2021-09-24 11:59:12 +08:00
|
|
|
// RUN: torch-mlir-opt <%s -convert-torch-to-std | FileCheck %s
|
2021-04-28 06:15:50 +08:00
|
|
|
|
2021-06-17 06:53:15 +08:00
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.dim(
|
2021-06-17 06:53:15 +08:00
|
|
|
// CHECK-SAME: %[[ARG:.*]]: !torch.vtensor<*,f32>) -> !torch.int {
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[BUILTIN_TENSOR:.*]] = torch_c.to_builtin_tensor %[[ARG]] : !torch.vtensor<*,f32> -> tensor<*xf32>
|
2022-01-05 00:00:48 +08:00
|
|
|
// CHECK: %[[RANK:.*]] = tensor.rank %[[BUILTIN_TENSOR]] : tensor<*xf32>
|
2021-10-16 02:34:29 +08:00
|
|
|
// CHECK: %[[RANK_I64:.*]] = arith.index_cast %[[RANK]] : index to i64
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[RANK_TORCH_INT:.*]] = torch_c.from_i64 %[[RANK_I64]]
|
2021-06-17 06:53:15 +08:00
|
|
|
// CHECK: return %[[RANK_TORCH_INT]] : !torch.int
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.dim(%arg0: !torch.vtensor<*,f32>) -> !torch.int {
|
2021-06-17 06:53:15 +08:00
|
|
|
%0 = torch.aten.dim %arg0 : !torch.vtensor<*,f32> -> !torch.int
|
|
|
|
|
return %0 : !torch.int
|
2021-04-28 06:15:50 +08:00
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.runtime.assert(
|
2022-02-08 22:18:08 +08:00
|
|
|
// CHECK-SAME: %[[X:.*]]: !torch.int,
|
|
|
|
|
// CHECK-SAME: %[[Y:.*]]: !torch.int) {
|
|
|
|
|
// CHECK: %[[X_I64:.*]] = torch_c.to_i64 %[[X]]
|
|
|
|
|
// CHECK: %[[Y_I64:.*]] = torch_c.to_i64 %[[Y]]
|
|
|
|
|
// CHECK: %[[CMP:.*]] = arith.cmpi ne, %[[X_I64]], %[[Y_I64]] : i64
|
|
|
|
|
// CHECK: assert %[[CMP]], "x must not be equal to y"
|
|
|
|
|
// CHECK: return
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.runtime.assert(%arg0: !torch.int, %arg1: !torch.int) {
|
2022-02-08 22:18:08 +08:00
|
|
|
%0 = torch.aten.ne.int %arg0, %arg1 : !torch.int, !torch.int -> !torch.bool
|
|
|
|
|
torch.runtime.assert %0, "x must not be equal to y"
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.ne.int(
|
2021-06-17 06:53:15 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.int,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.int) -> !torch.bool {
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[LHS_I64:.*]] = torch_c.to_i64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_I64:.*]] = torch_c.to_i64 %[[RHS]]
|
2021-10-16 02:34:29 +08:00
|
|
|
// CHECK: %[[CMP:.*]] = arith.cmpi ne, %[[LHS_I64]], %[[RHS_I64]] : i64
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[CMP_TORCH_BOOL:.*]] = torch_c.from_i1 %[[CMP]]
|
2021-06-17 06:53:15 +08:00
|
|
|
// CHECK: return %[[CMP_TORCH_BOOL]] : !torch.bool
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.ne.int(%arg0: !torch.int, %arg1: !torch.int) -> !torch.bool {
|
2021-06-17 06:53:15 +08:00
|
|
|
%0 = torch.aten.ne.int %arg0, %arg1 : !torch.int, !torch.int -> !torch.bool
|
2021-06-16 07:47:53 +08:00
|
|
|
return %0 : !torch.bool
|
2022-02-12 05:39:30 +08:00
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.eq.int(
|
2022-02-12 05:39:30 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.int,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.int) -> !torch.bool {
|
|
|
|
|
// CHECK: %[[LHS_I64:.*]] = torch_c.to_i64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_I64:.*]] = torch_c.to_i64 %[[RHS]]
|
|
|
|
|
// CHECK: %[[CMP:.*]] = arith.cmpi eq, %[[LHS_I64]], %[[RHS_I64]] : i64
|
|
|
|
|
// CHECK: %[[CMP_TORCH_BOOL:.*]] = torch_c.from_i1 %[[CMP]]
|
|
|
|
|
// CHECK: return %[[CMP_TORCH_BOOL]] : !torch.bool
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.eq.int(%arg0: !torch.int, %arg1: !torch.int) -> !torch.bool {
|
2022-02-12 05:39:30 +08:00
|
|
|
%0 = torch.aten.eq.int %arg0, %arg1 : !torch.int, !torch.int -> !torch.bool
|
|
|
|
|
return %0 : !torch.bool
|
2021-04-28 06:15:50 +08:00
|
|
|
}
|
2021-04-30 06:38:11 +08:00
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.gt.int(
|
2021-06-17 06:53:15 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.int,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.int) -> !torch.bool {
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[LHS_I64:.*]] = torch_c.to_i64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_I64:.*]] = torch_c.to_i64 %[[RHS]]
|
2021-10-16 02:34:29 +08:00
|
|
|
// CHECK: %[[CMP:.*]] = arith.cmpi sgt, %[[LHS_I64]], %[[RHS_I64]] : i64
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[CMP_TORCH_BOOL:.*]] = torch_c.from_i1 %[[CMP]]
|
2021-06-17 06:53:15 +08:00
|
|
|
// CHECK: return %[[CMP_TORCH_BOOL]] : !torch.bool
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.gt.int(%arg0: !torch.int, %arg1: !torch.int) -> !torch.bool {
|
2021-06-17 06:53:15 +08:00
|
|
|
%0 = torch.aten.gt.int %arg0, %arg1 : !torch.int, !torch.int -> !torch.bool
|
2021-06-16 07:47:53 +08:00
|
|
|
return %0 : !torch.bool
|
2021-04-30 06:38:11 +08:00
|
|
|
}
|
Introduce `!torch.tensor` / `!torch.vtensor` types.
This removes our reliance on the numpy dialect and avoids our off-label
use of the builtin tnesor type for modeling unknown dtypes. The
`!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor.
The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic
tensor. The new types look as follows syntactically:
```
// Least-static-information, non-value-semantic tensor.
!torch.tensor
// Explicit form of least-static-information variant.
!torch.tensor<*,unk>
// Least-static-information, value-semantic tensor.
!torch.vtensor
// Explicit form of least-static-information variant.
!torch.vtensor<*,unk>
// Fixed-set of allowable element types, with first-class support for
// Torch's frontend signedness semantics.
!torch.tensor<*,si32>
// First-class support for unknown dtypes.
!torch.tensor<[?,?,?],unk>
// Standard MLIR representation of `?` for unknown dimensions.
!torch.tensor<[?,2,?,4],unk>
// Statically shaped / dtyped example.
!torch.vtensor<[1,2,3,4],f32>
```
This required fairly significant changes throughout the compiler, but
overall it is a big cleanup. We now have a much clearer layering of "the
Torch frontend lowering" vs "lowering to std + linalg + etc.".
At the C++ level, there is `ValueTensorType`, `NonValueTensorType`.
We also have a helper `BaseTensorType` (kind of like ShapedType) which
interoperates with those two.
Included changes:
- New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for
creating torch tensor literals in the frontend.
- Consistently use signedness for the types (except i1 which I didn't
touch -- we need to sort out the situation with !basicpy.BoolType
there anyway so will be attending to that soon)
- Frontend can annotate whether an argument to the function has value
semantics. We currently require this, as our backend contract does not
currently allow us to even model the non-value-semantic case. Before,
the value-semantic assumption was randomly injected in the middle of
the pass pipeline.
- Move ArrayToTensor (now called MaximizeValueSemantics) and
RefinePublicReturn passes to torch dialect.
- The TorchToStd and TorchToLinalg passes are now type conversions from
`!torch.vtensor` to `tensor` and use the dialect conversion infra.
The overall conversion pipeline is set up following the best practices
of the "Type Conversions the Not-So-Hard Way" talk. This required
introducing `torch-func-builtin-tensorize` and
`torch-finalizing-builtin-tensorize` passes analogous to the upstream
bufferization passes with the corresponding names (mostly just
copypasta from there).
- Misc Torch-level canonicalizations -- we now cleanly layer the
lowering to std later in the pipeline, so we are gradually lessening
our reliance on random std constant folding before we get to that
point.
Recommended review order:
- New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp
- New ops in TorchOps.td / TorchOps.cpp
- Less important / more mechanical stuff
- Frontend changes.
- Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.vtensor.literal() -> !torch.vtensor<[],f32> {
|
2021-10-16 02:34:29 +08:00
|
|
|
// CHECK: %[[CST:.*]] = arith.constant dense<0.000000e+00> : tensor<f32>
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[VTENSOR:.*]] = torch_c.from_builtin_tensor %[[CST]] : tensor<f32> -> !torch.vtensor<[],f32>
|
Introduce `!torch.tensor` / `!torch.vtensor` types.
This removes our reliance on the numpy dialect and avoids our off-label
use of the builtin tnesor type for modeling unknown dtypes. The
`!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor.
The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic
tensor. The new types look as follows syntactically:
```
// Least-static-information, non-value-semantic tensor.
!torch.tensor
// Explicit form of least-static-information variant.
!torch.tensor<*,unk>
// Least-static-information, value-semantic tensor.
!torch.vtensor
// Explicit form of least-static-information variant.
!torch.vtensor<*,unk>
// Fixed-set of allowable element types, with first-class support for
// Torch's frontend signedness semantics.
!torch.tensor<*,si32>
// First-class support for unknown dtypes.
!torch.tensor<[?,?,?],unk>
// Standard MLIR representation of `?` for unknown dimensions.
!torch.tensor<[?,2,?,4],unk>
// Statically shaped / dtyped example.
!torch.vtensor<[1,2,3,4],f32>
```
This required fairly significant changes throughout the compiler, but
overall it is a big cleanup. We now have a much clearer layering of "the
Torch frontend lowering" vs "lowering to std + linalg + etc.".
At the C++ level, there is `ValueTensorType`, `NonValueTensorType`.
We also have a helper `BaseTensorType` (kind of like ShapedType) which
interoperates with those two.
Included changes:
- New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for
creating torch tensor literals in the frontend.
- Consistently use signedness for the types (except i1 which I didn't
touch -- we need to sort out the situation with !basicpy.BoolType
there anyway so will be attending to that soon)
- Frontend can annotate whether an argument to the function has value
semantics. We currently require this, as our backend contract does not
currently allow us to even model the non-value-semantic case. Before,
the value-semantic assumption was randomly injected in the middle of
the pass pipeline.
- Move ArrayToTensor (now called MaximizeValueSemantics) and
RefinePublicReturn passes to torch dialect.
- The TorchToStd and TorchToLinalg passes are now type conversions from
`!torch.vtensor` to `tensor` and use the dialect conversion infra.
The overall conversion pipeline is set up following the best practices
of the "Type Conversions the Not-So-Hard Way" talk. This required
introducing `torch-func-builtin-tensorize` and
`torch-finalizing-builtin-tensorize` passes analogous to the upstream
bufferization passes with the corresponding names (mostly just
copypasta from there).
- Misc Torch-level canonicalizations -- we now cleanly layer the
lowering to std later in the pipeline, so we are gradually lessening
our reliance on random std constant folding before we get to that
point.
Recommended review order:
- New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp
- New ops in TorchOps.td / TorchOps.cpp
- Less important / more mechanical stuff
- Frontend changes.
- Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
|
|
|
// CHECK: return %[[VTENSOR]] : !torch.vtensor<[],f32>
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.vtensor.literal() -> !torch.vtensor<[],f32> {
|
2021-06-17 23:52:13 +08:00
|
|
|
%0 = torch.vtensor.literal(dense<0.0> : tensor<f32>) : !torch.vtensor<[],f32>
|
Introduce `!torch.tensor` / `!torch.vtensor` types.
This removes our reliance on the numpy dialect and avoids our off-label
use of the builtin tnesor type for modeling unknown dtypes. The
`!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor.
The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic
tensor. The new types look as follows syntactically:
```
// Least-static-information, non-value-semantic tensor.
!torch.tensor
// Explicit form of least-static-information variant.
!torch.tensor<*,unk>
// Least-static-information, value-semantic tensor.
!torch.vtensor
// Explicit form of least-static-information variant.
!torch.vtensor<*,unk>
// Fixed-set of allowable element types, with first-class support for
// Torch's frontend signedness semantics.
!torch.tensor<*,si32>
// First-class support for unknown dtypes.
!torch.tensor<[?,?,?],unk>
// Standard MLIR representation of `?` for unknown dimensions.
!torch.tensor<[?,2,?,4],unk>
// Statically shaped / dtyped example.
!torch.vtensor<[1,2,3,4],f32>
```
This required fairly significant changes throughout the compiler, but
overall it is a big cleanup. We now have a much clearer layering of "the
Torch frontend lowering" vs "lowering to std + linalg + etc.".
At the C++ level, there is `ValueTensorType`, `NonValueTensorType`.
We also have a helper `BaseTensorType` (kind of like ShapedType) which
interoperates with those two.
Included changes:
- New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for
creating torch tensor literals in the frontend.
- Consistently use signedness for the types (except i1 which I didn't
touch -- we need to sort out the situation with !basicpy.BoolType
there anyway so will be attending to that soon)
- Frontend can annotate whether an argument to the function has value
semantics. We currently require this, as our backend contract does not
currently allow us to even model the non-value-semantic case. Before,
the value-semantic assumption was randomly injected in the middle of
the pass pipeline.
- Move ArrayToTensor (now called MaximizeValueSemantics) and
RefinePublicReturn passes to torch dialect.
- The TorchToStd and TorchToLinalg passes are now type conversions from
`!torch.vtensor` to `tensor` and use the dialect conversion infra.
The overall conversion pipeline is set up following the best practices
of the "Type Conversions the Not-So-Hard Way" talk. This required
introducing `torch-func-builtin-tensorize` and
`torch-finalizing-builtin-tensorize` passes analogous to the upstream
bufferization passes with the corresponding names (mostly just
copypasta from there).
- Misc Torch-level canonicalizations -- we now cleanly layer the
lowering to std later in the pipeline, so we are gradually lessening
our reliance on random std constant folding before we get to that
point.
Recommended review order:
- New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp
- New ops in TorchOps.td / TorchOps.cpp
- Less important / more mechanical stuff
- Frontend changes.
- Pass changes/additions in `Torch/Transforms` and `Conversion/`
2021-05-21 08:07:18 +08:00
|
|
|
return %0 : !torch.vtensor<[],f32>
|
|
|
|
|
}
|
2021-06-19 01:44:43 +08:00
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.constant.bool() -> !torch.bool {
|
2021-10-16 02:34:29 +08:00
|
|
|
// CHECK: %[[CST:.*]] = arith.constant true
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[BOOL:.*]] = torch_c.from_i1 %[[CST]]
|
2021-06-19 01:44:43 +08:00
|
|
|
// CHECK: return %[[BOOL]] : !torch.bool
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.constant.bool() -> !torch.bool {
|
2021-06-19 01:44:43 +08:00
|
|
|
%true = torch.constant.bool true
|
|
|
|
|
return %true : !torch.bool
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.constant.float() -> !torch.float {
|
2021-10-16 02:34:29 +08:00
|
|
|
// CHECK: %[[CST:.*]] = arith.constant 1.000000e+00 : f64
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[FLOAT:.*]] = torch_c.from_f64 %[[CST]]
|
2021-06-19 01:44:43 +08:00
|
|
|
// CHECK: return %[[FLOAT]] : !torch.float
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.constant.float() -> !torch.float {
|
2021-06-19 01:44:43 +08:00
|
|
|
%float = torch.constant.float 1.000000e+00
|
|
|
|
|
return %float : !torch.float
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.constant.int() -> !torch.int {
|
2021-10-16 02:34:29 +08:00
|
|
|
// CHECK: %[[CST:.*]] = arith.constant 1 : i64
|
Add TorchToIREE and factor out TorchConversion dialect.
This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
2021-08-12 05:40:08 +08:00
|
|
|
// CHECK: %[[INT:.*]] = torch_c.from_i64 %[[CST]]
|
2021-06-19 01:44:43 +08:00
|
|
|
// CHECK: return %[[INT]] : !torch.int
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.constant.int() -> !torch.int {
|
2021-06-19 01:44:43 +08:00
|
|
|
%int1 = torch.constant.int 1
|
|
|
|
|
return %int1 : !torch.int
|
|
|
|
|
}
|
2021-11-24 08:16:32 +08:00
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.add.int(
|
2022-02-17 00:34:03 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.int,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.int) -> !torch.int {
|
2021-11-24 08:16:32 +08:00
|
|
|
// CHECK: %[[LHS_I64:.*]] = torch_c.to_i64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_I64:.*]] = torch_c.to_i64 %[[RHS]]
|
2022-02-17 00:34:03 +08:00
|
|
|
// CHECK: %[[ADD:.*]] = arith.addi %[[LHS_I64:.*]], [[RHS_I64:.*]] : i64
|
|
|
|
|
// CHECK: %[[OUT:.*]] = torch_c.from_i64 %[[INT:.*]]
|
|
|
|
|
// CHECK: return %[[OUT:.*]] : !torch.int
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.add.int(%arg0: !torch.int, %arg1: !torch.int) -> !torch.int {
|
2021-12-17 23:24:56 +08:00
|
|
|
%0 = torch.aten.add.int %arg0, %arg1 : !torch.int, !torch.int -> !torch.int
|
|
|
|
|
return %0 : !torch.int
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.sub.int(
|
2022-02-17 00:34:03 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.int,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.int) -> !torch.int {
|
2021-12-17 23:24:56 +08:00
|
|
|
// CHECK: %[[LHS_I64:.*]] = torch_c.to_i64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_I64:.*]] = torch_c.to_i64 %[[RHS]]
|
2022-02-17 00:34:03 +08:00
|
|
|
// CHECK: %[[SUB:.*]] = arith.subi %[[LHS_I64:.*]], [[RHS_I64:.*]] : i64
|
|
|
|
|
// CHECK: %[[OUT:.*]] = torch_c.from_i64 %[[INT:.*]]
|
|
|
|
|
// CHECK: return %[[OUT:.*]] : !torch.int
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.sub.int(%arg0: !torch.int, %arg1: !torch.int) -> !torch.int {
|
2021-12-17 23:24:56 +08:00
|
|
|
%0 = torch.aten.sub.int %arg0, %arg1 : !torch.int, !torch.int -> !torch.int
|
|
|
|
|
return %0 : !torch.int
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.sub.float(
|
2022-02-17 00:34:03 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.float,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.float) -> !torch.float {
|
|
|
|
|
// CHECK: %[[LHS_F64:.*]] = torch_c.to_f64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_F64:.*]] = torch_c.to_f64 %[[RHS]]
|
|
|
|
|
// CHECK: %[[SUB:.*]] = arith.subf %[[LHS_F64:.*]], [[RHS_F64:.*]] : f64
|
|
|
|
|
// CHECK: %[[OUT:.*]] = torch_c.from_f64 %[[SUB:.*]]
|
|
|
|
|
// CHECK: return %[[OUT:.*]] : !torch.float
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.sub.float(%arg0: !torch.float, %arg1: !torch.float) -> !torch.float {
|
2022-02-17 00:34:03 +08:00
|
|
|
%0 = torch.aten.sub.float %arg0, %arg1 : !torch.float, !torch.float -> !torch.float
|
|
|
|
|
return %0 : !torch.float
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.mul.int(
|
2022-02-17 00:34:03 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.int,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.int) -> !torch.int {
|
2021-12-17 23:24:56 +08:00
|
|
|
// CHECK: %[[LHS_I64:.*]] = torch_c.to_i64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_I64:.*]] = torch_c.to_i64 %[[RHS]]
|
2022-02-17 00:34:03 +08:00
|
|
|
// CHECK: %[[MUL:.*]] = arith.muli %[[LHS_I64:.*]], [[RHS_I64:.*]] : i64
|
|
|
|
|
// CHECK: %[[OUT:.*]] = torch_c.from_i64 %[[MUL:.*]]
|
|
|
|
|
// CHECK: return %[[OUT:.*]] : !torch.int
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.mul.int(%arg0: !torch.int, %arg1: !torch.int) -> !torch.int {
|
2021-12-17 23:24:56 +08:00
|
|
|
%0 = torch.aten.mul.int %arg0, %arg1 : !torch.int, !torch.int -> !torch.int
|
2021-11-24 08:16:32 +08:00
|
|
|
return %0 : !torch.int
|
|
|
|
|
}
|
2022-04-25 20:06:41 +08:00
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.div.float(
|
2022-04-25 20:06:41 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.float,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.float) -> !torch.float {
|
|
|
|
|
// CHECK: %[[LHS_F64:.*]] = torch_c.to_f64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_F64:.*]] = torch_c.to_f64 %[[RHS]]
|
|
|
|
|
// CHECK: %[[SUB:.*]] = arith.divf %[[LHS_F64:.*]], [[RHS_F64:.*]] : f64
|
|
|
|
|
// CHECK: %[[OUT:.*]] = torch_c.from_f64 %[[SUB:.*]]
|
|
|
|
|
// CHECK: return %[[OUT:.*]] : !torch.float
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.div.float(%arg0: !torch.float, %arg1: !torch.float) -> !torch.float {
|
2022-04-25 20:06:41 +08:00
|
|
|
%0 = torch.aten.div.float %arg0, %arg1 : !torch.float, !torch.float -> !torch.float
|
|
|
|
|
return %0 : !torch.float
|
|
|
|
|
}
|
2022-04-25 21:12:45 +08:00
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.ge.float(
|
2022-04-25 21:12:45 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.float,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.float) -> !torch.bool {
|
|
|
|
|
// CHECK: %[[LHS_F64:.*]] = torch_c.to_f64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_F64:.*]] = torch_c.to_f64 %[[RHS]]
|
|
|
|
|
// CHECK: %[[CMP:.*]] = arith.cmpf uge, %[[LHS_F64]], %[[RHS_F64]] : f64
|
|
|
|
|
// CHECK: %[[CMP_TORCH_BOOL:.*]] = torch_c.from_i1 %[[CMP]]
|
|
|
|
|
// CHECK: return %[[CMP_TORCH_BOOL]] : !torch.bool
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.ge.float(%arg0: !torch.float, %arg1: !torch.float) -> !torch.bool {
|
2022-04-25 21:12:45 +08:00
|
|
|
%0 = torch.aten.ge.float %arg0, %arg1 : !torch.float, !torch.float -> !torch.bool
|
|
|
|
|
return %0 : !torch.bool
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.ge.float_int(
|
2022-04-25 21:12:45 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.float,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.int) -> !torch.bool {
|
|
|
|
|
// CHECK: %[[LHS_F64:.*]] = torch_c.to_f64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_I64:.*]] = torch_c.to_i64 %[[RHS]]
|
|
|
|
|
// CHECK: %[[RHS_F64:.*]] = arith.sitofp %[[RHS_I64]] : i64 to f64
|
|
|
|
|
// CHECK: %[[CMP:.*]] = arith.cmpf uge, %[[LHS_F64]], %[[RHS_F64]] : f64
|
|
|
|
|
// CHECK: %[[CMP_TORCH_BOOL:.*]] = torch_c.from_i1 %[[CMP]]
|
|
|
|
|
// CHECK: return %[[CMP_TORCH_BOOL]] : !torch.bool
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.ge.float_int(%arg0: !torch.float, %arg1: !torch.int) -> !torch.bool {
|
2022-04-25 21:12:45 +08:00
|
|
|
%0 = torch.aten.ge.float_int %arg0, %arg1 : !torch.float, !torch.int -> !torch.bool
|
|
|
|
|
return %0 : !torch.bool
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.ne.float_int(
|
2022-04-25 21:12:45 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.float,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.int) -> !torch.bool {
|
|
|
|
|
// CHECK: %[[LHS_F64:.*]] = torch_c.to_f64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_I64:.*]] = torch_c.to_i64 %[[RHS]]
|
|
|
|
|
// CHECK: %[[RHS_F64:.*]] = arith.sitofp %[[RHS_I64]] : i64 to f64
|
|
|
|
|
// CHECK: %[[CMP:.*]] = arith.cmpf une, %[[LHS_F64]], %[[RHS_F64]] : f64
|
|
|
|
|
// CHECK: %[[CMP_TORCH_BOOL:.*]] = torch_c.from_i1 %[[CMP]]
|
|
|
|
|
// CHECK: return %[[CMP_TORCH_BOOL]] : !torch.bool
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.ne.float_int(%arg0: !torch.float, %arg1: !torch.int) -> !torch.bool {
|
2022-04-25 21:12:45 +08:00
|
|
|
%0 = torch.aten.ne.float_int %arg0, %arg1 : !torch.float, !torch.int -> !torch.bool
|
|
|
|
|
return %0 : !torch.bool
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.ceil.float(
|
2022-04-25 21:12:45 +08:00
|
|
|
// CHECK-SAME: %[[ARG:.*]]: !torch.float) -> !torch.int {
|
|
|
|
|
// CHECK: %[[ARG_F64:.*]] = torch_c.to_f64 %[[ARG]]
|
|
|
|
|
// CHECK: %[[CEIL:.*]] = math.ceil %[[ARG_F64]] : f64
|
|
|
|
|
// CHECK: %[[CEIL_I64:.*]] = arith.fptosi %[[CEIL]] : f64 to i64
|
|
|
|
|
// CHECK: %[[OUT:.*]] = torch_c.from_i64 %[[CEIL_I64]]
|
|
|
|
|
// CHECK: return %[[OUT]] : !torch.int
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.ceil.float(%arg0: !torch.float) -> !torch.int {
|
2022-04-25 21:12:45 +08:00
|
|
|
%0 = torch.aten.ceil.float %arg0 : !torch.float -> !torch.int
|
|
|
|
|
return %0 : !torch.int
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-17 03:54:35 +08:00
|
|
|
// CHECK-LABEL: func.func @torch.aten.gt.float_int(
|
2022-04-25 21:12:45 +08:00
|
|
|
// CHECK-SAME: %[[LHS:.*]]: !torch.float,
|
|
|
|
|
// CHECK-SAME: %[[RHS:.*]]: !torch.int) -> !torch.bool {
|
|
|
|
|
// CHECK: %[[LHS_F64:.*]] = torch_c.to_f64 %[[LHS]]
|
|
|
|
|
// CHECK: %[[RHS_I64:.*]] = torch_c.to_i64 %[[RHS]]
|
|
|
|
|
// CHECK: %[[RHS_F64:.*]] = arith.sitofp %[[RHS_I64]] : i64 to f64
|
|
|
|
|
// CHECK: %[[CMP:.*]] = arith.cmpf ugt, %[[LHS_F64]], %[[RHS_F64]] : f64
|
|
|
|
|
// CHECK: %[[CMP_TORCH_BOOL:.*]] = torch_c.from_i1 %[[CMP]]
|
|
|
|
|
// CHECK: return %[[CMP_TORCH_BOOL]] : !torch.bool
|
2022-05-17 03:54:35 +08:00
|
|
|
func.func @torch.aten.gt.float_int(%arg0: !torch.float, %arg1: !torch.int) -> !torch.bool {
|
2022-04-25 21:12:45 +08:00
|
|
|
%0 = torch.aten.gt.float_int %arg0, %arg1 : !torch.float, !torch.int -> !torch.bool
|
|
|
|
|
return %0 : !torch.bool
|
|
|
|
|
}
|
2022-05-19 22:54:16 +08:00
|
|
|
|
|
|
|
|
// CHECK-LABEL: func.func @torch.aten.sqrt.int(
|
|
|
|
|
// CHECK-SAME: %[[ARG:.*]]: !torch.int) -> !torch.float {
|
|
|
|
|
// CHECK: %[[ARG_I64:.*]] = torch_c.to_i64 %[[ARG]]
|
|
|
|
|
// CHECK: %[[ARG_F64:.*]] = arith.sitofp %[[ARG_I64]] : i64 to f64
|
|
|
|
|
// CHECK: %[[SQRT:.*]] = math.sqrt %[[ARG_F64]] : f64
|
|
|
|
|
// CHECK: %[[SQRT_TORCH_FLOAT:.*]] = torch_c.from_f64 %[[SQRT]]
|
|
|
|
|
// CHECK: return %[[SQRT_TORCH_FLOAT]] : !torch.float
|
|
|
|
|
func.func @torch.aten.sqrt.int(%arg0: !torch.int) -> !torch.float {
|
|
|
|
|
%0 = torch.aten.sqrt.int %arg0 : !torch.int -> !torch.float
|
|
|
|
|
return %0 : !torch.float
|
|
|
|
|
}
|
2022-05-23 23:26:02 +08:00
|
|
|
|
|
|
|
|
// CHECK-LABEL: func.func @torch.aten.any.bool() -> !torch.bool {
|
|
|
|
|
// CHECK: %[[CST_FALSE:.*]] = arith.constant false
|
|
|
|
|
// CHECK: %[[FALSE:.*]] = torch_c.from_i1 %[[CST_FALSE]]
|
|
|
|
|
// CHECK: %[[CST_TRUE:.*]] = arith.constant true
|
|
|
|
|
// CHECK: %[[TRUE:.*]] = torch_c.from_i1 %[[CST_TRUE]]
|
|
|
|
|
// CHECK: %[[INPUT:.*]] = torch.prim.ListConstruct %[[FALSE]], %[[TRUE]], %[[FALSE]] : (!torch.bool, !torch.bool, !torch.bool) -> !torch.list<bool>
|
|
|
|
|
// CHECK: %[[CST_RESULT:.*]] = arith.constant true
|
|
|
|
|
// CHECK: %[[RESULT:.*]] = torch_c.from_i1 %[[CST_RESULT]]
|
|
|
|
|
// CHECK: return %[[RESULT]] : !torch.bool
|
|
|
|
|
func.func @torch.aten.any.bool() -> !torch.bool {
|
|
|
|
|
%false = torch.constant.bool false
|
|
|
|
|
%true = torch.constant.bool true
|
|
|
|
|
%input = torch.prim.ListConstruct %false, %true, %false : (!torch.bool, !torch.bool, !torch.bool) -> !torch.list<bool>
|
|
|
|
|
%0 = torch.aten.any.bool %input : !torch.list<bool> -> !torch.bool
|
|
|
|
|
return %0 : !torch.bool
|
|
|
|
|
}
|
