[AutoBump] Merge with fixes of aca33f17 (Oct 22, needs LLVM Oct 19) (92) #476
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…llvm#3807) I've upstreamed the necessary quantized linalg Op with the "channel-first" ordering used by torch (llvm/llvm-project#107740) for 2d convolution. This patch changes the lowering for the quantized 2d case of `aten.convolution` accordingly, which saves three transpositions per convolution (input, weights, result) and therefore removes the requirement to try to optimize these away in downstream passes.
…lculate` region (llvm#3812) Reports a match failure for the pattern `FullyUnrollPrimLoop` when the loop op is not in a region defined by a `torch.shape.calculate` op. This is needed to avoid unrolling prim loops generated by ONNX IR, since we are applying shape refinement in the `torch-onnx-to-torch-backend-pipeline` introduced in fa4794d . See also the discussion in <iree-org/iree#18867 (comment)>
As discussed in llvm#3652, we should replace maxpool3dwithindices with maxpool3d if indices have no user.
Added autopad. and passed 3 tests test_maxpool_2d_precomputed_same_upper test_maxpool_2d_same_lower' test_maxpool_2d_same_upper Address : nod-ai/SHARK-ModelDev#843 2 attributes yet to complete : storage_order, indices output
### Changes
1. Folders for view-like ops: `aten.view`, `aten.flatten.using_ints`,
and `aten.unflatten.int`
2. Folder for transpose
3. Extended support for the `aten.slice.Tensor` op folder to include
negative strides.
### Motivation
The biggest motivation for this patch is to fold the extremely
convoluted ir that gets generated when exporting a pytorch model with an
`aten.pad` op to ONNX, then re-importing and lowering back to torch. For
example, the verbose output of the e2e test `PadModule_basic` with `-c
onnx`:
```mlir
module {
func.func @main_graph(%arg0: !torch.vtensor<[?,?,?,?],f32>) -> !torch.vtensor<[?,?,?,?],f32> attributes {torch.onnx_meta.ir_version = 9 : si64, torch.onnx_meta.opset_version = 20 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.5.0"} {
%none = torch.constant.none
%0 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<_> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%1 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__1> : tensor<4xsi64>} : () -> !torch.vtensor<[4],si64>
%2 = torch.operator "onnx.ConstantOfShape"(%0) {torch.onnx.value = dense_resource<__2> : tensor<1xsi64>} : (!torch.vtensor<[1],si64>) -> !torch.vtensor<[4],si64>
%3 = torch.operator "onnx.Concat"(%1, %2) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[4],si64>, !torch.vtensor<[4],si64>) -> !torch.vtensor<[8],si64>
%4 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__3> : tensor<2xsi64>} : () -> !torch.vtensor<[2],si64>
%5 = torch.operator "onnx.Reshape"(%3, %4) {torch.onnx.allowzero = 0 : si64} : (!torch.vtensor<[8],si64>, !torch.vtensor<[2],si64>) -> !torch.vtensor<[4,2],si64>
%6 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__4> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%7 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__5> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%8 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__6> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%9 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__7> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%10 = torch.operator "onnx.Slice"(%5, %7, %8, %6, %9) : (!torch.vtensor<[4,2],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[4,2],si64>
%11 = torch.operator "onnx.Transpose"(%10) {torch.onnx.perm = [1 : si64, 0 : si64]} : (!torch.vtensor<[4,2],si64>) -> !torch.vtensor<[2,4],si64>
%12 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__8> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
%13 = torch.operator "onnx.Reshape"(%11, %12) {torch.onnx.allowzero = 0 : si64} : (!torch.vtensor<[2,4],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[8],si64>
%14 = torch.operator "onnx.Cast"(%13) {torch.onnx.to = 7 : si64} : (!torch.vtensor<[8],si64>) -> !torch.vtensor<[8],si64>
%15 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__9> : tensor<f32>} : () -> !torch.vtensor<[],f32>
%16 = torch.operator "onnx.Pad"(%arg0, %14, %15) {torch.onnx.mode = "constant"} : (!torch.vtensor<[?,?,?,?],f32>, !torch.vtensor<[8],si64>, !torch.vtensor<[],f32>) -> !torch.vtensor<[?,?,?,?],f32>
return %16 : !torch.vtensor<[?,?,?,?],f32>
}
}
{-#
dialect_resources: {
builtin: {
_: "0x080000000400000000000000",
__1: "0x080000000000000000000000010000000000000002000000000000000300000000000000",
__2: "0x080000000000000000000000",
__3: "0x08000000FFFFFFFFFFFFFFFF0200000000000000",
__4: "0x080000000000000000000000",
__5: "0x08000000FFFFFFFFFFFFFFFF",
__6: "0x080000000100000000000080",
__7: "0x08000000FFFFFFFFFFFFFFFF",
__8: "0x08000000FFFFFFFFFFFFFFFF",
__9: "0x080000000000C03F"
}
}
#-}
```
Get's converted to the torch IR:
```mlir
module {
func.func @main_graph(%arg0: !torch.vtensor<[?,?,?,?],f32>) -> !torch.vtensor<[?,?,?,?],f32> attributes {torch.onnx_meta.ir_version = 9 : si64, torch.onnx_meta.opset_version = 20 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.5.0"} {
%float1.500000e00 = torch.constant.float 1.500000e+00
%int-9223372036854775807 = torch.constant.int -9223372036854775807
%int-1 = torch.constant.int -1
%int7 = torch.constant.int 7
%int6 = torch.constant.int 6
%int5 = torch.constant.int 5
%int3 = torch.constant.int 3
%int8 = torch.constant.int 8
%int1 = torch.constant.int 1
%int2 = torch.constant.int 2
%int4 = torch.constant.int 4
%int0 = torch.constant.int 0
%0 = torch.vtensor.literal(dense<[0, 1, 2, 3, 0, 0, 0, 0]> : tensor<8xsi64>) : !torch.vtensor<[8],si64>
%1 = torch.prim.ListConstruct %int4, %int2 : (!torch.int, !torch.int) -> !torch.list<int>
%2 = torch.aten.view %0, %1 : !torch.vtensor<[8],si64>, !torch.list<int> -> !torch.vtensor<[4,2],si64>
%3 = torch.aten.slice.Tensor %2, %int0, %int-1, %int-9223372036854775807, %int-1 : !torch.vtensor<[4,2],si64>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[4,2],si64>
%4 = torch.aten.transpose.int %3, %int0, %int1 : !torch.vtensor<[4,2],si64>, !torch.int, !torch.int -> !torch.vtensor<[2,4],si64>
%5 = torch.prim.ListConstruct %int-1 : (!torch.int) -> !torch.list<int>
%6 = torch.aten.view %4, %5 : !torch.vtensor<[2,4],si64>, !torch.list<int> -> !torch.vtensor<[8],si64>
%7 = torch.aten.slice.Tensor %6, %int0, %int0, %int1, %int1 : !torch.vtensor<[8],si64>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%8 = torch.aten.item %7 : !torch.vtensor<[1],si64> -> !torch.int
%9 = torch.aten.slice.Tensor %6, %int0, %int1, %int2, %int1 : !torch.vtensor<[8],si64>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%10 = torch.aten.item %9 : !torch.vtensor<[1],si64> -> !torch.int
%11 = torch.aten.slice.Tensor %6, %int0, %int2, %int3, %int1 : !torch.vtensor<[8],si64>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%12 = torch.aten.item %11 : !torch.vtensor<[1],si64> -> !torch.int
%13 = torch.aten.slice.Tensor %6, %int0, %int3, %int4, %int1 : !torch.vtensor<[8],si64>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%14 = torch.aten.item %13 : !torch.vtensor<[1],si64> -> !torch.int
%15 = torch.aten.slice.Tensor %6, %int0, %int4, %int5, %int1 : !torch.vtensor<[8],si64>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%16 = torch.aten.item %15 : !torch.vtensor<[1],si64> -> !torch.int
%17 = torch.aten.slice.Tensor %6, %int0, %int5, %int6, %int1 : !torch.vtensor<[8],si64>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%18 = torch.aten.item %17 : !torch.vtensor<[1],si64> -> !torch.int
%19 = torch.aten.slice.Tensor %6, %int0, %int6, %int7, %int1 : !torch.vtensor<[8],si64>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%20 = torch.aten.item %19 : !torch.vtensor<[1],si64> -> !torch.int
%21 = torch.aten.slice.Tensor %6, %int0, %int7, %int8, %int1 : !torch.vtensor<[8],si64>, !torch.int, !torch.int, !torch.int, !torch.int -> !torch.vtensor<[1],si64>
%22 = torch.aten.item %21 : !torch.vtensor<[1],si64> -> !torch.int
%23 = torch.prim.ListConstruct %14, %22, %12, %20, %10, %18, %8, %16 : (!torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
%24 = torch.aten.constant_pad_nd %arg0, %23, %float1.500000e00 : !torch.vtensor<[?,?,?,?],f32>, !torch.list<int>, !torch.float -> !torch.vtensor<[?,?,?,?],f32>
return %24 : !torch.vtensor<[?,?,?,?],f32>
}
}
```
***All of these operations are useless***. It is literally the result of
needing to reverse (and change the lexicographic order hierarchy of)
padding ints provided via torch vs. ONNX pad ops, which is then
subsequently UNDONE by our ONNX->Torch lowering (represented in the
ordering of the generated list construct).
With the added folders in this patch, the torch IR becomes:
```
module {
func.func @main_graph(%arg0: !torch.vtensor<[?,?,?,?],f32>) -> !torch.vtensor<[?,?,?,?],f32> attributes {torch.onnx_meta.ir_version = 9 : si64, torch.onnx_meta.opset_version = 20 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.5.0"} {
%float1.500000e00 = torch.constant.float 1.500000e+00
%int0 = torch.constant.int 0
%int2 = torch.constant.int 2
%int3 = torch.constant.int 3
%int1 = torch.constant.int 1
%0 = torch.prim.ListConstruct %int0, %int1, %int2, %int3, %int0, %int0, %int0, %int0 : (!torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
%1 = torch.aten.constant_pad_nd %arg0, %0, %float1.500000e00 : !torch.vtensor<[?,?,?,?],f32>, !torch.list<int>, !torch.float -> !torch.vtensor<[?,?,?,?],f32>
return %1 : !torch.vtensor<[?,?,?,?],f32>
}
}
```
…llvm#3815) Continuation of llvm#3809 for the matmul tests.
Compiling with clang 16.0 on macOS I have warnings about incorrect
printf format (see below).
Values to be printed are `int64_t`, but they are printed with `%zu` and
`%ld`, which are not portable way to print this type.
```
<...>/torch-mlir/test/CAPI/torch.c:52:3: warning: format specifies type 'size_t' (aka 'unsigned long') but the argument has type 'int64_t' (aka 'long long') [-Wformat]
52 | DEFINE_CHECK(NonValueTensor)
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
<...>/torch-mlir/test/CAPI/torch.c:37:13: note: expanded from macro 'DEFINE_CHECK'
36 | fprintf(stderr, #TTT "Type %s rank: %zu\n", testName, \
| ~~~
37 | torchMlirTorch##TTT##TypeGetRank(TTT##Type)); \
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<scratch space>:78:1: note: expanded from here
78 | torchMlirTorchNonValueTensorTypeGetRank
| ^
<...>/torch-mlir/test/CAPI/torch.c:52:3: warning: format specifies type 'long' but the argument has type 'int64_t' (aka 'long long') [-Wformat]
52 | DEFINE_CHECK(NonValueTensor)
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
<...>/torch-mlir/test/CAPI/torch.c:42:15: note: expanded from macro 'DEFINE_CHECK'
41 | fprintf(stderr, #TTT "Type %s pos %d size: %ld\n", testName, i, \
| ~~~
42 | TTT##Sizes[i]); \
| ^~~~~~~~~~~~~
<scratch space>:85:1: note: expanded from here
85 | NonValueTensorSizes
| ^
<...>/torch-mlir/test/CAPI/torch.c:53:3: warning: format specifies type 'size_t' (aka 'unsigned long') but the argument has type 'int64_t' (aka 'long long') [-Wformat]
53 | DEFINE_CHECK(ValueTensor)
| ^~~~~~~~~~~~~~~~~~~~~~~~~
<...>/torch-mlir/test/CAPI/torch.c:37:13: note: expanded from macro 'DEFINE_CHECK'
36 | fprintf(stderr, #TTT "Type %s rank: %zu\n", testName, \
| ~~~
37 | torchMlirTorch##TTT##TypeGetRank(TTT##Type)); \
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<scratch space>:112:1: note: expanded from here
112 | torchMlirTorchValueTensorTypeGetRank
| ^
<...>/torch-mlir/test/CAPI/torch.c:53:3: warning: format specifies type 'long' but the argument has type 'int64_t' (aka 'long long') [-Wformat]
53 | DEFINE_CHECK(ValueTensor)
| ^~~~~~~~~~~~~~~~~~~~~~~~~
<...>/torch-mlir/test/CAPI/torch.c:42:15: note: expanded from macro 'DEFINE_CHECK'
41 | fprintf(stderr, #TTT "Type %s pos %d size: %ld\n", testName, i, \
| ~~~
42 | TTT##Sizes[i]); \
| ^~~~~~~~~~~~~
<scratch space>:119:1: note: expanded from here
119 | ValueTensorSizes
| ^
4 warnings generated.
```
… torch.aten.rrelu_with_noise_backward ops (fix) (llvm#3748)
…ensor_hacked_twin for TorchToTosa lowering. (llvm#3790) 1. Negative indices for tensor indexing is handled by wrapping around the index values by checking their values at run time. Without the fix, there was a runtime error. 2. Added a lit test to lock down the behavior. 3. Updated the `xfails_set` for `fx_importer_tosa` config to lockdown the behavior with e2e test as well. "THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY."
This commit sets the PyTorch and TorchVision version to nightly release 2024-10-29. This commit also fixes the CI failure after this commit llvm@54d9e24 got merged. The issue was that the CI checks in the PR were run before the previous roll pytorch update but the PR was actually merged after the roll pytorch update. Hence, the failure was not caught before merging the PR. While exporting the fx_graph through fx_importer for `rrelu` and `rrelu_with_noise` op for train mode, it decomposes the `aten.rrelu_with_noise` op based on the PyTorch decomposition which is the default behavior. However, the decomposition contains an input mutation specifically here https://github.com/pytorch/pytorch/blob/9bbe4a67ad137032add6a3b0b74bda66f5ef83d2/torch/_decomp/decompositions.py#L325, resulting in the runtime failure. This issue would probably be fixed by pytorch/pytorch#138503. Until then, the failing tests are added to the xfail set. Also, after the roll pytorch update following tests started passing for fx_importer, and fx_importer_stablehlo config. - "ElementwiseRreluTrainModule_basic" - "ElementwiseRreluTrainStaticModule_basic" - "ElementwiseRreluWithNoiseTrainModule_basic" - "ElementwiseRreluWithNoiseTrainStaticModule_basic" This commit also updates the dtype check for the `aten.linear` op since the op now expects both the input tensors to have the same dtype. Signed-Off By: Vivek Khandelwal <[email protected]>
Removes a boolean variable that is used only for an assertion, and inlines the condition into the assertion. Signed-off-by: Max Dawkins <[email protected]>
- bumps llvm-project to llvm/llvm-project@6c64c8a - bumps stablehlo to openxla/stablehlo@6e403b1 - Updates type conversion materialization functions to return Value after API change in llvm-project. --------- Signed-off-by: Max Dawkins <[email protected]>
mgehre-amd
changed the title
3 tasks
mgehre-amd
force-pushed
the
bump_to_aca33f17
branch
from
5ce612e to
0038abc
Compare
jorickert
approved these changes
Jan 29, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Requires llvm bump