-
Notifications
You must be signed in to change notification settings - Fork 31
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[BufferizeToAllocation] Handle pack operands dependent on other op ty…
…pes (#972)
- Loading branch information
Showing
7 changed files
with
164 additions
and
41 deletions.
There are no files selected for viewing
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
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
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
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
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
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
111 changes: 111 additions & 0 deletions
111
...ler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/test/bufferize_to_allocation_pack.mlir
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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,111 @@ | ||
| // RUN: iree-opt --pass-pipeline='builtin.module(func.func(iree-amdaie-bufferize-to-allocation{memory-space=1 bufferize-operand=pack-input pack-depth=2}))' --split-input-file --verify-diagnostics %s | FileCheck %s | ||
|
|
||
| // Checks for packs with interleaved extract slices. | ||
| // CHECK-LABEL: @matmul_tensor_extract_slice | ||
| // CHECK: scf.forall | ||
| // CHECK: %[[ALLOC_0:.+]] = memref.alloc() : memref<4x8x32x64xbf16, 1 : i32> | ||
| // CHECK: %[[TO_TENSOR_0:.+]] = bufferization.to_tensor %[[ALLOC_0]] | ||
| // CHECK: %[[PACK_0:.+]] = tensor.pack | ||
| // CHECK-SAME: into %[[TO_TENSOR_0]] | ||
| // CHECK: %[[ALLOC_1:.+]] = memref.alloc() : memref<4x8x64x32xbf16, 1 : i32> | ||
| // CHECK: %[[TO_TENSOR_1:.+]] = bufferization.to_tensor %[[ALLOC_1]] | ||
| // CHECK: %[[PACK_1:.+]] = tensor.pack | ||
| // CHECK-SAME: into %[[TO_TENSOR_1]] | ||
| // CHECK: scf.forall | ||
| // CHECK: %[[SLICE_0:.+]] = tensor.extract_slice %[[PACK_0]] | ||
| // CHECK: %[[SLICE_1:.+]] = tensor.extract_slice %[[PACK_1]] | ||
| // CHECK: linalg.fill | ||
| // CHECK: scf.for | ||
| // CHECK: %[[SLICE_2:.+]] = tensor.extract_slice %[[SLICE_0]] | ||
| // CHECK: %[[PACK_2:.+]] = tensor.pack %[[SLICE_2]] | ||
| // CHECK: %[[SLICE_3:.+]] = tensor.extract_slice %[[SLICE_1]] | ||
| // CHECK: %[[PACK_3:.+]] = tensor.pack %[[SLICE_3]] | ||
| // CHECK: linalg.generic | ||
| func.func @matmul_tensor_extract_slice() { | ||
| %c1 = arith.constant 1 : index | ||
| %c8 = arith.constant 8 : index | ||
| %cst = arith.constant 0.000000e+00 : f32 | ||
| %c0 = arith.constant 0 : index | ||
| %0 = hal.interface.binding.subspan layout(<bindings = [#hal.pipeline.binding<storage_buffer, "ReadOnly|Indirect">, #hal.pipeline.binding<storage_buffer, "ReadOnly|Indirect">, #hal.pipeline.binding<storage_buffer, Indirect>], flags = Indirect>) binding(0) alignment(64) offset(%c0) flags("ReadOnly|Indirect") : !flow.dispatch.tensor<readonly:tensor<512x512xbf16>> | ||
| %1 = hal.interface.binding.subspan layout(<bindings = [#hal.pipeline.binding<storage_buffer, "ReadOnly|Indirect">, #hal.pipeline.binding<storage_buffer, "ReadOnly|Indirect">, #hal.pipeline.binding<storage_buffer, Indirect>], flags = Indirect>) binding(1) alignment(64) offset(%c0) flags("ReadOnly|Indirect") : !flow.dispatch.tensor<readonly:tensor<512x4096xbf16>> | ||
| %2 = hal.interface.binding.subspan layout(<bindings = [#hal.pipeline.binding<storage_buffer, "ReadOnly|Indirect">, #hal.pipeline.binding<storage_buffer, "ReadOnly|Indirect">, #hal.pipeline.binding<storage_buffer, Indirect>], flags = Indirect>) binding(2) alignment(64) offset(%c0) flags(Indirect) : !flow.dispatch.tensor<writeonly:tensor<512x4096xf32>> | ||
| %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [512, 512], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<512x512xbf16>> -> tensor<512x512xbf16> | ||
| %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [512, 4096], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<512x4096xbf16>> -> tensor<512x4096xbf16> | ||
| %5 = tensor.empty() : tensor<512x4096xf32> | ||
| %6 = scf.forall (%arg0, %arg1) = (0, 0) to (512, 4096) step (128, 128) shared_outs(%arg2 = %5) -> (tensor<512x4096xf32>) { | ||
| %extracted_slice = tensor.extract_slice %3[%arg0, 0] [128, 512] [1, 1] : tensor<512x512xbf16> to tensor<128x512xbf16> | ||
| %extracted_slice_0 = tensor.extract_slice %4[0, %arg1] [512, 128] [1, 1] : tensor<512x4096xbf16> to tensor<512x128xbf16> | ||
| %extracted_slice_1 = tensor.extract_slice %arg2[%arg0, %arg1] [128, 128] [1, 1] : tensor<512x4096xf32> to tensor<128x128xf32> | ||
| %7 = tensor.empty() : tensor<4x8x32x64xbf16> | ||
| %pack = tensor.pack %extracted_slice inner_dims_pos = [0, 1] inner_tiles = [32, 64] into %7 : tensor<128x512xbf16> -> tensor<4x8x32x64xbf16> | ||
| %8 = tensor.empty() : tensor<4x8x64x32xbf16> | ||
| %pack_2 = tensor.pack %extracted_slice_0 outer_dims_perm = [1, 0] inner_dims_pos = [0, 1] inner_tiles = [64, 32] into %8 : tensor<512x128xbf16> -> tensor<4x8x64x32xbf16> | ||
| %alloc = memref.alloc() : memref<4x4x32x32xf32, 1 : i32> | ||
| %9 = bufferization.to_tensor %alloc restrict writable : memref<4x4x32x32xf32, 1 : i32> to tensor<4x4x32x32xf32> | ||
| %10 = tensor.empty() : tensor<4x4x8x8x4x4xf32> | ||
| %11 = scf.forall (%arg3, %arg4) = (0, 0) to (4, 4) step (2, 2) shared_outs(%arg5 = %10) -> (tensor<4x4x8x8x4x4xf32>) { | ||
| %extracted_slice_4 = tensor.extract_slice %pack[%arg3, 0, 0, 0] [2, 8, 32, 64] [1, 1, 1, 1] : tensor<4x8x32x64xbf16> to tensor<2x8x32x64xbf16> | ||
| %12 = tensor.empty() : tensor<2x8x8x8x4x8xbf16> | ||
| %extracted_slice_5 = tensor.extract_slice %pack_2[%arg4, 0, 0, 0] [2, 8, 64, 32] [1, 1, 1, 1] : tensor<4x8x64x32xbf16> to tensor<2x8x64x32xbf16> | ||
| %13 = tensor.empty() : tensor<2x8x8x8x8x4xbf16> | ||
| %extracted_slice_6 = tensor.extract_slice %arg5[%arg3, %arg4, 0, 0, 0, 0] [2, 2, 8, 8, 4, 4] [1, 1, 1, 1, 1, 1] : tensor<4x4x8x8x4x4xf32> to tensor<2x2x8x8x4x4xf32> | ||
| %14 = linalg.fill ins(%cst : f32) outs(%extracted_slice_6 : tensor<2x2x8x8x4x4xf32>) -> tensor<2x2x8x8x4x4xf32> | ||
| %15 = scf.for %arg6 = %c0 to %c8 step %c1 iter_args(%arg7 = %14) -> (tensor<2x2x8x8x4x4xf32>) { | ||
| %extracted_slice_7 = tensor.extract_slice %extracted_slice_4[0, %arg6, 0, 0] [2, 1, 32, 64] [1, 1, 1, 1] : tensor<2x8x32x64xbf16> to tensor<2x1x32x64xbf16> | ||
| %extracted_slice_8 = tensor.extract_slice %12[0, %arg6, 0, 0, 0, 0] [2, 1, 8, 8, 4, 8] [1, 1, 1, 1, 1, 1] : tensor<2x8x8x8x4x8xbf16> to tensor<2x1x8x8x4x8xbf16> | ||
| %pack_9 = tensor.pack %extracted_slice_7 outer_dims_perm = [0, 1, 3, 2] inner_dims_pos = [2, 3] inner_tiles = [4, 8] into %extracted_slice_8 : tensor<2x1x32x64xbf16> -> tensor<2x1x8x8x4x8xbf16> | ||
| %extracted_slice_10 = tensor.extract_slice %extracted_slice_5[0, %arg6, 0, 0] [2, 1, 64, 32] [1, 1, 1, 1] : tensor<2x8x64x32xbf16> to tensor<2x1x64x32xbf16> | ||
| %extracted_slice_11 = tensor.extract_slice %13[0, %arg6, 0, 0, 0, 0] [2, 1, 8, 8, 8, 4] [1, 1, 1, 1, 1, 1] : tensor<2x8x8x8x8x4xbf16> to tensor<2x1x8x8x8x4xbf16> | ||
| %pack_12 = tensor.pack %extracted_slice_10 outer_dims_perm = [0, 1, 3, 2] inner_dims_pos = [2, 3] inner_tiles = [8, 4] into %extracted_slice_11 : tensor<2x1x64x32xbf16> -> tensor<2x1x8x8x8x4xbf16> | ||
| %16 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3, d4, d5, d6, d7, d8) -> (d0, d2, d5, d3, d6, d8)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7, d8) -> (d1, d2, d4, d5, d8, d7)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7, d8) -> (d0, d1, d4, d3, d6, d7)>], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel", "reduction", "parallel", "parallel", "reduction"]} ins(%pack_9, %pack_12 : tensor<2x1x8x8x4x8xbf16>, tensor<2x1x8x8x8x4xbf16>) outs(%arg7 : tensor<2x2x8x8x4x4xf32>) attrs = {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[128, 128], [2, 2, 0], [0, 0, 1], [1, 1, 0, 0, 0, 0]]>, packing_config = #amdaie.packing_config<packing_config = [{packedSizes = [32, 32, 64], transposePackIndices = [1], unpackEmpty = [false], innerPerm = [[1, 0]], outerPerm = [[1, 0]]}, {packedSizes = [0, 0, 0, 4, 4, 8], transposePackIndices = [0, 1, 2], unpackEmpty = [false, false, true], innerPerm = [[0, 1], [1, 0], [0, 1]], outerPerm = [[0, 1, 3, 2], [0, 1, 3, 2], [0, 1, 3, 2]]}]>} { | ||
| ^bb0(%in: bf16, %in_13: bf16, %out: f32): | ||
| %17 = arith.extf %in : bf16 to f32 | ||
| %18 = arith.extf %in_13 : bf16 to f32 | ||
| %19 = arith.mulf %17, %18 : f32 | ||
| %20 = arith.addf %out, %19 : f32 | ||
| linalg.yield %20 : f32 | ||
| } -> tensor<2x2x8x8x4x4xf32> | ||
| scf.yield %16 : tensor<2x2x8x8x4x4xf32> | ||
| } | ||
| scf.forall.in_parallel { | ||
| tensor.parallel_insert_slice %15 into %arg5[%arg3, %arg4, 0, 0, 0, 0] [2, 2, 8, 8, 4, 4] [1, 1, 1, 1, 1, 1] : tensor<2x2x8x8x4x4xf32> into tensor<4x4x8x8x4x4xf32> | ||
| } | ||
| } {mapping = [#gpu.block<y>, #gpu.block<x>]} | ||
| %unpack = tensor.unpack %11 outer_dims_perm = [0, 1, 3, 2] inner_dims_pos = [2, 3] inner_tiles = [4, 4] into %9 : tensor<4x4x8x8x4x4xf32> -> tensor<4x4x32x32xf32> | ||
| %unpack_3 = tensor.unpack %unpack inner_dims_pos = [0, 1] inner_tiles = [32, 32] into %extracted_slice_1 : tensor<4x4x32x32xf32> -> tensor<128x128xf32> | ||
| memref.dealloc %alloc : memref<4x4x32x32xf32, 1 : i32> | ||
| scf.forall.in_parallel { | ||
| tensor.parallel_insert_slice %unpack_3 into %arg2[%arg0, %arg1] [128, 128] [1, 1] : tensor<128x128xf32> into tensor<512x4096xf32> | ||
| } | ||
| } {mapping = [#gpu.block<y>, #gpu.block<x>]} | ||
| flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [512, 4096], strides = [1, 1] : tensor<512x4096xf32> -> !flow.dispatch.tensor<writeonly:tensor<512x4096xf32>> | ||
| return | ||
| } | ||
|
|
||
| // ----- | ||
|
|
||
| #map = affine_map<(d3, d4, d5, d6, d7, d8) -> (d3, d5, d6, d8)> | ||
| #map1 = affine_map<(d3, d4, d5, d6, d7, d8) -> (d5, d4, d7, d8)> | ||
| #map2 = affine_map<(d3, d4, d5, d6, d7, d8) -> (d3, d4, d6, d7)> | ||
| func.func @pack_error(%arg0 : tensor<1024x2048xi32>, %arg1 : tensor<2048x512xi32>) -> tensor<1024x512xi32> | ||
| { | ||
| %c0_i32 = arith.constant 0 : i32 | ||
| %c0 = arith.constant 0 : index | ||
| %5 = tensor.empty() : tensor<1024x512xi32> | ||
| %6 = tensor.empty() : tensor<16x32x64x64xi32> | ||
| %pack = tensor.pack %arg0 inner_dims_pos = [0, 1] inner_tiles = [64, 64] into %6 : tensor<1024x2048xi32> -> tensor<16x32x64x64xi32> | ||
| %7 = tensor.empty() : tensor<32x8x64x64xi32> | ||
| %pack_0 = tensor.pack %arg1 outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [64, 64] into %7 : tensor<2048x512xi32> -> tensor<32x8x64x64xi32> | ||
| %8 = tensor.empty() : tensor<16x8x64x64xi32> | ||
| %9 = linalg.fill ins(%c0_i32 : i32) outs(%8 : tensor<16x8x64x64xi32>) -> tensor<16x8x64x64xi32> | ||
| // expected-error @+2 {{could not fetch operands to bufferize}} | ||
| // expected-error @+1 {{'linalg.generic' op operand #0 only has pack ops to depth 1, but request is for a depth 2 pack op}} | ||
| %10 = linalg.generic {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction", "parallel", "parallel", "reduction"]} ins(%pack, %pack_0 : tensor<16x32x64x64xi32>, tensor<32x8x64x64xi32>) outs(%9 : tensor<16x8x64x64xi32>) { | ||
| ^bb0(%in: i32, %in_4: i32, %out: i32): | ||
| %14 = arith.muli %in, %in_4 : i32 | ||
| %15 = arith.addi %out, %14 : i32 | ||
| linalg.yield %15 : i32 | ||
| } -> tensor<16x8x64x64xi32> | ||
| %unpack = tensor.unpack %10 inner_dims_pos = [0, 1] inner_tiles = [64, 64] into %5 : tensor<16x8x64x64xi32> -> tensor<1024x512xi32> | ||
| return %unpack : tensor<1024x512xi32> | ||
| } |