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

[OPTIMIZER] Improved layout conversion simplification algorithm (#97)

Browse Source 
This PR both simplifies the layout conversion simplification algorithm, and also improves it to make it work with vectorized element-wise ops. The conversion optimizer still has a lot of room for improvements, and other PRs will address its limitations (ideally via some sort of explicit cost model)
This commit is contained in:
Philippe Tillet
2022-09-02 16:52:44 -07:00
committed by GitHub
parent 3c635449e5
commit d0b4c67b05
2 changed files with 75 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
test/TritonGPU/combine.mlir
View File

@@ -173,3 +173,34 @@ func @loop(%arg0: !tt.ptr<f32>, %arg1: i32, %arg2: !tt.ptr<f32>, %arg3: i32, %ar
tt.store %20, %21, %22 : tensor<64x64xf32, #blocked1>
return
}
// CHECK-LABEL: vecadd
func @vecadd(%arg0: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg1: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg2: !tt.ptr<f32> {tt.divisibility = 16 : i32}, %arg3: i32) {
// CHECK-NOT: triton_gpu.convert_layout
%c256_i32 = arith.constant 256 : i32
%0 = tt.get_program_id {axis = 0 : i32} : i32
%1 = arith.muli %0, %c256_i32 : i32
%2 = tt.splat %1 : (i32) -> tensor<256xi32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%3 = tt.make_range {end = 256 : i32, start = 0 : i32} : tensor<256xi32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%4 = tt.splat %1 : (i32) -> tensor<256xi32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%5 = tt.make_range {end = 256 : i32, start = 0 : i32} : tensor<256xi32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%6 = tt.splat %1 : (i32) -> tensor<256xi32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%7 = tt.make_range {end = 256 : i32, start = 0 : i32} : tensor<256xi32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%8 = tt.splat %arg0 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%9 = arith.addi %6, %7 : tensor<256xi32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%10 = tt.splat %arg1 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%11 = arith.addi %4, %5 : tensor<256xi32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%12 = tt.getelementptr %8, %9 : tensor<256x!tt.ptr<f32>, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%13 = tt.load %12 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%14 = triton_gpu.convert_layout %13 : (tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>) -> tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [2], order = [0]}>>
%15 = tt.getelementptr %10, %11 : tensor<256x!tt.ptr<f32>, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%16 = tt.load %15 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%17 = triton_gpu.convert_layout %16 : (tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>) -> tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [2], order = [0]}>>
%18 = arith.addf %14, %17 : tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [2], order = [0]}>>
%19 = tt.splat %arg2 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%20 = arith.addi %2, %3 : tensor<256xi32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%21 = tt.getelementptr %19, %20 : tensor<256x!tt.ptr<f32>, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
%22 = triton_gpu.convert_layout %18 : (tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [2], order = [0]}>>) -> tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
tt.store %21, %22 : tensor<256xf32, #triton_gpu.blocked<{sizePerThread = [4], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>>
return
}