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

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>