tlopex commented on code in PR #18609:
URL: https://github.com/apache/tvm/pull/18609#discussion_r2647033801
##########
tests/python/relax/test_transform_legalize_ops_nn.py:
##########
@@ -2091,6 +2110,62 @@ def batch_norm(var_x: T.handle, var_gamma: T.handle,
var_beta: T.handle, var_mov
T.reads(x[v_ax0, v_ax1, v_ax2, v_ax3],
T_reshape[T.int64(0), v_ax1, T.int64(0), T.int64(0)])
T.writes(T_subtract[v_ax0, v_ax1, v_ax2,
v_ax3])
T_subtract[v_ax0, v_ax1, v_ax2, v_ax3] =
x[v_ax0, v_ax1, v_ax2, v_ax3] - T_reshape[T.int64(0), v_ax1, T.int64(0),
T.int64(0)]
+ for ax0 in range(T.int64(2)):
+ for ax1 in range(T.int64(3)):
+ for ax2 in range(T.int64(28)):
+ for ax3 in range(T.int64(28)):
+ with T.block("T_subtract_1"):
+ v_ax0 = T.axis.spatial(T.int64(2), ax0)
+ v_ax1 = T.axis.spatial(T.int64(3), ax1)
+ v_ax2 = T.axis.spatial(T.int64(28), ax2)
+ v_ax3 = T.axis.spatial(T.int64(28), ax3)
+ T.reads(x[v_ax0, v_ax1, v_ax2, v_ax3],
T_reshape[T.int64(0), v_ax1, T.int64(0), T.int64(0)])
+ T.writes(T_subtract_1[v_ax0, v_ax1, v_ax2,
v_ax3])
+ T_subtract_1[v_ax0, v_ax1, v_ax2, v_ax3] =
x[v_ax0, v_ax1, v_ax2, v_ax3] - T_reshape[T.int64(0), v_ax1, T.int64(0),
T.int64(0)]
+ for ax0 in range(T.int64(2)):
+ for ax1 in range(T.int64(3)):
+ for ax2 in range(T.int64(28)):
+ for ax3 in range(T.int64(28)):
+ with T.block("T_subtract_2"):
+ v_ax0 = T.axis.spatial(T.int64(2), ax0)
+ v_ax1 = T.axis.spatial(T.int64(3), ax1)
+ v_ax2 = T.axis.spatial(T.int64(28), ax2)
+ v_ax3 = T.axis.spatial(T.int64(28), ax3)
+ T.reads(x[v_ax0, v_ax1, v_ax2, v_ax3],
T_reshape[T.int64(0), v_ax1, T.int64(0), T.int64(0)])
+ T.writes(T_subtract_2[v_ax0, v_ax1, v_ax2,
v_ax3])
+ T_subtract_2[v_ax0, v_ax1, v_ax2, v_ax3] =
x[v_ax0, v_ax1, v_ax2, v_ax3] - T_reshape[T.int64(0), v_ax1, T.int64(0),
T.int64(0)]
+ for ax0 in range(T.int64(2)):
+ for ax1 in range(T.int64(3)):
+ for ax2 in range(T.int64(28)):
+ for ax3 in range(T.int64(28)):
+ with T.block("T_multiply"):
+ v_ax0 = T.axis.spatial(T.int64(2), ax0)
+ v_ax1 = T.axis.spatial(T.int64(3), ax1)
+ v_ax2 = T.axis.spatial(T.int64(28), ax2)
+ v_ax3 = T.axis.spatial(T.int64(28), ax3)
+ T.reads(T_subtract_1[v_ax0, v_ax1, v_ax2,
v_ax3], T_subtract_2[v_ax0, v_ax1, v_ax2, v_ax3])
+ T.writes(T_multiply[v_ax0, v_ax1, v_ax2,
v_ax3])
+ T_multiply[v_ax0, v_ax1, v_ax2, v_ax3] =
T_subtract_1[v_ax0, v_ax1, v_ax2, v_ax3] * T_subtract_2[v_ax0, v_ax1, v_ax2,
v_ax3]
+ for ax0 in range(T.int64(3)):
+ for k0 in range(T.int64(2)):
+ for k2 in range(T.int64(28)):
+ for k3 in range(T.int64(28)):
+ with T.block("T_multiply_red"):
+ v_ax0 = T.axis.spatial(T.int64(3), ax0)
+ v_k0 = T.axis.reduce(T.int64(2), k0)
+ v_k2 = T.axis.reduce(T.int64(28), k2)
+ v_k3 = T.axis.reduce(T.int64(28), k3)
+ T.reads(T_multiply[v_k0, v_ax0, v_k2,
v_k3])
+ T.writes(T_multiply_red[v_ax0])
+ with T.init():
+ T_multiply_red[v_ax0] = T.float32(0.0)
+ T_multiply_red[v_ax0] =
T_multiply_red[v_ax0] + T_multiply[v_k0, v_ax0, v_k2, v_k3]
+ for ax0 in range(T.int64(3)):
+ with T.block("T_divide_1"):
+ v_ax0 = T.axis.spatial(T.int64(3), ax0)
+ T.reads(T_multiply_red[v_ax0])
+ T.writes(T_divide_1[v_ax0])
+ T_divide_1[v_ax0] = T_multiply_red[v_ax0] /
T.float32(1568.0)
Review Comment:
I see. Thank you for telling me about it
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
