comaniac commented on a change in pull request #8234:
URL: https://github.com/apache/tvm/pull/8234#discussion_r660255574
##########
File path: python/tvm/topi/nn/dense.py
##########
@@ -51,37 +65,120 @@ def dense(data, weight, bias=None, out_dtype=None,
auto_scheduler_rewritten_layo
assert len(bias.shape) == 1
if out_dtype is None:
out_dtype = data.dtype
- batch, in_dim = data.shape
+ if data_transposed:
+ in_dim, batch = data.shape
+ else:
+ batch, in_dim = data.shape
if auto_scheduler_rewritten_layout:
# Infer shape for the rewritten layout
out_dim, red_dim = auto_scheduler.get_shape_from_rewritten_layout(
- auto_scheduler_rewritten_layout, ["j", "k"]
+ auto_scheduler_rewritten_layout, ["j", "k"] if weight_transposed
else ["k", "j"]
)
auto_scheduler.remove_index_check(weight)
- else:
+ elif weight_transposed:
out_dim, red_dim = weight.shape
+ else:
+ red_dim, out_dim = weight.shape
assert in_dim == red_dim
k = te.reduce_axis((0, in_dim), name="k")
- matmul = te.compute(
+ if data_transposed:
+ if weight_transposed:
+ compute_lambda = lambda i, j: te.sum(
+ data[k, i].astype(out_dtype) * weight[j, k].astype(out_dtype),
axis=k
+ )
+ compute_name = "T_matmul_TT"
+ else:
+ compute_lambda = lambda i, j: te.sum(
+ data[k, i].astype(out_dtype) * weight[k, j].astype(out_dtype),
axis=k
+ )
+ compute_name = "T_matmul_TN"
+ compute_tag = "matmul"
+ else:
+ if weight_transposed:
+ compute_lambda = lambda i, j: te.sum(
+ data[i, k].astype(out_dtype) * weight[j, k].astype(out_dtype),
axis=k
+ )
+ compute_name = "T_dense"
Review comment:
I personally vote for B.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
