Skip to content

misc GenericOp ODS changes #414

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

Merged
alexbaden merged 2 commits into from
May 26, 2026
Merged

misc GenericOp ODS changes #414

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
59a7c67
Adjust generic assembly format
alexbaden May 22, 2026
0fc2b9d
Introduce GenericOp::toHeader to print generic op info without body
alexbaden May 22, 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
10 changes: 6 additions & 4 deletions cpu/include/Dialect/TritonCPU/IR/TritonCPUOps.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -213,7 +213,7 @@ def TTC_GenericOp : TTC_Op<"generic", [AttrSizedOperandSegments, RecursiveMemory

Example — elementwise chain (no reductions):
```
ttc.generic (%src_ptrs, %dst_ptrs) blocks [%c128 : i32]
ttc.generic ins (%src_ptrs, %dst_ptrs) blocks [%c128 : i32]
{tileShape = array<i32: 16>, reductionDims = array<i32>}
body {
^bb0(%iv: i32, %src: tensor<16x!tt.ptr<f32>>, %dst: tensor<16x!tt.ptr<f32>>):
Expand All @@ -226,7 +226,7 @@ def TTC_GenericOp : TTC_Op<"generic", [AttrSizedOperandSegments, RecursiveMemory

Example — scalar reduction (softmax max pass):
```
%max = ttc.generic (%ptrs, %n_cols) init (%neg_inf : f32)
%max = ttc.generic ins (%ptrs, %n_cols) init (%neg_inf : f32)
blocks [%c128 : i32]
{tileShape = array<i32: 16>, reductionDims = array<i32: 0>}
body {
Expand All @@ -240,7 +240,7 @@ def TTC_GenericOp : TTC_Op<"generic", [AttrSizedOperandSegments, RecursiveMemory

Example — mixed scalar + tensor result (softmax sum + exp):
```
%sum, %exp_out = ttc.generic (%ptrs, %n_cols) init (%zero : f32)
%sum, %exp_out = ttc.generic ins (%ptrs, %n_cols) init (%zero : f32)
blocks [%c128 : i32]
{tileShape = array<i32: 16>, reductionDims = array<i32: 0>}
body {
Expand Down Expand Up @@ -276,7 +276,7 @@ def TTC_GenericOp : TTC_Op<"generic", [AttrSizedOperandSegments, RecursiveMemory
];

let assemblyFormat = [{
`(` $ins `)` (`init` `(` $init_vals^ `:` type($init_vals) `)`)? `blocks` `[` $blockShape `:` type($blockShape) `]` attr-dict-with-keyword `body` $body `:` functional-type($ins, $results)
(`init` `(` $init_vals^ `:` type($init_vals) `)`)? (`ins` `(` $ins^ `:` type($ins) `)`)? `blocks` `[` $blockShape `:` type($blockShape) `]` attr-dict-with-keyword `body` $body (`->` type($results)^)?
}];

let extraClassDeclaration = [{
Expand All @@ -293,6 +293,8 @@ def TTC_GenericOp : TTC_Op<"generic", [AttrSizedOperandSegments, RecursiveMemory
unsigned getInsArgOffset() {
return getNumInductionVars() + getNumIterArgs();
}

std::string getHeader();
}];
}

Expand Down
7 changes: 7 additions & 0 deletions cpu/lib/Dialect/TritonCPU/IR/Ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -135,6 +135,13 @@ LogicalResult GenericOp::verify() {
return success();
}

std::string GenericOp::getHeader() {
std::string s;
llvm::raw_string_ostream os(s);
print(os, OpPrintingFlags().skipRegions());
return s;
}

LogicalResult MakeDynamicRangeOp::verify() {
auto resultTensorTy = cast<RankedTensorType>(getResult().getType());
if (resultTensorTy.getShape().size() != 1)
Expand Down
6 changes: 1 addition & 5 deletions cpu/lib/TritonCPUToLLVM/GenericOpToLLVM.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -776,11 +776,7 @@ struct GenericOpConversion : public ConvertOpToLLVMPattern<cpu::GenericOp> {
vectorSize *= tileShape[d];
}

// TODO: put this into extraClassDefinitions?
std::string s;
llvm::raw_string_ostream os(s);
op->print(os, OpPrintingFlags().skipRegions());
LDBG("Lowering generic op: " << s
LDBG("Lowering generic op: " << op.getHeader()
<< "\n with vectorSize = " << vectorSize);

SmallVector<ArgInfo> argInfos =
Expand Down
13 changes: 7 additions & 6 deletions test/Analysis/axis-info-generic.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,11 +17,11 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shar
%x_3 = tt.addptr %x, %offsets_1 : tensor<512x!tt.ptr<f32>, #blocked>, tensor<512xi32, #blocked>

// CHECK-COUNT-128: ttc.masked_load {{.*}} -> vector<4xf32>
ttc.generic (%x_3) blocks [%c512_i32 : i32] attributes {tileShape = array<i32: 4>, reductionDims = array<i32>} body {
ttc.generic ins (%x_3 : tensor<512x!tt.ptr<f32>, #blocked>) blocks [%c512_i32 : i32] attributes {tileShape = array<i32: 4>, reductionDims = array<i32>} body {
^bb0(%offset:i32, %arg0: tensor<4x!tt.ptr<f32>, #blocked>):
%x_10 = tt.load %arg0 : tensor<4x!tt.ptr<f32>, #blocked>
ttc.yield
}: (tensor<512x!tt.ptr<f32>, #blocked>) -> ()
}
tt.return
}
}
Expand All @@ -32,7 +32,7 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shar
module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shared = 0 : i32, ttg.target = "cpu", "ttg.threads-per-warp" = 1 : i32} {
tt.func public @load_scalar(%x_ptr: !tt.ptr<f32> {tt.divisibility = 16 : i32}) {
%c1024_i32 = arith.constant 1024 : i32
ttc.generic (%x_ptr) blocks [%c1024_i32 : i32] attributes {tileShape = array<i32: 4>, reductionDims = array<i32>} body {
ttc.generic ins (%x_ptr : !tt.ptr<f32>) blocks [%c1024_i32 : i32] attributes {tileShape = array<i32: 4>, reductionDims = array<i32>} body {
^bb0(%tileOffset: i32, %ptr: !tt.ptr<f32>):
%offsets = ttc.make_dynamic_range %tileOffset : tensor<4xi32, #blocked>
%ptrs = tt.splat %ptr : !tt.ptr<f32> -> tensor<4x!tt.ptr<f32>, #blocked>
Expand All @@ -46,7 +46,7 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shar
// CHECK: ttc.masked_load {{.*}} -> vector<4xf32>
%ret = tt.load %offset_ptrs : tensor<4x!tt.ptr<f32>, #blocked>
ttc.yield
}: (!tt.ptr<f32>) -> ()
}

tt.return
}
Expand All @@ -69,7 +69,8 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shar
// CHECK-NOT: ttc.masked_store {{.*}} : (!llvm.ptr<1>, vector<1xf16>
%c4_i32 = arith.constant 4 : i32
%c8_i32 = arith.constant 8 : i32
ttc.generic (%data, %c_ptr, %stride_cm, %M, %N, %offs_am, %offs_bn) blocks
ttc.generic ins (%data, %c_ptr, %stride_cm, %M, %N, %offs_am, %offs_bn :
tensor<4x8xf16, #blocked>, !tt.ptr<f16>, i32, i32, i32, i32, i32) blocks
[%c4_i32, %c8_i32 : i32, i32] attributes {tileShape = array<i32: 1, 8>, reductionDims = array<i32>} body {
^bb0(%tile_m: i32, %tile_n: i32,
%tile_data: tensor<1x8xf16, #blocked>,
Expand Down Expand Up @@ -116,7 +117,7 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shar

tt.store %ptrs, %tile_data, %mask : tensor<1x8x!tt.ptr<f16>, #blocked>
ttc.yield
} : (tensor<4x8xf16, #blocked>, !tt.ptr<f16>, i32, i32, i32, i32, i32) -> ()
}

tt.return
}
Expand Down
16 changes: 8 additions & 8 deletions test/Conversion/generic-op-loops.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,13 +17,13 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shar
tt.func public @scale_1d(%scale: f32) {
%c16_i32 = arith.constant 16 : i32

%0 = ttc.generic (%scale) blocks [%c16_i32 : i32] attributes {tileShape = array<i32: 4>, reductionDims = array<i32>} body {
%0 = ttc.generic ins (%scale : f32) blocks [%c16_i32 : i32] attributes {tileShape = array<i32: 4>, reductionDims = array<i32>} body {
^bb0(%offset: i32, %s: f32):
%cst = arith.constant dense<1.0> : tensor<4xf32, #blocked>
%splat = tt.splat %s : f32 -> tensor<4xf32, #blocked>
%result = arith.mulf %cst, %splat : tensor<4xf32, #blocked>
ttc.yield %result : tensor<4xf32, #blocked>
} : (f32) -> tensor<16xf32, #blocked>
} -> tensor<16xf32, #blocked>
tt.return
}
}
Expand All @@ -40,13 +40,13 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shar
tt.func public @scale_2d(%scale: f32) {
%c4_i32 = arith.constant 4 : i32
%c8_i32 = arith.constant 8 : i32
%0 = ttc.generic (%scale) blocks [%c4_i32, %c8_i32 : i32, i32] attributes {tileShape = array<i32: 1, 4>, reductionDims = array<i32>} body {
%0 = ttc.generic ins (%scale : f32) blocks [%c4_i32, %c8_i32 : i32, i32] attributes {tileShape = array<i32: 1, 4>, reductionDims = array<i32>} body {
^bb0(%dim0: i32, %dim1: i32, %s: f32):
%cst = arith.constant dense<1.0> : tensor<1x4xf32, #blocked>
%splat = tt.splat %s : f32 -> tensor<1x4xf32, #blocked>
%result = arith.mulf %cst, %splat : tensor<1x4xf32, #blocked>
ttc.yield %result : tensor<1x4xf32, #blocked>
} : (f32) -> tensor<4x8xf32, #blocked>
} -> tensor<4x8xf32, #blocked>
tt.return
}
}
Expand All @@ -63,13 +63,13 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shar
tt.func public @scale_3d(%scale: f32) {
%c4_i32 = arith.constant 4 : i32
%c8_i32 = arith.constant 8 : i32
%0 = ttc.generic (%scale) blocks [%c4_i32, %c4_i32, %c8_i32 : i32, i32, i32] attributes {tileShape = array<i32: 1, 1, 4>, reductionDims = array<i32>} body {
%0 = ttc.generic ins (%scale : f32) blocks [%c4_i32, %c4_i32, %c8_i32 : i32, i32, i32] attributes {tileShape = array<i32: 1, 1, 4>, reductionDims = array<i32>} body {
^bb0(%dim0: i32, %dim1: i32, %dim2: i32, %s: f32):
%cst = arith.constant dense<1.0> : tensor<1x1x4xf32, #blocked>
%splat = tt.splat %s : f32 -> tensor<1x1x4xf32, #blocked>
%result = arith.mulf %cst, %splat : tensor<1x1x4xf32, #blocked>
ttc.yield %result : tensor<1x1x4xf32, #blocked>
} : (f32) -> tensor<4x4x8xf32, #blocked>
} -> tensor<4x4x8xf32, #blocked>
tt.return
}
}
Expand All @@ -87,13 +87,13 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, ttg.shar
%c2_i32 = arith.constant 2 : i32
%c4_i32 = arith.constant 4 : i32
%c8_i32 = arith.constant 8 : i32
%0 = ttc.generic (%scale) blocks [%c2_i32, %c2_i32, %c4_i32, %c8_i32 : i32, i32, i32, i32] attributes {tileShape = array<i32: 1, 1, 1, 4>, reductionDims = array<i32>} body {
%0 = ttc.generic ins (%scale : f32) blocks [%c2_i32, %c2_i32, %c4_i32, %c8_i32 : i32, i32, i32, i32] attributes {tileShape = array<i32: 1, 1, 1, 4>, reductionDims = array<i32>} body {
^bb0(%dim0: i32, %dim1: i32, %dim2: i32, %dim3: i32, %s: f32):
%cst = arith.constant dense<1.0> : tensor<1x1x1x4xf32, #blocked>
%splat = tt.splat %s : f32 -> tensor<1x1x1x4xf32, #blocked>
%result = arith.mulf %cst, %splat : tensor<1x1x1x4xf32, #blocked>
ttc.yield %result : tensor<1x1x1x4xf32, #blocked>
} : (f32) -> tensor<2x2x4x8xf32, #blocked>
} -> tensor<2x2x4x8xf32, #blocked>
tt.return
}
}
Loading