-
Notifications
You must be signed in to change notification settings - Fork 1.6k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[TritonIR] Convert Triton dialect's
Combine
pass to MLIR DRR based (#…
…16)
- Loading branch information
Showing
4 changed files
with
162 additions
and
132 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,6 +1,11 @@ | ||
set(LLVM_TARGET_DEFINITIONS Combine.td) | ||
mlir_tablegen(TritonCombine.inc -gen-rewriters) | ||
add_public_tablegen_target(TritonCombineIncGen) | ||
|
||
add_mlir_dialect_library(TritonTransforms | ||
Combine.cpp | ||
|
||
DEPENDS | ||
TritonTransformsIncGen | ||
TritonCombineIncGen | ||
) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,53 @@ | ||
#ifndef TRITON_PATTERNS | ||
#define TRITON_PATTERNS | ||
|
||
include "mlir/Dialect/StandardOps/IR/Ops.td" | ||
include "mlir/Dialect/Arithmetic/IR/ArithmeticOps.td" | ||
include "triton/Dialect/Triton/IR/TritonOps.td" | ||
|
||
|
||
// AddIOp(DotOp(a, b, c), d) and c==0 => DotOp(a, b, d) | ||
// AddFOp(DotOp(a, b, c), d) and c==0 => DotOp(a, b, d) | ||
|
||
// AddIOp(d, DotOp(a, b, c)) and c==0 => DotOp(a, b, d) | ||
// AddFOp(d, DotOp(a, b, c)) and c==0 => DotOp(a, b, d) | ||
def CombineDotAddIPattern : Pat< | ||
(Arith_AddIOp $d, (TT_DotOp:$res $a, $b, $c, $allowTF32)), | ||
(TT_DotOp $a, $b, $d, $allowTF32), | ||
[(Constraint<CPred<"isZero($0)">> $c)]>; | ||
def CombineDotAddFPattern : Pat< | ||
(Arith_AddFOp $d, (TT_DotOp:$res $a, $b, $c, $allowTF32)), | ||
(TT_DotOp $a, $b, $d, $allowTF32), | ||
[(Constraint<CPred<"isZero($0)">> $c)]>; | ||
|
||
def CombineDotAddIRevPattern : Pat< | ||
(Arith_AddIOp (TT_DotOp:$res $a, $b, $c, $allowTF32), $d), | ||
(TT_DotOp $a, $b, $d, $allowTF32), | ||
[(Constraint<CPred<"isZero($0)">> $c)]>; | ||
def CombineDotAddFRevPattern : Pat< | ||
(Arith_AddFOp (TT_DotOp:$res $a, $b, $c, $allowTF32), $d), | ||
(TT_DotOp $a, $b, $d, $allowTF32), | ||
[(Constraint<CPred<"isZero($0)">> $c)]>; | ||
|
||
|
||
// gep(gep(%ptr, %idx0), %idx1) => gep(%ptr, AddI(%idx0, %idx1)) | ||
// Note: leave (sub %c0, %c0) canceling to ArithmeticDialect | ||
// (ref: ArithmeticCanonicalization.td) | ||
def CombineGEPPattern : Pat< | ||
(TT_GEPOp (TT_GEPOp $ptr, $idx0), $idx1), | ||
(TT_GEPOp $ptr, (Arith_AddIOp $idx0, $idx1))>; | ||
|
||
// select(cond, load(ptrs, broadcast(cond), ???), other) | ||
// => load(ptrs, broadcast(cond), other) | ||
def CombineSelectMaskedLoadPattern : Pat< | ||
(SelectOp $cond, (TT_LoadOp $ptrs, (TT_BroadcastOp:$bcast_res $cond), $other, $cache, $evict, $isVolatile), $falseValue), | ||
(TT_LoadOp $ptrs, $bcast_res, $falseValue, $cache, $evict, $isVolatile)>; | ||
|
||
// broadcast(cst) => cst | ||
def getConstantValue : NativeCodeCall<"getConstantValue($_builder, $0, $1)">; | ||
def CombineBroadcastConstantPattern : Pat< | ||
(TT_BroadcastOp:$bcast_res (Arith_ConstantOp $value)), | ||
(Arith_ConstantOp (getConstantValue $value, $bcast_res)), | ||
[(Constraint<CPred<"isBroadcastConstantCombinable($0)">> $value)]>; | ||
|
||
#endif |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,68 @@ | ||
// RUN: triton-opt %s -split-input-file -canonicalize -triton-combine | ||
// RUN: triton-opt %s -split-input-file -canonicalize -triton-combine | FileCheck %s | ||
|
||
// CHECK-LABEL: @test_combine_dot_add_pattern | ||
func @test_combine_dot_add_pattern() -> (tensor<128x128xf32>, tensor<128x128xf32>) { | ||
// CHECK: %[[d:.*]] = arith.constant dense<3.000000e+00> : tensor<128x128xf32> | ||
// CHECK: %[[b:.*]] = arith.constant dense<2.000000e+00> : tensor<128x128xf32> | ||
// CHECK: %[[a:.*]] = arith.constant dense<1.000000e+00> : tensor<128x128xf32> | ||
%a = arith.constant dense<1.0> : tensor<128x128xf32> | ||
%b = arith.constant dense<2.0> : tensor<128x128xf32> | ||
%zero = arith.constant dense<0.0> : tensor<128x128xf32> | ||
%d = arith.constant dense<3.0> : tensor<128x128xf32> | ||
|
||
%dot_out = tt.dot %a, %b, %zero {allowTF32 = true} : tensor<128x128xf32> * tensor<128x128xf32> -> tensor<128x128xf32> | ||
|
||
// CHECK-NEXT: %[[res0:.*]] = tt.dot %[[a]], %[[b]], %[[d]] {allowTF32 = true} : tensor<128x128xf32> * tensor<128x128xf32> -> tensor<128x128xf32> | ||
%res0 = arith.addf %dot_out, %d : tensor<128x128xf32> | ||
|
||
// CHECK-NEXT: %[[res1:.*]] = tt.dot %[[a]], %[[b]], %[[d]] {allowTF32 = true} : tensor<128x128xf32> * tensor<128x128xf32> -> tensor<128x128xf32> | ||
%res1 = arith.addf %d, %dot_out : tensor<128x128xf32> | ||
|
||
return %res0, %res1 : tensor<128x128xf32>, tensor<128x128xf32> | ||
} | ||
|
||
// CHECK-LABEL: @test_combine_gep_pattern | ||
func @test_combine_gep_pattern(%base: !tt.ptr<f32>) -> tensor<8x!tt.ptr<f32>> { | ||
%off0 = arith.constant 10 : i32 | ||
%off1 = arith.constant 15 : i32 | ||
|
||
// 10 + 15 = 25 | ||
// CHECK-NEXT: %[[cst:.*]] = arith.constant dense<25> : tensor<8xi32> | ||
|
||
%base_ = tt.broadcast %base : (!tt.ptr<f32>) -> tensor<8x!tt.ptr<f32>> | ||
|
||
// CHECK-NEXT: %[[tmp0:.*]] = tt.broadcast %{{.*}} : (!tt.ptr<f32>) -> tensor<8x!tt.ptr<f32>> | ||
|
||
%idx0 = tt.broadcast %off0 : (i32) -> tensor<8xi32> | ||
%idx1 = tt.broadcast %off1 : (i32) -> tensor<8xi32> | ||
|
||
// CHECK-NEXT: %1 = tt.getelementptr %[[tmp0]], %[[cst]] : tensor<8x!tt.ptr<f32>> | ||
%ptr0 = tt.getelementptr %base_, %idx0 : tensor<8x!tt.ptr<f32>> | ||
%ptr1 = tt.getelementptr %ptr0, %idx1 : tensor<8x!tt.ptr<f32>> | ||
|
||
return %ptr1 : tensor<8x!tt.ptr<f32>> | ||
} | ||
|
||
// CHECK-LABEL: @test_combine_select_masked_load_pattern | ||
func @test_combine_select_masked_load_pattern(%ptr: tensor<8x!tt.ptr<f32>>, %cond: i1) -> tensor<8xf32> { | ||
%mask = tt.broadcast %cond : (i1) -> tensor<8xi1> | ||
%false_val = arith.constant dense<0.0> : tensor<8xf32> | ||
|
||
// CHECK: %[[res:.*]] = tt.load %{{.*}}, %{{.*}}, %{{.*}} {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32> | ||
%x = tt.load %ptr, %mask, %false_val {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32> | ||
%0 = select %cond, %x, %false_val : tensor<8xf32> | ||
|
||
// CHECK: return %[[res]] : tensor<8xf32> | ||
return %0 : tensor<8xf32> | ||
} | ||
|
||
// CHECK-LABEL: @test_combine_broadcast_constant_pattern | ||
func @test_combine_broadcast_constant_pattern(%cst : f32) -> tensor<8x2xf32> { | ||
// CHECK: %[[cst:.*]] = arith.constant dense<0.000000e+00> : tensor<8x2xf32> | ||
%const = arith.constant dense<1.0> : tensor<8xf32> | ||
%bst_out = tt.broadcast %const : (tensor<8xf32>) -> tensor<8x2xf32> | ||
|
||
// CHECK-NEXT: return %[[cst]] : tensor<8x2xf32> | ||
return %bst_out : tensor<8x2xf32> | ||
} |