torch-mlir
torch-mlir copied to clipboard
llvm
can we support float16(half) datatype?
my model is float16 model which failed to transform in torch-mlir https://github.com/llvm/torch-mlir/blob/32159c4e545359d4ad9e3b492c07e07528f5b47e/python/torch_mlir/dialects/torch/importer/jit_ir/csrc/torch_to_mlir_utils.cpp#L327
case ScalarType::Int: case ScalarType::Long: case ScalarType::Float: case ScalarType::Double: case ScalarType::Bool: case ScalarType::QInt8: case ScalarType::BFloat16: default: throwUnsupportedTensorError(); }
can we support fp16?
This is probably easy given the Half is a native APFloat type in LLVM. You probably want to look at @ashay 's commit adding BF16 here: https://github.com/llvm/torch-mlir/pull/761
maybe not.
below are interface exposed from MLIR, as you can see, it doesnot support Half :( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrBoolGet( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrUInt8Get( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrInt8Get( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrUInt16Get( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrInt16Get( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrUInt32Get( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrInt32Get( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrUInt64Get( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrInt64Get( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrFloatGet( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrDoubleGet( MLIR_CAPI_EXPORTED MlirAttribute mlirDenseElementsAttrBFloat16Get(
@FullZing Yeah, we were in a similar situation for BFloat16, so I added the mlirDenseElementsAttrBFloat16Get() to the MLIR C API in https://reviews.llvm.org/D123981. Perhaps you could add a similar function upstream for Half.
+1, this should be quite easy to add by repeating https://reviews.llvm.org/D123981 + https://github.com/llvm/torch-mlir/pull/761 but for float16
@FullZing -- can you do that?
@ashay @silvasean I have added a patch: https://reviews.llvm.org/D130069; please review for me.
@fortianyou That patch looks good to me, but could you add "ashay-github" as a reviewer? I botched the signup process on Phabricator, which is why it says that the email address for the account "ashay" was not verified. Sorry about the hassle!