[OneDNN Graph Dialect] Use Broadcast Trait and organize data types #81
Conversation
| auto ret = | ||
| inferBroadcastShape<ValueShapeRange>(operands, outShape, getShapeIdx); | ||
| llvm::SmallVector<int64_t> input1, input2; | ||
| getShapeIdx(operands, 0).getDims(input1); |
There was a problem hiding this comment.
Can we use something like getShape to get the ArrayRef of shape for performance?
There was a problem hiding this comment.
Also, maybe dyn_cast<ShapedType>() should be better, in case we need to support other shaped types in the future.
| // final shape | ||
| auto retShape = ShapedTypeComponents(outShape, lhsShape.getElementType()); | ||
| inferredReturnShapes.push_back(retShape); | ||
| // check for bias broadcasting | ||
| if (adaptor.getBias()) { | ||
| ShapeAdaptor biasShape(adaptor.getBias().getType()); | ||
| ShapeAdaptor matShape(retShape); | ||
| llvm::SmallVector<int64_t> matmulShape; | ||
| matShape.getDims(matmulShape); |
There was a problem hiding this comment.
Can we use something like getShape to get the ArrayRef of shape for performance?
ZhennanQin
changed the title
LongshengDu
changed the title
| // Floating-point types. | ||
| //===----------------------------------------------------------------------===// | ||
| def OneDNNGraph_Float : AnyTypeOf<[F32, | ||
| F16, | ||
| BF16]>; | ||
|
|
||
| //===----------------------------------------------------------------------===// | ||
| // Integer types. | ||
| //===----------------------------------------------------------------------===// | ||
|
|
||
| def OneDNNGraph_Int : AnyTypeOf<[SI<8>, | ||
| UI<8>]>; |
There was a problem hiding this comment.
What is the rationale for having separate types? Also, why do we only have fp16 and int8?
There was a problem hiding this comment.
Easier to construct Quantize/Dequantize ops later. https://oneapi-src.github.io/oneDNN/dev_guide_op_quantize.html#supported-data-types
There was a problem hiding this comment.
For ops like matmul, only f32/bf16/f16 are supported: https://oneapi-src.github.io/oneDNN/dev_guide_op_matmul.html#supported-data-types
|
This will still need to change when fp8eX types are added. |
Some small changes：