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

[Triton-IR] Added type inference and verifier for Triton-IR operations (#767)

Browse Source
This commit is contained in:
Philippe Tillet
2022-10-11 18:16:41 -07:00
committed by GitHub
parent b6e5a231e5
commit 623c99609f
27 changed files with 494 additions and 348 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
test/Analysis/test-allocation.mlir
View File

@@ -1,6 +1,7 @@
// RUN: triton-opt %s -split-input-file --mlir-disable-threading -test-print-allocation 2>&1 | FileCheck %s
#AL = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [4, 8], warpsPerCTA = [4, 1], order = [1, 0]}>
#sliceAd0 = #triton_gpu.slice<{dim = 0, parent = #AL}>
#BL = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [1, 32], warpsPerCTA = [4, 1], order = [1, 0]}>
#A = #triton_gpu.shared<{vec = 2, perPhase = 2, maxPhase = 4, order = [1, 0]}>
#B = #triton_gpu.shared<{vec = 2, perPhase = 2, maxPhase = 4, order = [1, 0]}>
@@ -164,7 +165,7 @@ func @alloc(%A : !tt.ptr<f16>) {
func @scratch() {
%cst0 = arith.constant dense<0.000000e+00> : tensor<16x16xf16, #AL>
// CHECK: scratch offset = 0, size = 512
%b = tt.reduce %cst0 {redOp = 1 : i32, axis = 0 : i32} : tensor<16x16xf16, #AL> -> tensor<16xf16, #AL>
%b = tt.reduce %cst0 {redOp = 1 : i32, axis = 0 : i32} : tensor<16x16xf16, #AL> -> tensor<16xf16, #sliceAd0>
return
// CHECK-NEXT: size = 512
}

4
test/Analysis/test-membar.mlir
View File

@@ -1,6 +1,7 @@
// RUN: triton-opt %s -split-input-file --mlir-disable-threading -test-print-membar 2>&1 | FileCheck %s
#AL = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [4, 8], warpsPerCTA = [4, 1], order = [1, 0]}>
#sliceAd0 = #triton_gpu.slice<{dim = 0, parent = #AL}>
#BL = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [1, 32], warpsPerCTA = [4, 1], order = [1, 0]}>
#A = #triton_gpu.shared<{vec = 2, perPhase = 2, maxPhase = 4, order = [1, 0]}>
#B = #triton_gpu.shared<{vec = 2, perPhase = 2, maxPhase = 4, order = [1, 0]}>
@@ -69,7 +70,8 @@ func @scratch() {
// CHECK: Membar 1
%a = tt.cat %cst0, %cst0 {axis = 0} : (tensor<16x16xf16, #A>, tensor<16x16xf16, #A>) -> tensor<32x16xf16, #A>
// CHECK-NEXT: Membar 3
%b = tt.reduce %a {redOp = 1 : i32, axis = 0 : i32} : tensor<32x16xf16, #A> -> tensor<16xf16, #A>
%aa = triton_gpu.convert_layout %a : (tensor<32x16xf16, #A>) -> tensor<32x16xf16, #AL>
%b = tt.reduce %aa {redOp = 1 : i32, axis = 0 : i32} : tensor<32x16xf16, #AL> -> tensor<16xf16, #sliceAd0>
return
}

30
test/Conversion/tritongpu_to_llvm.mlir
View File

@@ -348,15 +348,17 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
#block1 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [8], warpsPerCTA = [4], order = [0]}>
#block2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [4, 1], warpsPerCTA = [4, 1], order = [1, 0]}>
#block3 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 8], warpsPerCTA = [1, 4], order = [1, 0]}>
#slice2d1 = #triton_gpu.slice<{dim = 1, parent=#block2}>
#slice3d0 = #triton_gpu.slice<{dim = 0, parent=#block3}>
#AL = #triton_gpu.blocked<{sizePerThread = [1, 8], threadsPerWarp = [4, 8], warpsPerCTA = [4, 1], order = [1, 0]}>
#A = #triton_gpu.shared<{vec = 4, perPhase = 1, maxPhase = 4, order = [1, 0]}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
// CHECK-LABEL: basic_insert_slice_async_v4
func @basic_insert_slice_async_v4(%arg0: !tt.ptr<f32> {tt.divisibility = 4 : i32}) {
%off0_ = tt.make_range {end = 16 : i32, start = 0 : i32} : tensor<16xi32, #block0>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<64xi32, #block1>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<16xi32, #block0>) -> tensor<16x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<64xi32, #block1>) -> tensor<1x64xi32, #block3>
%off0_ = tt.make_range {end = 16 : i32, start = 0 : i32} : tensor<16xi32, #slice2d1>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<64xi32, #slice3d0>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<16xi32, #slice2d1>) -> tensor<16x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<64xi32, #slice3d0>) -> tensor<1x64xi32, #block3>
%broadcast_off0_scalar = tt.broadcast %off0 : (tensor<16x1xi32, #block2>) -> tensor<16x64xi32, #block2>
%cst_scalar = arith.constant 64 : i32
%cst = tt.splat %cst_scalar : (i32) -> tensor<16x64xi32, #block2>
@@ -387,15 +389,17 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
#block1 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [8], warpsPerCTA = [4], order = [0]}>
#block2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [4, 1], warpsPerCTA = [4, 1], order = [1, 0]}>
#block3 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 8], warpsPerCTA = [1, 4], order = [1, 0]}>
#slice2d1 = #triton_gpu.slice<{dim = 1, parent=#block2}>
#slice3d0 = #triton_gpu.slice<{dim = 0, parent=#block3}>
#AL = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [4, 8], warpsPerCTA = [4, 1], order = [1, 0]}>
#A = #triton_gpu.shared<{vec = 1, perPhase = 1, maxPhase = 4, order = [1, 0]}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
// CHECK-LABEL: basic_insert_slice_async_v1
func @basic_insert_slice_async_v1(%arg0: !tt.ptr<f32> {tt.divisibility = 4 : i32}) {
%off0_ = tt.make_range {end = 16 : i32, start = 0 : i32} : tensor<16xi32, #block0>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #block1>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<16xi32, #block0>) -> tensor<16x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<32xi32, #block1>) -> tensor<1x32xi32, #block3>
%off0_ = tt.make_range {end = 16 : i32, start = 0 : i32} : tensor<16xi32, #slice2d1>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #slice3d0>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<16xi32, #slice2d1>) -> tensor<16x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<32xi32, #slice3d0>) -> tensor<1x32xi32, #block3>
%broadcast_off0_scalar = tt.broadcast %off0 : (tensor<16x1xi32, #block2>) -> tensor<16x32xi32, #block2>
%cst_scalar = arith.constant 32 : i32
%cst = tt.splat %cst_scalar : (i32) -> tensor<16x32xi32, #block2>
@@ -429,15 +433,17 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
#block0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [8], warpsPerCTA = [4], order = [0]}>
#block2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [8, 1], warpsPerCTA = [4, 1], order = [1, 0]}>
#block3 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 8], warpsPerCTA = [1, 4], order = [1, 0]}>
#slice2d1 = #triton_gpu.slice<{dim = 1, parent=#block2}>
#slice3d0 = #triton_gpu.slice<{dim = 0, parent=#block3}>
#AL = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [4, 8], warpsPerCTA = [4, 1], order = [1, 0]}>
#A = #triton_gpu.shared<{vec = 1, perPhase = 1, maxPhase = 4, order = [1, 0]}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
// CHECK-LABEL: basic_insert_slice_async_v1_multictas
func @basic_insert_slice_async_v1_multictas(%arg0: !tt.ptr<f32> {tt.divisibility = 4 : i32}) {
%off0_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #block0>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #block0>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<32xi32, #block0>) -> tensor<32x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<32xi32, #block0>) -> tensor<1x32xi32, #block3>
%off0_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #slice2d1>
%off1_ = tt.make_range {end = 32 : i32, start = 0 : i32} : tensor<32xi32, #slice3d0>
%off0 = tt.expand_dims %off0_ {axis = 1 : i32} : (tensor<32xi32, #slice2d1>) -> tensor<32x1xi32, #block2>
%off1 = tt.expand_dims %off1_ {axis = 0 : i32} : (tensor<32xi32, #slice3d0>) -> tensor<1x32xi32, #block3>
%broadcast_off0_scalar = tt.broadcast %off0 : (tensor<32x1xi32, #block2>) -> tensor<32x32xi32, #block2>
%cst_scalar = arith.constant 32 : i32
%cst = tt.splat %cst_scalar : (i32) -> tensor<32x32xi32, #block2>

11
test/TritonGPU/coalesce.mlir
View File

@@ -1,8 +1,10 @@
// RUN: triton-opt %s -split-input-file -tritongpu-coalesce -canonicalize -tritongpu-verifier | FileCheck %s
// RUN: triton-opt %s -split-input-file -tritongpu-coalesce -canonicalize | FileCheck %s
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [32, 1], warpsPerCTA = [4, 1], order = [0, 1]}>
#blocked2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 32], warpsPerCTA = [1, 4], order = [0, 1]}>
#slice1dim1 = #triton_gpu.slice<{dim = 1, parent = #blocked1}>
#slice2dim0 = #triton_gpu.slice<{dim = 0, parent = #blocked2}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
@@ -23,13 +25,14 @@ func @transpose(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%arg3: i32 {tt.divisibility = 16 : i32}) {
%cst = arith.constant dense<true> : tensor<64x64xi1, #blocked1>
%cst_0 = arith.constant dense<0.000000e+00> : tensor<64x64xf32, #blocked1>
%0 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #blocked0>
%1 = tt.expand_dims %0 {axis = 1 : i32} : (tensor<64xi32, #blocked0>) -> tensor<64x1xi32, #blocked1>
%00 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #slice1dim1>
%01 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #slice2dim0>
%1 = tt.expand_dims %00 {axis = 1 : i32} : (tensor<64xi32, #slice1dim1>) -> tensor<64x1xi32, #blocked1>
%2 = tt.splat %arg1 : (i32) -> tensor<64x1xi32, #blocked1>
%3 = arith.muli %1, %2 : tensor<64x1xi32, #blocked1>
%4 = tt.splat %arg0 : (!tt.ptr<f32>) -> tensor<64x1x!tt.ptr<f32>, #blocked1>
%5 = tt.addptr %4, %3 : tensor<64x1x!tt.ptr<f32>, #blocked1>
%6 = tt.expand_dims %0 {axis = 0 : i32} : (tensor<64xi32, #blocked0>) -> tensor<1x64xi32, #blocked2>
%6 = tt.expand_dims %01 {axis = 0 : i32} : (tensor<64xi32, #slice2dim0>) -> tensor<1x64xi32, #blocked2>
%7 = tt.broadcast %5 : (tensor<64x1x!tt.ptr<f32>, #blocked1>) -> tensor<64x64x!tt.ptr<f32>, #blocked1>
%8 = tt.broadcast %6 : (tensor<1x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked2>
%9 = triton_gpu.convert_layout %8 : (tensor<64x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked1>

16
test/TritonGPU/combine.mlir
View File

@@ -53,7 +53,9 @@ func @remat(%arg0: i32) -> tensor<1024xi32, #layout1> {
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [32, 1], warpsPerCTA = [4, 1], order = [0, 1]}>
#slice1dim1 = #triton_gpu.slice<{dim = 1, parent = #blocked1}>
#blocked2 = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 32], warpsPerCTA = [1, 4], order = [0, 1]}>
#slice2dim0 = #triton_gpu.slice<{dim = 0, parent = #blocked2}>
#blocked3 = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [2, 16], warpsPerCTA = [1, 4], order = [1, 0]}>
#blocked4 = #triton_gpu.blocked<{sizePerThread = [4, 1], threadsPerWarp = [16, 2], warpsPerCTA = [4, 1], order = [0, 1]}>
@@ -90,13 +92,14 @@ func @transpose(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: i32 {tt
// CHECK: return
%cst = arith.constant dense<0.000000e+00> : tensor<64x64xf32, #blocked1>
%cst_0 = arith.constant dense<true> : tensor<64x64xi1, #blocked1>
%0 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #blocked0>
%1 = tt.expand_dims %0 {axis = 1 : i32} : (tensor<64xi32, #blocked0>) -> tensor<64x1xi32, #blocked1>
%00 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #slice1dim1>
%01 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #slice2dim0>
%1 = tt.expand_dims %00 {axis = 1 : i32} : (tensor<64xi32, #slice1dim1>) -> tensor<64x1xi32, #blocked1>
%2 = tt.splat %arg1 : (i32) -> tensor<64x1xi32, #blocked1>
%3 = arith.muli %1, %2 : tensor<64x1xi32, #blocked1>
%4 = tt.splat %arg0 : (!tt.ptr<f32>) -> tensor<64x1x!tt.ptr<f32>, #blocked1>
%5 = tt.addptr %4, %3 : tensor<64x1x!tt.ptr<f32>, #blocked1>
%6 = tt.expand_dims %0 {axis = 0 : i32} : (tensor<64xi32, #blocked0>) -> tensor<1x64xi32, #blocked2>
%6 = tt.expand_dims %01 {axis = 0 : i32} : (tensor<64xi32, #slice2dim0>) -> tensor<1x64xi32, #blocked2>
%7 = tt.broadcast %5 : (tensor<64x1x!tt.ptr<f32>, #blocked1>) -> tensor<64x64x!tt.ptr<f32>, #blocked1>
%8 = tt.broadcast %6 : (tensor<1x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked2>
%9 = triton_gpu.convert_layout %8 : (tensor<64x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked1>
@@ -138,13 +141,14 @@ func @loop(%arg0: !tt.ptr<f32>, %arg1: i32, %arg2: !tt.ptr<f32>, %arg3: i32, %ar
%c32 = arith.constant 32 : index
%c0 = arith.constant 0 : index
%cst_1 = arith.constant dense<0.000000e+00> : tensor<64x64xf32, #blocked1>
%0 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #blocked0>
%1 = tt.expand_dims %0 {axis = 1 : i32} : (tensor<64xi32, #blocked0>) -> tensor<64x1xi32, #blocked1>
%00 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #slice1dim1>
%01 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32, #slice2dim0>
%1 = tt.expand_dims %00 {axis = 1 : i32} : (tensor<64xi32, #slice1dim1>) -> tensor<64x1xi32, #blocked1>
%2 = tt.splat %arg1 : (i32) -> tensor<64x1xi32, #blocked1>
%3 = arith.muli %1, %2 : tensor<64x1xi32, #blocked1>
%4 = tt.splat %arg0 : (!tt.ptr<f32>) -> tensor<64x1x!tt.ptr<f32>, #blocked1>
%5 = tt.addptr %4, %3 : tensor<64x1x!tt.ptr<f32>, #blocked1>
%6 = tt.expand_dims %0 {axis = 0 : i32} : (tensor<64xi32, #blocked0>) -> tensor<1x64xi32, #blocked2>
%6 = tt.expand_dims %01 {axis = 0 : i32} : (tensor<64xi32, #slice2dim0>) -> tensor<1x64xi32, #blocked2>
%7 = tt.broadcast %5 : (tensor<64x1x!tt.ptr<f32>, #blocked1>) -> tensor<64x64x!tt.ptr<f32>, #blocked1>
%8 = tt.broadcast %6 : (tensor<1x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked2>
%9 = triton_gpu.convert_layout %8 : (tensor<64x64xi32, #blocked2>) -> tensor<64x64xi32, #blocked1>

2
test/TritonGPU/loop-pipeline.mlir
View File

@@ -1,4 +1,4 @@
// RUN: triton-opt %s -split-input-file -tritongpu-pipeline=num-stages=3 -canonicalize -tritongpu-verifier | FileCheck %s
// RUN: triton-opt %s -split-input-file -tritongpu-pipeline=num-stages=3 -canonicalize | FileCheck %s
// 4 warps
// matmul: 128x32 @ 32x128 -> 128x128