61 lines · python
1# RUN: %PYTHON %s | FileCheck %s2 3from mlir.ir import *4from mlir.dialects import builtin5from mlir.dialects import func6from mlir.dialects import linalg7 8from mlir.dialects.linalg.opdsl.lang import *9 10T1 = TV.T111T2 = TV.T212 13 14@linalg_structured_op15def conv_poly(16 I=TensorDef(T1, S.N, S.IH, S.IW, S.C),17 K=TensorDef(T2, S.KH, S.KW, S.C),18 O=TensorDef(U, S.N, S.OH, S.OW, S.C, output=True),19 strides=IndexAttrDef(S.SH, S.SW, default=[1, 1]),20 dilations=IndexAttrDef(S.DH, S.DW, default=[1, 2]),21):22 domain(D.n, D.oh, D.ow, D.kh, D.kw, D.c)23 O[D.n, D.oh, D.ow, D.c] += TypeFn.cast_signed(24 U, I[D.n, D.oh * S.SH + D.kh * S.DH, D.ow * S.SW + D.kw * S.DW, D.c]25 ) * TypeFn.cast_signed(U, K[D.kh, D.kw, D.c])26 27 28with Context() as ctx, Location.unknown():29 module = Module.create()30 f32 = F32Type.get()31 i32 = IntegerType.get_signless(32)32 with InsertionPoint(module.body):33 34 # Convolution indexing maps.35 # CHECK: #[[$CONV_MAP_I:.+]] = affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1 * 2 + d3, d2 * 4 + d4 * 2, d5)>36 # CHECK: #[[$CONV_MAP_K:.+]] = affine_map<(d0, d1, d2, d3, d4, d5) -> (d3, d4, d5)>37 # CHECK: #[[$CONV_MAP_O:.+]] = affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2, d5)>38 39 # CHECK-LABEL: @test_f32i32_conv40 # CHECK: linalg.generic41 # CHECK-SAME: indexing_maps = [#[[$CONV_MAP_I]], #[[$CONV_MAP_K]], #[[$CONV_MAP_O]]]42 # CHECK-SAME: iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction", "parallel"]43 # CHECK: ^{{.*}}(%[[IN:.+]]: f32, %[[FILTER:.+]]: f32, %[[OUT:.+]]: i32)44 # CHECK-NEXT: %[[IN_CAST:.+]] = arith.fptosi %[[IN:.+]] : f32 to i3245 # CHECK-NEXT: %[[FILTER_CAST:.+]] = arith.fptosi %[[FILTER:.+]] : f32 to i3246 # CHECK-NEXT: %[[PROD:.+]] = arith.muli %[[IN_CAST]], %[[FILTER_CAST]] : i3247 # CHECK-NEXT: %[[SUM:.+]] = arith.addi %[[OUT]], %[[PROD]] : i3248 # CHECK-NEXT: linalg.yield %[[SUM]] : i3249 # CHECK-NEXT: -> tensor<1x2x4x1xi32>50 @func.FuncOp.from_py_func(51 RankedTensorType.get((1, 4, 16, 1), f32),52 RankedTensorType.get((2, 2, 1), f32),53 RankedTensorType.get((1, 2, 4, 1), i32),54 )55 def test_f32i32_conv(input, filter, init_result):56 # Use default dilations and set non-default strides.57 return conv_poly(input, filter, outs=[init_result], strides=[2, 4])58 59 60print(module)61