descriptor_invalid_size_and_stride_sizes(%base: !amdgpu.tdm_base<i32>) {
404404
// expected-error@+1 {{'amdgpu.make_dma_descriptor' op strides and sizes must have same rank.}}
405-
amdgpu.make_dma_descriptor %base globalSize [1, 1, 1] globalStride [1, 1] sharedSize [1, 0] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
405+
amdgpu.make_dma_descriptor %base globalSize [1, 1, 1] globalStride [1, 1] sharedSize [1, 0] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
406406
func.return
407407
}
408408

@@ -412,7 +412,7 @@ func.func @make_dma_descriptor_invalid_size_and_stride_sizes(%base: !amdgpu.tdm_
412412
// CHECK-SAME: (%[[BASE:.+]]: !amdgpu.tdm_base<i32>)
413413
func.func @make_dma_descriptor_invalid_shared_and_global_rank(%base: !amdgpu.tdm_base<i32>) {
414414
// expected-error@+1 {{'amdgpu.make_dma_descriptor' op tensor must have same rank as tile.}}
415-
amdgpu.make_dma_descriptor %base globalSize [4, 4] globalStride [1, 1] sharedSize [1, 2, 3] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
415+
amdgpu.make_dma_descriptor %base globalSize [4, 4] globalStride [1, 1] sharedSize [1, 2, 3] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
416416
func.return
417417
}
418418

‎mlir/test/Dialect/AMDGPU/ops.mlir‎

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -706,8 +706,8 @@ func.func @make_dma_descriptor(%base: !amdgpu.tdm_base<i32>, %wg_mask: i16, %tim
706706
globalSize [64, 64]
707707
// CHECK-SAME: globalStride [64, 1]
708708
globalStride [64, 1]
709-
// CHECK-SAME: sharedSize [64, 64] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
710-
sharedSize [64, 64] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
709+
// CHECK-SAME: sharedSize [64, 64] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
710+
sharedSize [64, 64] : !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
711711

712712
// CHECK: amdgpu.make_dma_descriptor %[[BASE]]
713713
amdgpu.make_dma_descriptor %base
@@ -719,7 +719,7 @@ func.func @make_dma_descriptor(%base: !amdgpu.tdm_base<i32>, %wg_mask: i16, %tim
719719
sharedSize [64, 64]
720720
// CHECK-SAME: padShared(%[[IDX]] every %[[IDX]])
721721
padShared(%idx every %idx)
722-
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
722+
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
723723

724724
// CHECK: amdgpu.make_dma_descriptor %[[BASE]]
725725
amdgpu.make_dma_descriptor %base
@@ -757,7 +757,7 @@ func.func @make_dma_descriptor(%base: !amdgpu.tdm_base<i32>, %wg_mask: i16, %tim
757757
sharedSize [64, 64]
758758
// CHECK-SAME: atomicBarrier(%[[BARRIER]][%[[IDX]]] : memref<8xi32, #gpu.address_space<workgroup>>)
759759
atomicBarrier(%barrier[%idx] : memref<8xi32, #gpu.address_space<workgroup>>)
760-
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<2>
760+
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
761761

762762
// CHECK: amdgpu.make_dma_descriptor %[[BASE]]
763763
amdgpu.make_dma_descriptor %base
@@ -769,7 +769,7 @@ func.func @make_dma_descriptor(%base: !amdgpu.tdm_base<i32>, %wg_mask: i16, %tim
769769
sharedSize [64, 64]
770770
// CHECK-SAME: iterate %[[IDX]], %[[IDX]], %[[IDX]]
771771
iterate %idx, %idx, %idx
772-
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor<4>
772+
: !amdgpu.tdm_base<i32> -> !amdgpu.tdm_descriptor
773773

774774
func.return
775775
}

0 commit comments

Comments
 (0)