gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[FRONTEND][BACKEND] Fixes for cat / reshape / addptr (#959)

Browse Source 
Most notably, this PR:
- changes the traits (and assembly format) of addptr so it can handle offsets that have arbitrary integer width.
- adds support for `cat`
This commit is contained in:
Philippe Tillet
2022-12-06 23:29:50 -08:00
committed by GitHub
parent 981aee7f1e
commit b2b793dfb5
24 changed files with 199 additions and 132 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
test/Triton/combine.mlir
View File

@@ -22,28 +22,30 @@ func @test_combine_dot_add_pattern() -> (tensor<128x128xf32>, tensor<128x128xf32
return %res0, %res1 : tensor<128x128xf32>, tensor<128x128xf32>
}
// CHECK-LABEL: @test_combine_addptr_pattern
// COM: CHECK-LABEL: @test_combine_addptr_pattern
func @test_combine_addptr_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>
// COM: 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>>
// COM: 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.addptr %[[tmp0]], %[[cst]] : tensor<8x!tt.ptr<f32>>
%ptr0 = tt.addptr %base_, %idx0 : tensor<8x!tt.ptr<f32>>
%ptr1 = tt.addptr %ptr0, %idx1 : tensor<8x!tt.ptr<f32>>
// COM: CHECK-NEXT: %1 = tt.addptr %[[tmp0]], %[[cst]] : tensor<8x!tt.ptr<f32>>, tensor<8xi32>
%ptr0 = tt.addptr %base_, %idx0 : tensor<8x!tt.ptr<f32>>, tensor<8xi32>
%ptr1 = tt.addptr %ptr0, %idx1 : tensor<8x!tt.ptr<f32>>, tensor<8xi32>
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>, tensor<8xf32>) {
%mask = tt.broadcast %cond : (i1) -> tensor<8xi1>

36
test/Triton/vecadd.mlir
View File

@@ -11,9 +11,9 @@ module {
%5 = tt.broadcast %arg3 : (i32) -> tensor<256xi32>
%6 = arith.cmpi slt, %4, %5 : tensor<256xi32>
%7 = tt.broadcast %arg0 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>>
%8 = tt.addptr %7, %4 : tensor<256x!tt.ptr<f32>>
%8 = tt.addptr %7, %4 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
%9 = tt.broadcast %arg1 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>>
%10 = tt.addptr %9, %4 : tensor<256x!tt.ptr<f32>>
%10 = tt.addptr %9, %4 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
%cst = arith.constant 0.000000e+00 : f32
%11 = tt.broadcast %cst : (f32) -> tensor<256xf32>
%c0_i32 = arith.constant 0 : i32
@@ -31,13 +31,13 @@ module {
%22 = arith.addf %19, %21 : tensor<256xf32>
%23 = arith.addf %arg7, %22 : tensor<256xf32>
%24 = tt.broadcast %arg5 : (i32) -> tensor<256xi32>
%25 = tt.addptr %arg8, %24 : tensor<256x!tt.ptr<f32>>
%25 = tt.addptr %arg8, %24 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
%26 = tt.broadcast %arg5 : (i32) -> tensor<256xi32>
%27 = tt.addptr %arg9, %26 : tensor<256x!tt.ptr<f32>>
%27 = tt.addptr %arg9, %26 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
scf.yield %23, %25, %27 : tensor<256xf32>, tensor<256x!tt.ptr<f32>>, tensor<256x!tt.ptr<f32>>
}
%16 = tt.broadcast %arg2 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>>
%17 = tt.addptr %16, %4 : tensor<256x!tt.ptr<f32>>
%17 = tt.addptr %16, %4 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
tt.store %17, %15#0, %6 : tensor<256xf32>
return
}
@@ -57,9 +57,9 @@ module {
// %5 = tt.broadcast %arg3 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %6 = "triton_gpu.cmpi"(%4, %5) {predicate = 2 : i64} : (tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %7 = tt.broadcast %arg0 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %8 = tt.addptr %7, %4, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %8 = tt.addptr %7, %4, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %9 = tt.broadcast %arg1 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %10 = tt.addptr %9, %4, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %10 = tt.addptr %9, %4, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %11 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %12 = arith.index_cast %arg4 : i32 to index
// %13 = arith.cmpi slt, %c0, %12 : index
@@ -72,9 +72,9 @@ module {
// %20 = arith.andi %6, %19 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %21 = triton_gpu.copy_async %10, %20, %18 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %22 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %23 = tt.addptr %8, %22, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %23 = tt.addptr %8, %22, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %24 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %25 = tt.addptr %10, %24, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %25 = tt.addptr %10, %24, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %26 = arith.cmpi slt, %c32, %12 : index
// %27 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %28 = tt.broadcast %26 : (i1) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
@@ -85,9 +85,9 @@ module {
// %33 = arith.andi %6, %32 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %34 = triton_gpu.copy_async %25, %33, %31 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %35 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %36 = tt.addptr %23, %35, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %36 = tt.addptr %23, %35, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %37 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %38 = tt.addptr %25, %37, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %38 = tt.addptr %25, %37, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %39 = arith.cmpi slt, %c64, %12 : index
// %40 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %41 = tt.broadcast %39 : (i1) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
@@ -98,16 +98,16 @@ module {
// %46 = arith.andi %6, %45 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %47 = triton_gpu.copy_async %38, %46, %44 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %48 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %49 = tt.addptr %36, %48, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %49 = tt.addptr %36, %48, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %50 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %51 = tt.addptr %38, %50, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %51 = tt.addptr %38, %50, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %52:12 = scf.for %arg6 = %c0 to %12 step %c32 iter_args(%arg7 = %11, %arg8 = %8, %arg9 = %10, %arg10 = %17, %arg11 = %30, %arg12 = %43, %arg13 = %21, %arg14 = %34, %arg15 = %47, %arg16 = %51, %arg17 = %49, %arg18 = %c64) -> (tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, index) {
// %55 = arith.addf %arg10, %arg13 : tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %56 = arith.addf %arg7, %55 : tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %57 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %58 = tt.addptr %arg8, %57, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %58 = tt.addptr %arg8, %57, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %59 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %60 = tt.addptr %arg9, %59, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %60 = tt.addptr %arg9, %59, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %61 = arith.addi %arg18, %c32 : index
// %62 = arith.cmpi slt, %61, %12 : index
// %63 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
@@ -117,13 +117,13 @@ module {
// %67 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %68 = triton_gpu.copy_async %arg16, %65, %67 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %69 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %70 = tt.addptr %arg17, %69, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %70 = tt.addptr %arg17, %69, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// %71 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %72 = tt.addptr %arg16, %71, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %72 = tt.addptr %arg16, %71, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// scf.yield %56, %58, %60, %arg11, %arg12, %66, %arg14, %arg15, %68, %72, %70, %61 : tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, index
// }
// %53 = tt.broadcast %arg2 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %54 = tt.addptr %53, %4, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// %54 = tt.addptr %53, %4, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
// tt.store %54, %52#0, %6 : tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
// return
// }