[Dialect] [OneDNNGraph] Add onednn_graph ops for llama2 mlp #92
Conversation
LongshengDu
force-pushed
the
longsheng/add_onednn_ops
branch
from
21992a7 to
0781b4f
Compare
| [SameOperandsAndResultElementType, InferTensorTypeAdaptor]> { | ||
| let summary = [{ | ||
| MatMul operation computes the product of two tensors with optional bias addition. | ||
| }]; | ||
| let description = [{ | ||
| `https://oneapi-src.github.io/oneDNN/dev_guide_op_matmul.html` | ||
| }]; | ||
|
|
||
| let arguments = (ins OneDNNGraph_FloatTensor:$input_a, | ||
| OneDNNGraph_FloatTensor:$input_b, | ||
| Optional<OneDNNGraph_LogicalTensor>:$bias, |
There was a problem hiding this comment.
The optional bias here shall also be OneDNNGraph_FloatTensor
|
can you add some new ops in the test of |
|
https://github.com/intel/graph-compiler/blob/main/lib/gc/Dialect/OneDNNGraph/OneDNNGraphOps.cpp#L48 |
OP Lowering PR is WIP |
|
|
||
| // CHECK-LABEL: @relu | ||
| func.func @relu(%arg0: tensor<128x512xbf16>) -> tensor<128x512xbf16> { | ||
| // CHECK: onednn_graph.relu |
There was a problem hiding this comment.
Do we need these at all? What do they test?
There was a problem hiding this comment.
Test for basic format, attrs, auto verify and canonicalize, etc. Similar to llvm-project/mlir/test/Dialect/Linalg/named-ops.mlir
Noted, added a TODO. Will fix this after issues resolved in lowering pass |
Add llama2 mlp ops: sigmoid, type_cast, pow, mul, sub, div, reduce_sum, reduce_mean