Last active
October 11, 2023 13:53
-
-
Save matthias-springer/b664feb23be0159f72726025923bb9ca to your computer and use it in GitHub Desktop.
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
| // RUN: mlir-opt %s -test-transform-dialect-interpreter="debug-payload-root-tag=payload" -test-transform-dialect-erase-schedule | |
| #map = affine_map<(d0) -> (-d0 + 17, 5)> | |
| #map1 = affine_map<(d0) -> (-d0 + 29, 7)> | |
| module attributes {transform.target_tag="payload"} { | |
| // Tiled batched matrix multiplication. %A and %B are the inputs, %C is the | |
| // output. | |
| func.func @test_matmul(%A: memref<1x17x19xf32>, %B: memref<1x19x29xf32>, | |
| %C: memref<1x17x29xf32>) { | |
| %c7 = arith.constant 7 : index | |
| %c29 = arith.constant 29 : index | |
| %c5 = arith.constant 5 : index | |
| %c17 = arith.constant 17 : index | |
| %c0 = arith.constant 0 : index | |
| %cst = arith.constant 0.000000e+00 : f32 | |
| %A_tensor = bufferization.to_tensor %A restrict : memref<1x17x19xf32> | |
| %B_tensor = bufferization.to_tensor %B restrict : memref<1x19x29xf32> | |
| %C_tensor = bufferization.to_tensor %C restrict writable : memref<1x17x29xf32> | |
| %3 = linalg.fill ins(%cst : f32) outs(%C_tensor : tensor<1x17x29xf32>) -> tensor<1x17x29xf32> | |
| %4 = scf.for %arg3 = %c0 to %c17 step %c5 iter_args(%arg4 = %3) -> (tensor<1x17x29xf32>) { | |
| %5 = affine.min #map(%arg3) | |
| %6 = scf.for %arg5 = %c0 to %c29 step %c7 iter_args(%arg6 = %arg4) -> (tensor<1x17x29xf32>) { | |
| %7 = affine.min #map1(%arg5) | |
| // Let's assume that something went wrong during tiling. Due to a bug, we | |
| // extract from %arg4 instead of %arg6. | |
| %extracted_slice = tensor.extract_slice %arg4[0, %arg3, %arg5] [1, %5, %7] [1, 1, 1] : tensor<1x17x29xf32> to tensor<1x?x?xf32> | |
| %extracted_slice_0 = tensor.extract_slice %A_tensor[0, %arg3, 0] [1, %5, 19] [1, 1, 1] : tensor<1x17x19xf32> to tensor<1x?x19xf32> | |
| %extracted_slice_1 = tensor.extract_slice %B_tensor[0, 0, %arg5] [1, 19, %7] [1, 1, 1] : tensor<1x19x29xf32> to tensor<1x19x?xf32> | |
| %8 = linalg.batch_matmul ins(%extracted_slice_0, %extracted_slice_1 : tensor<1x?x19xf32>, tensor<1x19x?xf32>) outs(%extracted_slice : tensor<1x?x?xf32>) -> tensor<1x?x?xf32> | |
| %inserted_slice = tensor.insert_slice %8 into %arg6[0, %arg3, %arg5] [1, %5, %7] [1, 1, 1] : tensor<1x?x?xf32> into tensor<1x17x29xf32> | |
| scf.yield %inserted_slice : tensor<1x17x29xf32> | |
| } | |
| scf.yield %6 : tensor<1x17x29xf32> | |
| } | |
| bufferization.materialize_in_destination %4 in restrict writable %C : (tensor<1x17x29xf32>, memref<1x17x29xf32>) -> () | |
| return | |
| } | |
| } | |
| transform.sequence failures(propagate) { | |
| ^bb0(%module1: !transform.any_op): | |
| // Tile matmul | |
| // ... | |
| %func2 = transform.structured.match ops{["func.func"]} in %module1 | |
| : (!transform.any_op) -> !transform.any_op | |
| // Bufferize | |
| %module2 = transform.bufferization.one_shot_bufferize %module1 | |
| {allow_return_allocs_from_loops = true} | |
| : (!transform.any_op) -> !transform.any_op | |
| // Canonicalize + CSE | |
| transform.apply_patterns to %module2 { | |
| transform.apply_patterns.canonicalization | |
| } {apply_cse} : !transform.any_op | |
| // Deallocate buffers. That's a no-op in this example because no temporary | |
| // buffers are allocated. | |
| %module3 = transform.apply_registered_pass | |
| "buffer-deallocation-pipeline" to %module2 | |
| : (!transform.any_op) -> !transform.any_op | |
| } |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment