transformations: (onnx) fix lower onnx.Relu lowering #2435
Conversation
kayode-gif
changed the title
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #2435 +/- ##
=======================================
Coverage 89.85% 89.85%
=======================================
Files 351 351
Lines 43757 43794 +37
Branches 6523 6530 +7
=======================================
+ Hits 39317 39351 +34
- Misses 3483 3485 +2
- Partials 957 958 +1 ☔ View full report in Codecov by Sentry. |
| // CHECK-NEXT: %2 = arith.maximumf %0, %res_relu_1 : f64 | ||
| // CHECK-NEXT: linalg.yield %2 : f64 | ||
| // CHECK-NEXT: } -> tensor<3x4xf64> | ||
| %t2 = "test.op"() : () -> (tensor<3x4xf32>) |
There was a problem hiding this comment.
Surely we should support all float types? Can you get the arg type from the parameter type in the rewrite pattern?
There was a problem hiding this comment.
I would say let's duplicate this test, to make sure that we handle both f32 and f64?
|
@kayode-gif, when you have a moment, could you please merge this? If you feel like it, minimising the diff would be nice, but not necessary |
should be all good now |
| @@ -31,10 +43,10 @@ | |||
| %res_gemm= "onnx.Gemm"(%t5, %t6, %t7) {onnx_node_name = "/Gemm", "alpha" = 1.000000e+00 : f32, "beta" = 1.000000e+00 : f32, "transA" = 0 : si64, "transB" = 1 : si64}: (tensor<1x320xf32>, tensor<50x320xf32>, tensor<50xf32>) -> tensor<1x50xf32> | |||
|
|
|||
| // CHECK-NEXT: %t5, %t6, %t7 = "test.op"() : () -> (tensor<1x320xf32>, tensor<50x320xf32>, tensor<50xf32>) | |||
| // CHECK-NEXT: %3 = tensor.empty() : tensor<320x50xf32> | |||
| // CHECK-NEXT: %4 = linalg.transpose ins(%t6:tensor<50x320xf32>) outs(%3:tensor<320x50xf32>) permutation = [1, 0] | |||
| // CHECK-NEXT: %6 = tensor.empty() : tensor<320x50xf32> | |||
There was a problem hiding this comment.
In order to avoid these spurious changes, I'd recommend using regex patterns:
| // CHECK-NEXT: %6 = tensor.empty() : tensor<320x50xf32> | |
| // CHECK-NEXT: %{{.*}} = tensor.empty() : tensor<320x50xf32> |
You can also pip install filecheckize and use it with --mlir-anonymize to convert MLIR IR to filechek tests, something like this:
xdsl-opt file.mlir -p my-awesome-pass | filecheckize --mlir-anonymize
There was a problem hiding this comment.
oops i merged it, ill keep these in mind for future, thank you
This PR fixes the issue where the affine map dimensions do not match the rank of the input operand
Resolves #2432