llvm
/
torch-mlir
mirror of https://github.com/llvm/torch-mlir
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
torch-mlir/test/Conversion/TorchToLinalg/flatten.mlir

85 lines
5.8 KiB
MLIR
Raw Normal View History

Move TorchConversion dialect and TorchTo* into torch-mlir
2021-09-24 11:59:12 +08:00
// RUN: torch-mlir-opt <%s -convert-torch-to-linalg -split-input-file -verify-diagnostics | FileCheck %s
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// -----
// CHECK-LABEL: func @torch.aten.flatten.using_ints$basic(
// CHECK-SAME: %[[TENSOR:.*]]: !torch.vtensor<[3,3,2,2,3,3,5],f32>) -> !torch.vtensor<[3,3,?,3,5],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: %[[BUILTIN_TENSOR:.*]] = torch_c.to_builtin_tensor %[[TENSOR]] : !torch.vtensor<[3,3,2,2,3,3,5],f32> -> tensor<3x3x2x2x3x3x5xf32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: %[[COLLAPSED:.*]] = linalg.tensor_collapse_shape %[[BUILTIN_TENSOR]] {{\[\[}}0], [1], [2, 3, 4], [5], [6]] : tensor<3x3x2x2x3x3x5xf32> into tensor<3x3x12x3x5xf32>
// CHECK: %[[DYNAMIC:.*]] = tensor.cast %[[COLLAPSED]] : tensor<3x3x12x3x5xf32> to tensor<3x3x?x3x5xf32>
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: %[[RESULT:.*]] = torch_c.from_builtin_tensor %[[DYNAMIC]] : tensor<3x3x?x3x5xf32> -> !torch.vtensor<[3,3,?,3,5],f32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: return %[[RESULT]] : !torch.vtensor<[3,3,?,3,5],f32>
func @torch.aten.flatten.using_ints$basic(%arg0: !torch.vtensor<[3,3,2,2,3,3,5],f32>) -> !torch.vtensor<[3,3,?,3,5],f32> {
%int2 = torch.constant.int 2
%int4 = torch.constant.int 4
%0 = torch.aten.flatten.using_ints %arg0, %int2, %int4 : !torch.vtensor<[3,3,2,2,3,3,5],f32>, !torch.int, !torch.int -> !torch.vtensor<[3,3,?,3,5],f32>
return %0 : !torch.vtensor<[3,3,?,3,5],f32>
}
// -----
// CHECK-LABEL: func @torch.aten.flatten.using_ints$basic_negative(
// CHECK-SAME: %[[TENSOR:.*]]: !torch.vtensor<[3,3,2,2,3,3,5],f32>) -> !torch.vtensor<[3,3,?,3,5],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: %[[BUILTIN_TENSOR:.*]] = torch_c.to_builtin_tensor %[[TENSOR]] : !torch.vtensor<[3,3,2,2,3,3,5],f32> -> tensor<3x3x2x2x3x3x5xf32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: %[[COLLAPSED:.*]] = linalg.tensor_collapse_shape %[[BUILTIN_TENSOR]] {{\[\[}}0], [1], [2, 3, 4], [5], [6]] : tensor<3x3x2x2x3x3x5xf32> into tensor<3x3x12x3x5xf32>
// CHECK: %[[DYNAMIC:.*]] = tensor.cast %[[COLLAPSED]] : tensor<3x3x12x3x5xf32> to tensor<3x3x?x3x5xf32>
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: %[[RESULT:.*]] = torch_c.from_builtin_tensor %[[DYNAMIC]] : tensor<3x3x?x3x5xf32> -> !torch.vtensor<[3,3,?,3,5],f32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: return %[[RESULT]] : !torch.vtensor<[3,3,?,3,5],f32>
func @torch.aten.flatten.using_ints$basic_negative(%arg0: !torch.vtensor<[3,3,2,2,3,3,5],f32>) -> !torch.vtensor<[3,3,?,3,5],f32> {
%int-5 = torch.constant.int -5
%int-3 = torch.constant.int -3
%0 = torch.aten.flatten.using_ints %arg0, %int-5, %int-3 : !torch.vtensor<[3,3,2,2,3,3,5],f32>, !torch.int, !torch.int -> !torch.vtensor<[3,3,?,3,5],f32>
return %0 : !torch.vtensor<[3,3,?,3,5],f32>
}
// -----
Update llvm-project to 830c0b9023cd0cf91955900e0d96283e7a8c3711 - builder.getSymbolRefAttr is gone. - OpAsmOpInterface's getAsmResultNames method needs explicit override - a bunch of churn for builtin.func needing to be made explicit (and sometimes implicit?) - operation printers no longer need to print the operation name themselves. - snuck in beneficial trivial addition to TmpDeleteDeadIREEListsPass to test a particular upstream change e2e with my local patchset.
2021-09-04 02:38:00 +08:00
// CHECK-LABEL: func @torch.aten.flatten.using_ints$flatten_front(
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK-SAME: %[[TENSOR:.*]]: !torch.vtensor<[3,3,2,2],f32>) -> !torch.vtensor<[?,?],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: %[[BUILTIN_TENSOR:.*]] = torch_c.to_builtin_tensor %[[TENSOR]] : !torch.vtensor<[3,3,2,2],f32> -> tensor<3x3x2x2xf32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: %[[COLLAPSED:.*]] = linalg.tensor_collapse_shape %[[BUILTIN_TENSOR]] {{\[\[}}0, 1, 2], [3]] : tensor<3x3x2x2xf32> into tensor<18x2xf32>
// CHECK: %[[DYNAMIC:.*]] = tensor.cast %[[COLLAPSED]] : tensor<18x2xf32> to tensor<?x?xf32>
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: %[[RESULT:.*]] = torch_c.from_builtin_tensor %[[DYNAMIC]] : tensor<?x?xf32> -> !torch.vtensor<[?,?],f32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: return %[[RESULT]] : !torch.vtensor<[?,?],f32>
func @torch.aten.flatten.using_ints$flatten_front(%arg0: !torch.vtensor<[3,3,2,2],f32>) -> !torch.vtensor<[?,?],f32> {
%int0 = torch.constant.int 0
%int2 = torch.constant.int 2
%0 = torch.aten.flatten.using_ints %arg0, %int0, %int2 : !torch.vtensor<[3,3,2,2],f32>, !torch.int, !torch.int -> !torch.vtensor<[?,?],f32>
return %0 : !torch.vtensor<[?,?],f32>
}
// -----
Update llvm-project to 830c0b9023cd0cf91955900e0d96283e7a8c3711 - builder.getSymbolRefAttr is gone. - OpAsmOpInterface's getAsmResultNames method needs explicit override - a bunch of churn for builtin.func needing to be made explicit (and sometimes implicit?) - operation printers no longer need to print the operation name themselves. - snuck in beneficial trivial addition to TmpDeleteDeadIREEListsPass to test a particular upstream change e2e with my local patchset.
2021-09-04 02:38:00 +08:00
// CHECK-LABEL: func @torch.aten.flatten.using_ints$flatten_back(
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK-SAME: %[[TENSOR:.*]]: !torch.vtensor<[3,3,2,2],f32>) -> !torch.vtensor<[?,12],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: %[[BUILTIN_TENSOR:.*]] = torch_c.to_builtin_tensor %[[TENSOR]] : !torch.vtensor<[3,3,2,2],f32> -> tensor<3x3x2x2xf32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: %[[COLLAPSED:.*]] = linalg.tensor_collapse_shape %[[BUILTIN_TENSOR]] {{\[\[}}0], [1, 2, 3]] : tensor<3x3x2x2xf32> into tensor<3x12xf32>
// CHECK: %[[DYNAMIC:.*]] = tensor.cast %[[COLLAPSED]] : tensor<3x12xf32> to tensor<?x12xf32>
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: %[[RESULT:.*]] = torch_c.from_builtin_tensor %[[DYNAMIC]] : tensor<?x12xf32> -> !torch.vtensor<[?,12],f32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: return %[[RESULT]] : !torch.vtensor<[?,12],f32>
func @torch.aten.flatten.using_ints$flatten_back(%arg0: !torch.vtensor<[3,3,2,2],f32>) -> !torch.vtensor<[?,12],f32> {
%int1 = torch.constant.int 1
%int-1 = torch.constant.int -1
%0 = torch.aten.flatten.using_ints %arg0, %int1, %int-1 : !torch.vtensor<[3,3,2,2],f32>, !torch.int, !torch.int -> !torch.vtensor<[?,12],f32>
return %0 : !torch.vtensor<[?,12],f32>
}
// -----
// CHECK-LABEL: func @torch.aten.flatten.using_ints$rank0(
// CHECK-SAME: %[[TENSOR:.*]]: !torch.vtensor<[],f32>) -> !torch.vtensor<[1],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: %[[BUILTIN_TENSOR:.*]] = torch_c.to_builtin_tensor %[[TENSOR]] : !torch.vtensor<[],f32> -> tensor<f32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: %[[COLLAPSED:.*]] = linalg.tensor_expand_shape %[[BUILTIN_TENSOR]] [] : tensor<f32> into tensor<1xf32>
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: %[[RESULT:.*]] = torch_c.from_builtin_tensor %[[COLLAPSED]] : tensor<1xf32> -> !torch.vtensor<[1],f32>
Add torch.aten.flatten.using_ints and aten.MaxPool2d linalg lowering - torch.aten.flatten.using_ints to linalg lowering - torch.aten.max_pool2d to linalg lowering - Support torch.aten.conv2d for more flexible dilation and strides values
2021-07-08 04:59:47 +08:00
// CHECK: return %[[RESULT]] : !torch.vtensor<[1],f32>
func @torch.aten.flatten.using_ints$rank0(%arg0: !torch.vtensor<[],f32>) -> !torch.vtensor<[1],f32> {
%int0 = torch.constant.int 0
%0 = torch.aten.flatten.using_ints %arg0, %int0, %int0 : !torch.vtensor<[],f32>, !torch.int, !torch.int -> !torch.vtensor<[1],f32>
return %0 : !torch.vtensor<[1],f32>
}