mirror of https://github.com/llvm/torch-mlir
496 lines
13 KiB
Python
496 lines
13 KiB
Python
# Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
# See https://llvm.org/LICENSE.txt for license information.
|
|
# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
# Also available under a BSD-style license. See LICENSE.
|
|
|
|
import torch
|
|
|
|
from torch_mlir_e2e_test.torchscript.framework import TestUtils
|
|
from torch_mlir_e2e_test.torchscript.registry import register_test_case
|
|
from torch_mlir_e2e_test.torchscript.annotations import annotate_args, export
|
|
|
|
# ==============================================================================
|
|
|
|
|
|
class MmModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1], torch.float32, True),
|
|
([-1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, lhs, rhs):
|
|
return torch.mm(lhs, rhs)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: MmModule())
|
|
def MmModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(4, 4), tu.rand(4, 4))
|
|
|
|
|
|
@register_test_case(module_factory=lambda: MmModule())
|
|
def MmModule_chained(module, tu: TestUtils):
|
|
res = module.forward(tu.rand(4, 4), tu.rand(4, 4))
|
|
module.forward(res, res)
|
|
|
|
# ==============================================================================
|
|
|
|
|
|
class BmmModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, -1], torch.float32, True),
|
|
([-1, -1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, lhs, rhs):
|
|
return torch.bmm(lhs, rhs)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: BmmModule())
|
|
def BmmModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(3, 4, 5), tu.rand(3, 5, 4))
|
|
|
|
|
|
# ==============================================================================
|
|
|
|
|
|
# A subgraph with multiple mm ops.
|
|
class MmDagModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([4, 4], torch.float32, True),
|
|
([4, 4], torch.float32, True),
|
|
])
|
|
def forward(self, lhs, rhs):
|
|
return torch.mm(lhs, torch.mm(lhs, rhs))
|
|
|
|
|
|
@register_test_case(module_factory=lambda: MmDagModule())
|
|
def MmDagModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(4, 4), tu.rand(4, 4))
|
|
|
|
|
|
# ==============================================================================
|
|
|
|
|
|
class MmTanhModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1], torch.float32, True),
|
|
([-1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, lhs, rhs):
|
|
return torch.tanh(self.matmul(lhs, rhs))
|
|
|
|
def matmul(self, lhs, rhs):
|
|
return torch.mm(lhs, rhs)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: MmTanhModule())
|
|
def MmTanhModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(4, 2), tu.rand(2, 4))
|
|
|
|
|
|
class AdaptiveAvgPool2dModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.aap2d = torch.nn.AdaptiveAvgPool2d((1, 1))
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, -1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, x):
|
|
return self.aap2d(x)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: AdaptiveAvgPool2dModule())
|
|
def AdaptiveAvgPool2dModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(10, 3, 8, 9))
|
|
|
|
|
|
class FlattenStaticModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.flat = torch.nn.Flatten(2, 4)
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([10, 3, 8, 9, 3, 4], torch.float32, True),
|
|
])
|
|
def forward(self, x):
|
|
return self.flat(x)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: FlattenStaticModule())
|
|
def FlattenStaticModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(10, 3, 8, 9, 3, 4))
|
|
|
|
|
|
class FlattenRank0Module(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.flat = torch.nn.Flatten(-1, -1)
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([], torch.float32, True),
|
|
])
|
|
def forward(self, x):
|
|
return self.flat(x)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: FlattenRank0Module())
|
|
def FlattenRank0Module_basic(module, tu: TestUtils):
|
|
module.forward(torch.tensor(4.0))
|
|
|
|
|
|
class FlattenDynamicModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.flat = torch.nn.Flatten(2, 4)
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, -1, 9, 3, -1], torch.float32, True),
|
|
])
|
|
def forward(self, x):
|
|
return self.flat(x)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: FlattenDynamicModule())
|
|
def FlattenDynamicModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(10, 3, 8, 9, 3, 4))
|
|
|
|
|
|
class MaxPool2dModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.mp2d = torch.nn.MaxPool2d(kernel_size=[6, 8],
|
|
stride=[2, 2],
|
|
padding=[3, 4],
|
|
dilation=2)
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, -1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, x):
|
|
return self.mp2d(x)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: MaxPool2dModule())
|
|
def MaxPool2dModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(1, 1, 20, 20) - 0.5)
|
|
|
|
|
|
class TransposeIntModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([3, 4, 2], torch.float32, True),
|
|
])
|
|
def forward(self, x):
|
|
return torch.transpose(x, 0, 1)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: TransposeIntModule())
|
|
def TransposeIntModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(3, 4, 2))
|
|
|
|
class PermuteModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([3, 4, 2], torch.float32, True)
|
|
])
|
|
def forward(self, x):
|
|
return x.permute(0, 2, 1)
|
|
|
|
@register_test_case(module_factory=lambda: PermuteModule())
|
|
def PermuteModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(3, 4, 2))
|
|
|
|
class TransposeIntNegDimsModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([3, 4, 2], torch.float32, True),
|
|
])
|
|
def forward(self, x):
|
|
return torch.transpose(x, -1, -2)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: TransposeIntNegDimsModule())
|
|
def TransposeIntNegDimsModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(3, 4, 2))
|
|
|
|
|
|
class PermuteNegativeIndexModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([3, 4, 2], torch.float32, True)
|
|
])
|
|
def forward(self, x):
|
|
return x.permute(0, -1, 1)
|
|
|
|
@register_test_case(module_factory=lambda: PermuteNegativeIndexModule())
|
|
def PermuteNegativeIndexModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(3, 4, 2))
|
|
class TensorsConcatModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, -1], torch.float32, True),
|
|
([-1, -1, -1], torch.float32, True),
|
|
([-1, -1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, x, y, z):
|
|
return torch.cat([x, y, z], 1)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: TensorsConcatModule())
|
|
def TensorsConcatModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(2, 2, 4), tu.rand(2, 1, 4), tu.rand(2, 3, 4))
|
|
|
|
|
|
class GatherModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, -1], torch.float32, True),
|
|
([-1, -1, -1], torch.int64, True),
|
|
])
|
|
def forward(self, tensor, indices):
|
|
return torch.gather(tensor, 2, indices)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: GatherModule())
|
|
def GatherModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(2, 3, 4), torch.tensor([[[1, 2, 3], [1, 2, 3]]]))
|
|
|
|
class AddSizeIntModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, tensor):
|
|
# This is a workaround for not supporting scalar arguments.
|
|
# TODO: pass in dim as an argument to the forward method when scalar
|
|
# arguments are supported.
|
|
return tensor.add(tensor, alpha=tensor.size(1))
|
|
|
|
|
|
@register_test_case(module_factory=lambda: AddSizeIntModule())
|
|
def AddSizeIntModule_basic(module, tu: TestUtils):
|
|
module.forward(torch.randn(3, 3))
|
|
|
|
|
|
class AddSizeIntNegDimModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, tensor):
|
|
# This is a workaround for not supporting scalar arguments.
|
|
# TODO: pass in dim as an argument to the forward method when scalar
|
|
# arguments are supported.
|
|
return tensor.add(tensor, alpha=tensor.size(-2))
|
|
|
|
|
|
@register_test_case(module_factory=lambda: AddSizeIntNegDimModule())
|
|
def AddSizeIntNegDimModule_basic(module, tu: TestUtils):
|
|
module.forward(torch.randn(3, 3))
|
|
|
|
class EmbeddingModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
torch.manual_seed(0)
|
|
self.embed = torch.nn.Embedding(num_embeddings=100,
|
|
embedding_dim=50,
|
|
padding_idx=4)
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1], torch.int64, True),
|
|
])
|
|
def forward(self, indices):
|
|
return self.embed.forward(indices)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: EmbeddingModule())
|
|
def EmbeddingModule_basic(module, tu: TestUtils):
|
|
module.forward(torch.randint(100, (3, 3)))
|
|
|
|
|
|
class SoftmaxIntModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
torch.manual_seed(0)
|
|
self.softmax = torch.nn.Softmax(2)
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, tensor):
|
|
return self.softmax.forward(tensor)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: SoftmaxIntModule())
|
|
def SoftmaxIntModule_basic(module, tu: TestUtils):
|
|
module.forward(torch.randn(3, 2, 4))
|
|
|
|
|
|
class SoftmaxIntNegDimModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
torch.manual_seed(0)
|
|
self.softmax = torch.nn.Softmax(-2)
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, tensor):
|
|
return self.softmax.forward(tensor)
|
|
|
|
|
|
@register_test_case(module_factory=lambda: SoftmaxIntNegDimModule())
|
|
def SoftmaxIntNegDimModule_basic(module, tu: TestUtils):
|
|
module.forward(torch.randn(3, 2, 4))
|
|
|
|
|
|
class SoftmaxIntArgTypeF64Module(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
torch.manual_seed(0)
|
|
self.softmax = torch.nn.Softmax(2)
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, -1], torch.float64, True),
|
|
])
|
|
def forward(self, tensor):
|
|
return self.softmax.forward(tensor)
|
|
|
|
@register_test_case(module_factory=lambda: SoftmaxIntArgTypeF64Module())
|
|
def SoftmaxIntArgTypeF64Module_basic(module, tu: TestUtils):
|
|
module.forward(torch.randn(3, 2, 4).double())
|
|
|
|
class BroadcastToModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1, 1], torch.float32, True),
|
|
])
|
|
def forward(self, x):
|
|
return torch.broadcast_to(x, [1, -1, -1, 4])
|
|
|
|
|
|
@register_test_case(module_factory=lambda: BroadcastToModule())
|
|
def BroadcastToModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(3, 1, 1))
|
|
|
|
|
|
class OnesModuleInt(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
])
|
|
def forward(self):
|
|
return torch.ones(3, 4, dtype=torch.int64)
|
|
|
|
@register_test_case(module_factory=lambda: OnesModuleInt())
|
|
def OnesModuleInt_basic(module, tu: TestUtils):
|
|
module.forward()
|
|
|
|
class OnesModuleFloat(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
])
|
|
def forward(self):
|
|
return torch.ones(3, 4, dtype=torch.float32)
|
|
|
|
@register_test_case(module_factory=lambda: OnesModuleFloat())
|
|
def OnesModuleFloat_basic(module, tu: TestUtils):
|
|
module.forward()
|
|
|
|
class ContiguousModule(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@export
|
|
@annotate_args([
|
|
None,
|
|
([-1, -1], torch.float32, True),
|
|
])
|
|
def forward(self, x):
|
|
return x.contiguous()
|
|
|
|
|
|
@register_test_case(module_factory=lambda: ContiguousModule())
|
|
def ContiguousModule_basic(module, tu: TestUtils):
|
|
module.forward(tu.rand(3, 1))
|