Skip to content

[AutoTVM][TOPI] AutoTVM incorrect measurement #5511

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
kevinthesun merged 3 commits into from
May 26, 2020
Merged

[AutoTVM][TOPI] AutoTVM incorrect measurement #5511

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ce6ea37
[AutoTVM][TOPI] AutoTVM incorrect measurement
cchung100m May 5, 2020
64d1e6a
create new placeholder with converted layout
cchung100m May 25, 2020
1629b56
update _schedule_winograd
cchung100m May 26, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 6 additions & 9 deletions topi/python/topi/mali/conv2d.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -138,13 +138,14 @@ def _schedule_spatial_pack(cfg, s, output, conv, data_vec, kernel_vec):
s[data_vec].unroll(vw)

if isinstance(kernel_vec.op, tvm.te.ComputeOp) and kernel_vec.name == 'kernel_vec':
co, ci, kh, kw, vc = s[kernel_vec].op.axis
if autotvm.GLOBAL_SCOPE.in_tuning:
# kernel packing will be pre-computed during compilation, so we skip
# this part to make tuning records correct
s[kernel_vec].pragma(s[kernel_vec].op.axis[0], 'debug_skip_region')
# Directly use modified data layout placeholder.
kvshape = (co // vc, ci, kh, kw, vc)
kernel_vec = tvm.te.placeholder(kvshape, kernel_vec.dtype, name="kernel")
s[kernel_vec] = kernel_vec
else:
max_threads = tvm.target.Target.current(allow_none=False).max_num_threads
co, ci, kh, kw, vc = s[kernel_vec].op.axis
fused = s[kernel_vec].fuse(co, ci, kh, kw, vc)
fused, vec = s[kernel_vec].split(fused, VC)
bb, tt = s[kernel_vec].split(fused, max_threads)
Expand Down Expand Up @@ -345,11 +346,7 @@ def _schedule_winograd(cfg, s, op):
kernel, G = s[U].op.input_tensors
s[G].compute_inline()
eps, nu, co, ci, vco, = s[U].op.axis
if autotvm.GLOBAL_SCOPE.in_tuning:
# kernel transformation will be pre-computed during compilation, so we skip
# this part to make tuning records correct
s[U].pragma(eps, 'debug_skip_region')
else:
if not autotvm.GLOBAL_SCOPE.in_tuning:
r_kh, r_kw = s[U].op.reduce_axis
s[U].reorder(co, ci, eps, nu, r_kh, r_kw, vco)
_ = [s[U].unroll(x) for x in [eps, nu, r_kh, r_kw]]
Expand Down