Skip to content

[ARITH] Canonicalize mul-coefficient to rhs #18031

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
spectrometerHBH merged 2 commits into from
Jun 2, 2025
Merged

[ARITH] Canonicalize mul-coefficient to rhs #18031

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
492fc5b
[ARITH] Canonicalize mul-coefficient to rhs
tqchen Jun 2, 2025
695b2da
Fix the grad testcase caused by the changed simplification behavior
tqchen Jun 2, 2025
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
12 changes: 6 additions & 6 deletions src/arith/rewrite_simplify.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -446,10 +446,10 @@ PrimExpr RewriteSimplifier::Impl::VisitExpr_(const AddNode* op) {
// mul co-efficient folding
TVM_TRY_REWRITE(x + x, x * 2);

TVM_TRY_REWRITE(matches_one_of(x * y + x, y * x + x, x + y * x, x + x * y), x * (y + 1));
TVM_TRY_REWRITE(matches_one_of(x * y + x, y * x + x, x + y * x, x + x * y), (y + 1) * x);

TVM_TRY_REWRITE(matches_one_of(x * y + x * z, y * x + x * z, x * y + z * x, y * x + z * x),
x * (y + z));
(y + z) * x);

// DivMod rules
// truc div
Expand Down Expand Up @@ -563,12 +563,12 @@ PrimExpr RewriteSimplifier::Impl::VisitExpr_(const SubNode* op) {
TVM_TRY_REWRITE(matches_one_of(max(x, y) - y, x - min(y, x)), max(x - y, 0));
TVM_TRY_REWRITE(matches_one_of(x - min(x, y), max(y, x) - y), max(0, x - y));

// mul co-efficient folding
// mul co-efficient folding: pefer co-effiicent to stay at rhs
TVM_TRY_REWRITE(x - x, ZeroWithTypeLike(x));
TVM_TRY_REWRITE(matches_one_of(x * y - x, y * x - x), x * (y - 1));
TVM_TRY_REWRITE(matches_one_of(x - y * x, x - x * y), x * (1 - y));
TVM_TRY_REWRITE(matches_one_of(x * y - x, y * x - x), (y - 1) * x);
TVM_TRY_REWRITE(matches_one_of(x - y * x, x - x * y), (1 - y) * x);
TVM_TRY_REWRITE(matches_one_of(x * y - x * z, y * x - x * z, x * y - z * x, y * x - z * x),
x * (y - z));
(y - z) * x);

// constant cancelation
TVM_TRY_REWRITE((x + c1) - c2, x + (c1 - c2));
Expand Down
28 changes: 14 additions & 14 deletions tests/python/arith/test_arith_rewrite_simplify.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -391,28 +391,28 @@ class TestAddIndex(BaseCompare):
TestCase(tvm.te.max(2 - x * 4, 0) + x * 4, tvm.te.max(x * 4, 2)),
TestCase(tvm.te.min(0, 1 - x * 4) + x * 4, tvm.te.min(x * 4, 1)),
TestCase(tvm.te.min(2 - x * 4, 0) + x * 4, tvm.te.min(x * 4, 2)),
TestCase(x * y + x * 10, x * (y + 10)),
TestCase(y * x + x * 10, x * (y + 10)),
TestCase(y * x + 10 * x, x * (y + 10)),
TestCase(x * y + 10 * x, x * (y + 10)),
TestCase(x * y + x * 10, (y + 10) * x),
TestCase(y * x + x * 10, (y + 10) * x),
TestCase(y * x + 10 * x, (y + 10) * x),
TestCase(x * y + 10 * x, (y + 10) * x),
TestCase((2 * z) + tvm.te.min(x, y - (2 * z)), tvm.te.min(x + (z * 2), y)),
TestCase(y * x + x, x * (y + 1)),
TestCase(x * y + x, x * (y + 1)),
TestCase(y * x + x, (y + 1) * x),
TestCase(x * y + x, (y + 1) * x),
TestCase((x + 10) + 13, x + 23),
TestCase((x + 10) + (13 + z), x + z + 23),
TestCase(x * y + 10 * x, x * (y + 10)),
TestCase(y * x + x * 3, x * (y + 3)),
TestCase(x * y + 10 * x, (y + 10) * x),
TestCase(y * x + x * 3, (y + 3) * x),
TestCase(x + 3 + y, x + y + 3),
TestCase((3 - y) + x, x - y + 3),
# canonicalization
TestCase(x + 2 + 3 + 4 + x, x * 2 + 9),
TestCase(x + 2 + 3 + 4 + x * 3, x * 4 + 9),
# DivMod rules
# trunc div
TestCase(y * tmod(x, 8) + 10 * tmod(x, 8), tmod(x, 8) * (y + 10)),
TestCase(y * tmod(x, 8) + 10 * tmod(x, 8), (y + 10) * tmod(x, 8)),
TestCase(tdiv(x, 8) * 8 + tmod(x, 8), x),
# floor div
TestCase(y * flm(x, 8) + 10 * flm(x, 8), flm(x, 8) * (y + 10)),
TestCase(y * flm(x, 8) + 10 * flm(x, 8), (y + 10) * flm(x, 8)),
TestCase(fld(x, 8) * 8 + flm(x, 8), x),
TestCase(fld(flm(x, 2) + 7, 2) + fld(x, 2), fld(x + 7, 2)),
)
Expand All @@ -436,10 +436,10 @@ class TestSubIndex(BaseCompare):
TestCase(y - tvm.te.max(x, y), tvm.te.min(y - x, 0)),
# mul co-efficient foldng
TestCase(x - x, 0),
TestCase(x * y - x, x * (y + (-1))),
TestCase(x * y - 10 * x, x * (y + (-10))),
TestCase(y * x - x * z, x * (y - z)),
TestCase(y * x - z * x, x * (y - z)),
TestCase(x * y - x, (y + (-1)) * x),
TestCase(x * y - 10 * x, (y + (-10)) * x),
TestCase(y * x - x * z, (y - z) * x),
TestCase(y * x - z * x, (y - z) * x),
TestCase(x + 10 - 20, x + (-10)),
# 4-operands pattern
TestCase((x + y) - (x + z), y - z),
Expand Down
13 changes: 13 additions & 0 deletions tests/python/arith/test_arith_simplify.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -131,5 +131,18 @@ def test_regression_simplify_inf_recursion():
ana.rewrite_simplify(res)


def test_simplify_floor_mod_with_linear_offset():
"""
Test that the floor_mod is simplified correctly when the offset is linear.
"""
ana = tvm.arith.Analyzer()
past_decoder_sequence_length = tir.Var("past_decoder_sequence_length", "int64")
expr1 = (past_decoder_sequence_length + 1) * 64
divisor1 = (past_decoder_sequence_length + 1) * 32
assert ana.can_prove_equal(tvm.tir.floormod(expr1, divisor1), 0)
divisor2 = 32 * (past_decoder_sequence_length + 1)
assert ana.can_prove_equal(tvm.tir.floormod(expr1, divisor2), 0)


if __name__ == "__main__":
tvm.testing.main()
3 changes: 2 additions & 1 deletion tests/python/relax/test_transform_legalize_ops_grad.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -282,7 +282,8 @@ def avg_pool2d_backward(rxplaceholder: T.Buffer((T.int64(3), T.int64(2), T.int64
T.writes(T_pool_grad[v_ax0, v_ax1, v_ax2, v_ax3])
with T.init():
T_pool_grad[v_ax0, v_ax1, v_ax2, v_ax3] = T.float32(0)
T_pool_grad[v_ax0, v_ax1, v_ax2, v_ax3] = T_pool_grad[v_ax0, v_ax1, v_ax2, v_ax3] + T.if_then_else(T.Select(v_ax2 < T.int64(3), T.int64(0), T.Div(v_ax2 - T.int64(3), T.int64(2)) + T.int64(1)) <= T.Div(v_ax2 + T.int64(2), T.int64(2)) - v_wh and T.Div(v_ax2 + T.int64(2), T.int64(2)) - v_wh < T.int64(6) and T.Select(v_ax3 < T.int64(4), T.int64(0), T.Div(v_ax3 - T.int64(4), T.int64(2)) + T.int64(1)) <= T.Div(v_ax3 + T.int64(1), T.int64(2)) - v_ww and T.Div(v_ax3 + T.int64(1), T.int64(2)) - v_ww < T.int64(5), rxplaceholder[v_ax0, v_ax1, T.Div(v_ax2 + T.int64(2), T.int64(2)) - v_wh, T.Div(v_ax3 + T.int64(1), T.int64(2)) - v_ww] / T.Cast("float32", T.max((T.min(T.Div(v_ax2 + T.int64(2), T.int64(2)) * T.int64(2) + T.int64(3) - v_wh * T.int64(2), T.int64(10)) - T.max(T.Div(v_ax2 + T.int64(2), T.int64(2)) * T.int64(2) - v_wh * T.int64(2) - T.int64(2), T.int64(0))) * (T.min(T.Div(v_ax3 + T.int64(1), T.int64(2)) * T.int64(2) + T.int64(4) - v_ww * T.int64(2), T.int64(10)) - T.max(T.Div(v_ax3 + T.int64(1), T.int64(2)) * T.int64(2) - v_ww * T.int64(2) - T.int64(1), T.int64(0))), T.int64(1))), T.float32(0))
T_pool_grad[v_ax0, v_ax1, v_ax2, v_ax3] = T_pool_grad[v_ax0, v_ax1, v_ax2, v_ax3] + T.if_then_else(T.Select(v_ax2 < T.int64(3), T.int64(0), T.Div(v_ax2 - T.int64(3), T.int64(2)) + T.int64(1)) <= T.Div(v_ax2 + T.int64(2), T.int64(2)) - v_wh and T.Div(v_ax2 + T.int64(2), T.int64(2)) - v_wh < T.int64(6) and T.Select(v_ax3 < T.int64(4), T.int64(0), T.Div(v_ax3 - T.int64(4), T.int64(2)) + T.int64(1)) <= T.Div(v_ax3 + T.int64(1), T.int64(2)) - v_ww and T.Div(v_ax3 + T.int64(1), T.int64(2)) - v_ww < T.int64(5), rxplaceholder[v_ax0, v_ax1, T.Div(v_ax2 + T.int64(2), T.int64(2)) - v_wh, T.Div(v_ax3 + T.int64(1), T.int64(2)) - v_ww] / T.Cast("float32", T.max((T.min(T.Div(v_ax2 + T.int64(2), T.int64(2)) * T.int64(2) + T.int64(3) - v_wh * T.int64(2), T.int64(10)) - T.max(T.Div(v_ax2 + T.int64(2), T.int64(2)) - v_wh - T.int64(1), T.int64(0)) * T.int64(2)) * (T.min(T.Div(v_ax3 + T.int64(1), T.int64(2)) * T.int64(2) + T.int64(4) - v_ww * T.int64(2), T.int64(10)) - T.max(T.Div(v_ax3 + T.int64(1), T.int64(2)) * T.int64(2) - v_ww * T.int64(2) - T.int64(1), T.int64(0))), T.int64(1))), T.float32(0.0))

@R.function
def main(output_grad: R.Tensor((3, 2, 6, 5), dtype="float32"), data: R.Tensor((3, 2, 10, 10), dtype="float32")) -> R.Tensor((3, 2, 10, 10), dtype="float32"):
cls = Expected
Expand Down
2 changes: 1 addition & 1 deletion tests/python/tir-transform/test_tir_transform_common_subexpr_elim.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -352,7 +352,7 @@ def test_no_normalization_without_commoning():
def func_distributivity(
B: T.Buffer((50,), "int32"), i1: T.int32, i2: T.int32, x: T.int32, y: T.int32, z: T.int32
) -> None:
B[i1] = x * (y + z)
B[i1] = (y + z) * x
B[i2] = x * y + x * z


Expand Down
Loading