Skip to content

tvm: fix tvm unrolling #124

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
guillon merged 1 commit into from
Sep 2, 2026
Merged

tvm: fix tvm unrolling #124

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
51 changes: 33 additions & 18 deletions src/xtc/backends/tvm/TVMScheduler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -375,21 +375,33 @@ def tvm_cache_read_factor_offset(

def tvm_update_loopnest_for_codegen(sched: LoopNest) -> LoopNest:
def _update_loopnode(node: LoopNestNode) -> LoopNestNode:
axis_dim = {axis: dim for dim, tiles in node.tiles.items() for axis in tiles}
adjusted_tiles = {}
adjusted_unrolling = {
k: v for k, v in node.unroll.items() if k not in node.vectorize
}
adjusted_unrolls = list(adjusted_unrolling)
adjusted_vectorization = node.vectorize[:]
adjusted_permutation = node.interchange[:]
for dim, dim_tiles in node.tiles.items():
dims_to_update = set(
[
axis_dim.get(axis, axis)
for axis in adjusted_unrolls + adjusted_vectorization
]
)
tiled_dims = list(node.tiles)
additional_dims = sorted(dims_to_update - set(tiled_dims))
for dim in tiled_dims + additional_dims:
dim_tiles = node.tiles.get(dim, {})
adjusted_dim_tiles = {}
for axis, size in dim_tiles.items():
adjusted_dim_tiles.update({axis: size})
for axis, size in {dim: 0, **dim_tiles}.items():
if size:
adjusted_dim_tiles.update({axis: size})
if axis in adjusted_unrolling:
assert axis not in adjusted_vectorization
assert axis in adjusted_permutation
unroll = adjusted_unrolling[axis]
if unroll < size:
if size == 0 or unroll < size:
axis_idx = adjusted_unrolls.index(axis)
new_axis = f"__u_{axis}"
adjusted_dim_tiles.update({new_axis: unroll})
Expand All @@ -402,20 +414,23 @@ def _update_loopnode(node: LoopNestNode) -> LoopNestNode:
)
elif axis in adjusted_vectorization:
assert axis not in adjusted_unrolling
pow2 = pow2divisor(size)
unroll = size // pow2
if unroll > 1:
axis_idx = adjusted_vectorization.index(axis)
new_axis = f"__v_{axis}"
adjusted_dim_tiles.update({new_axis: pow2})
adjusted_vectorization[axis_idx] = new_axis
adjusted_unrolls.append(axis)
adjusted_unrolling.update({axis: unroll})
adjusted_permutation.insert(
adjusted_permutation.index(axis) + 1,
new_axis,
)
adjusted_tiles[dim] = adjusted_dim_tiles
assert axis in adjusted_permutation
if size > 0:
pow2 = pow2divisor(size)
unroll = size // pow2
if unroll > 1:
axis_idx = adjusted_vectorization.index(axis)
new_axis = f"__v_{axis}"
adjusted_dim_tiles.update({new_axis: pow2})
adjusted_vectorization[axis_idx] = new_axis
adjusted_unrolls.append(axis)
adjusted_unrolling.update({axis: unroll})
adjusted_permutation.insert(
adjusted_permutation.index(axis) + 1,
new_axis,
)
if adjusted_dim_tiles:
adjusted_tiles[dim] = adjusted_dim_tiles
adjusted_unrolling = {u: adjusted_unrolling[u] for u in adjusted_unrolls}
updated_node = LoopNestNode(
root=node.root,
Expand Down
5 changes: 3 additions & 2 deletions tests/filecheck/backends/test_conv2d_r181_tvm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,11 +68,12 @@
# CHECK-NEXT: O = obj['O']
# CHECK-NEXT: b, h, w, f, = O.op.axis
# CHECK-NEXT: r, s, c, = O.op.reduce_axis
# CHECK-NEXT: c, __u_c = sch[O].split(c, factor=3)
# CHECK-NEXT: w, w1 = sch[O].split(w, factor=4)
# CHECK-NEXT: f, f1 = sch[O].split(f, factor=16)
# CHECK-NEXT: sch[O].reorder(b, h, w, f, r, s, c, w1, f1)
# CHECK-NEXT: sch[O].reorder(b, h, w, f, r, s, c, __u_c, w1, f1)
# CHECK-NEXT: sch[O].unroll(w1)
# CHECK-NEXT: sch[O].unroll(c)
# CHECK-NEXT: sch[O].unroll(__u_c)
# CHECK-NEXT: sch[O].vectorize(f1)
# CHECK-NEXT:
# CHECK-NEXT: # from tvm.script import ir as I
Expand Down
27 changes: 20 additions & 7 deletions tests/filecheck/backends/test_matmul_pack_tvm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,7 @@
sch.buffer_at("j")
sch.pack_at("k", 1, pad=True)
sch.vectorize(["j2"])
sch.unroll({"i3": 4})
sch.unroll({"i2": 4})
sched = sch.schedule()

comp = impl.get_compiler(
Expand Down Expand Up @@ -85,6 +85,7 @@
# CHECK-NEXT: sch[O_W0].reorder(k, i, j1, i1, k1, i2, j2)
# CHECK-NEXT: sch[I_R1].compute_at(sch[O_W0], k)
# CHECK-NEXT: sch[I_R1].storage_align(I_R1.op.axis[-2], factor=1024, offset=16)
# CHECK-NEXT: sch[O_W0].unroll(i2)
# CHECK-NEXT: sch[O_W0].vectorize(j2)
# CHECK-NEXT:
# CHECK-NEXT: # from tvm.script import ir as I
Expand All @@ -99,18 +100,30 @@
# CHECK-NEXT: _1_global = T.allocate([16640], "float32", "global")
# CHECK-NEXT: for j_outer in range(2):
# CHECK-NEXT: C_global_1 = T.Buffer((2048,), data=C_global)
# CHECK-NEXT: for i_c_outer_init, j_c_outer_init, i_c_inner_outer_init, i_c_inner_inner_init in T.grid(8, 2, 2, 4):
# CHECK-NEXT: C_global_1[i_c_outer_init * 256 + i_c_inner_outer_init * 128 + i_c_inner_inner_init * 32 + j_c_outer_init * 16:i_c_outer_init * 256 + i_c_inner_outer_init * 128 + i_c_inner_inner_init * 32 + j_c_outer_init * 16 + 16] = T.Broadcast(T.float32(0.0), 16)
# CHECK-NEXT: for i_c_outer_init, j_c_outer_init, i_c_inner_outer_init in T.grid(8, 2, 2):
# CHECK-NEXT: cse_var_1: T.int32 = i_c_outer_init * 256 + i_c_inner_outer_init * 128 + j_c_outer_init * 16
# CHECK-NEXT: C_global_1[cse_var_1:cse_var_1 + 16] = T.Broadcast(T.float32(0.0), 16)
# CHECK-NEXT: C_global_1[cse_var_1 + 32:cse_var_1 + 32 + 16] = T.Broadcast(T.float32(0.0), 16)
# CHECK-NEXT: C_global_1[cse_var_1 + 64:cse_var_1 + 64 + 16] = T.Broadcast(T.float32(0.0), 16)
# CHECK-NEXT: C_global_1[cse_var_1 + 96:cse_var_1 + 96 + 16] = T.Broadcast(T.float32(0.0), 16)
# CHECK-NEXT: for k_outer in range(4):
# CHECK-NEXT: _1_global_1 = T.Buffer((16640,), data=_1_global)
# CHECK-NEXT: for ax0, ax1 in T.grid(16, 32):
# CHECK-NEXT: _1_1 = T.Buffer((4096,), data=_1.data)
# CHECK-NEXT: _1_global_1[ax0 * 1040 + ax1] = _1_1[k_outer * 1024 + ax0 * 64 + j_outer * 32 + ax1]
# CHECK-NEXT: for i_c_outer, j_c_outer, i_c_inner_outer, k_inner, i_c_inner_inner in T.grid(8, 2, 2, 16, 4):
# CHECK-NEXT: cse_var_2: T.int32 = j_c_outer * 16
# CHECK-NEXT: cse_var_1: T.int32 = i_c_outer * 256 + i_c_inner_outer * 128 + i_c_inner_inner * 32 + cse_var_2
# CHECK-NEXT: for i_c_outer, j_c_outer, i_c_inner_outer, k_inner in T.grid(8, 2, 2, 16):
# CHECK-NEXT: cse_var_8: T.int32 = j_c_outer * 16
# CHECK-NEXT: cse_var_7: T.int32 = k_inner * 1040 + cse_var_8
# CHECK-NEXT: cse_var_6: T.int32 = i_c_outer * 256 + i_c_inner_outer * 128 + cse_var_8
# CHECK-NEXT: cse_var_5: T.int32 = i_c_outer * 512 + i_c_inner_outer * 256 + k_outer * 16 + k_inner
# CHECK-NEXT: cse_var_4: T.int32 = cse_var_6 + 96
# CHECK-NEXT: cse_var_3: T.int32 = cse_var_6 + 64
# CHECK-NEXT: cse_var_2: T.int32 = cse_var_6 + 32
# CHECK-NEXT: _0_1 = T.Buffer((4096,), data=_0.data)
# CHECK-NEXT: C_global_1[cse_var_1:cse_var_1 + 16] = C_global_1[cse_var_1:cse_var_1 + 16] + T.Broadcast(_0_1[i_c_outer * 512 + i_c_inner_outer * 256 + i_c_inner_inner * 64 + k_outer * 16 + k_inner], 16) * _1_global_1[k_inner * 1040 + cse_var_2:k_inner * 1040 + cse_var_2 + 16]
# CHECK-NEXT: C_global_1[cse_var_6:cse_var_6 + 16] = C_global_1[cse_var_6:cse_var_6 + 16] + T.Broadcast(_0_1[cse_var_5], 16) * _1_global_1[cse_var_7:cse_var_7 + 16]
# CHECK-NEXT: C_global_1[cse_var_2:cse_var_2 + 16] = C_global_1[cse_var_2:cse_var_2 + 16] + T.Broadcast(_0_1[cse_var_5 + 64], 16) * _1_global_1[cse_var_7:cse_var_7 + 16]
# CHECK-NEXT: C_global_1[cse_var_3:cse_var_3 + 16] = C_global_1[cse_var_3:cse_var_3 + 16] + T.Broadcast(_0_1[cse_var_5 + 128], 16) * _1_global_1[cse_var_7:cse_var_7 + 16]
# CHECK-NEXT: C_global_1[cse_var_4:cse_var_4 + 16] = C_global_1[cse_var_4:cse_var_4 + 16] + T.Broadcast(_0_1[cse_var_5 + 192], 16) * _1_global_1[cse_var_7:cse_var_7 + 16]
# CHECK-NEXT: for j_inner, i in T.grid(32, 64):
# CHECK-NEXT: C_1 = T.Buffer((4096,), data=C.data)
# CHECK-NEXT: C_1[i * 64 + j_outer * 32 + j_inner] = C_global_1[i * 32 + j_inner]
Expand Down
89 changes: 89 additions & 0 deletions tests/filecheck/backends/test_matmul_unroll_tvm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,89 @@
# RUN: python %s 2>&1 | filecheck %s
# REQUIRES: module_tvm

import xtc.graphs.xtc.op as O
from xtc.backends.tvm import Backend

I, J, K, dtype = 4, 32, 256, "float32"
a = O.tensor((I, K), dtype, name="A")
b = O.tensor((K, J), dtype, name="B")

with O.graph(name="matmul") as gb:
O.matmul(a, b, name="C")

graph = gb.graph
print(graph)

impl = Backend(graph)

sch = impl.get_scheduler()
sch.unroll({"k": 4})
sched = sch.schedule()

comp = impl.get_compiler(
shared_lib=True,
dump_file="matmul_unroll_tvm",
print_source_ir=True,
print_transformed_ir=True,
)

module = comp.compile(sched)
executor = module.get_executor(validate=True)
res = executor.execute()
print(f"CODE: {res}")

# CHECK: graph:
# CHECK-NEXT: name: matmul
# CHECK-NEXT: inputs:
# CHECK-NEXT: - %0 : 4x256xfloat32
# CHECK-NEXT: - %1 : 256x32xfloat32
# CHECK-NEXT: outputs:
# CHECK-NEXT: - %2 : 4x32xfloat32
# CHECK-NEXT: nodes:
# CHECK-NEXT: - %2: matmul(%0, %1) {name = 'C'} : [4x256xfloat32, 256x32xfloat32] -> [4x32xfloat32]
# CHECK-NEXT:
# CHECK-NEXT: # from tvm.script import ir as I
# CHECK-NEXT: # from tvm.script import tir as T
# CHECK-NEXT:
# CHECK-NEXT: @I.ir_module
# CHECK-NEXT: class Module:
# CHECK-NEXT: @T.prim_func
# CHECK-NEXT: def main(_0: T.Buffer((4, 256), "float32"), _1: T.Buffer((256, 32), "float32"), C: T.Buffer((4, 32), "float32")):
# CHECK-NEXT: T.func_attr({"from_legacy_te_schedule": T.bool(True), "tir.noalias": T.bool(True)})
# CHECK-NEXT: for i, j in T.grid(4, 32):
# CHECK-NEXT: C_1 = T.Buffer((128,), data=C.data)
# CHECK-NEXT: C_1[i * 32 + j] = T.float32(0.0)
# CHECK-NEXT: for k in range(256):
# CHECK-NEXT: cse_var_1: T.int32 = i * 32 + j
# CHECK-NEXT: _0_1 = T.Buffer((1024,), data=_0.data)
# CHECK-NEXT: _1_1 = T.Buffer((8192,), data=_1.data)
# CHECK-NEXT: C_1[cse_var_1] = C_1[cse_var_1] + _0_1[i * 256 + k] * _1_1[k * 32 + j]
# CHECK-NEXT: O = obj['C']
# CHECK-NEXT: i, j, = O.op.axis
# CHECK-NEXT: k, = O.op.reduce_axis
# CHECK-NEXT: k, __u_k = sch[O].split(k, factor=4)
# CHECK-NEXT: sch[O].reorder(i, j, k, __u_k)
# CHECK-NEXT: sch[O].unroll(__u_k)
# CHECK-NEXT:
# CHECK-NEXT: # from tvm.script import ir as I
# CHECK-NEXT: # from tvm.script import tir as T
# CHECK-NEXT:
# CHECK-NEXT: @I.ir_module
# CHECK-NEXT: class Module:
# CHECK-NEXT: @T.prim_func
# CHECK-NEXT: def main(_0: T.Buffer((4, 256), "float32"), _1: T.Buffer((256, 32), "float32"), C: T.Buffer((4, 32), "float32")):
# CHECK-NEXT: T.func_attr({"from_legacy_te_schedule": T.bool(True), "tir.noalias": T.bool(True)})
# CHECK-NEXT: for i, j in T.grid(4, 32):
# CHECK-NEXT: C_1 = T.Buffer((128,), data=C.data)
# CHECK-NEXT: C_1[i * 32 + j] = T.float32(0.0)
# CHECK-NEXT: for k_outer in range(64):
# CHECK-NEXT: cse_var_3: T.int32 = k_outer * 128 + j
# CHECK-NEXT: cse_var_2: T.int32 = i * 32 + j
# CHECK-NEXT: cse_var_1: T.int32 = i * 256 + k_outer * 4
# CHECK-NEXT: _0_1 = T.Buffer((1024,), data=_0.data)
# CHECK-NEXT: _1_1 = T.Buffer((8192,), data=_1.data)
# CHECK-NEXT: C_1[cse_var_2] = C_1[cse_var_2] + _0_1[cse_var_1] * _1_1[cse_var_3]
# CHECK-NEXT: C_1[cse_var_2] = C_1[cse_var_2] + _0_1[cse_var_1 + 1] * _1_1[cse_var_3 + 32]
# CHECK-NEXT: C_1[cse_var_2] = C_1[cse_var_2] + _0_1[cse_var_1 + 2] * _1_1[cse_var_3 + 64]
# CHECK-NEXT: C_1[cse_var_2] = C_1[cse_var_2] + _0_1[cse_var_1 + 3] * _1_1[cse_var_3 + 96]
# CHECK-NEXT: CODE: 0
99 changes: 99 additions & 0 deletions tests/filecheck/backends/test_matmul_vectorize_tvm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,99 @@
# RUN: python %s 2>&1 | filecheck %s
# REQUIRES: module_tvm

import xtc.graphs.xtc.op as O
from xtc.backends.tvm import Backend

I, J, K, dtype = 4, 32, 256, "float32"
a = O.tensor((I, K), dtype, name="A")
b = O.tensor((K, J), dtype, name="B")

with O.graph(name="matmul") as gb:
O.matmul(a, b, name="C")

graph = gb.graph
print(graph)

impl = Backend(graph)

sch = impl.get_scheduler()
sch.strip_mine("j", {"j0": 24})
sch.interchange(["i", "j", "k", "j0"])
sch.vectorize(["j0"])
sched = sch.schedule()

comp = impl.get_compiler(
shared_lib=True,
dump_file="matmul_vectorize_tvm",
print_source_ir=True,
print_transformed_ir=True,
)

module = comp.compile(sched)
executor = module.get_executor(validate=True)
res = executor.execute()
print(f"CODE: {res}")

# CHECK: graph:
# CHECK-NEXT: name: matmul
# CHECK-NEXT: inputs:
# CHECK-NEXT: - %0 : 4x256xfloat32
# CHECK-NEXT: - %1 : 256x32xfloat32
# CHECK-NEXT: outputs:
# CHECK-NEXT: - %2 : 4x32xfloat32
# CHECK-NEXT: nodes:
# CHECK-NEXT: - %2: matmul(%0, %1) {name = 'C'} : [4x256xfloat32, 256x32xfloat32] -> [4x32xfloat32]
# CHECK-NEXT:
# CHECK-NEXT: # from tvm.script import ir as I
# CHECK-NEXT: # from tvm.script import tir as T
# CHECK-NEXT:
# CHECK-NEXT: @I.ir_module
# CHECK-NEXT: class Module:
# CHECK-NEXT: @T.prim_func
# CHECK-NEXT: def main(_0: T.Buffer((4, 256), "float32"), _1: T.Buffer((256, 32), "float32"), C: T.Buffer((4, 32), "float32")):
# CHECK-NEXT: T.func_attr({"from_legacy_te_schedule": T.bool(True), "tir.noalias": T.bool(True)})
# CHECK-NEXT: for i, j in T.grid(4, 32):
# CHECK-NEXT: C_1 = T.Buffer((128,), data=C.data)
# CHECK-NEXT: C_1[i * 32 + j] = T.float32(0.0)
# CHECK-NEXT: for k in range(256):
# CHECK-NEXT: cse_var_1: T.int32 = i * 32 + j
# CHECK-NEXT: _0_1 = T.Buffer((1024,), data=_0.data)
# CHECK-NEXT: _1_1 = T.Buffer((8192,), data=_1.data)
# CHECK-NEXT: C_1[cse_var_1] = C_1[cse_var_1] + _0_1[i * 256 + k] * _1_1[k * 32 + j]
# CHECK-NEXT: O = obj['C']
# CHECK-NEXT: i, j, = O.op.axis
# CHECK-NEXT: k, = O.op.reduce_axis
# CHECK-NEXT: j, j0 = sch[O].split(j, factor=24)
# CHECK-NEXT: j0, __v_j0 = sch[O].split(j0, factor=8)
# CHECK-NEXT: sch[O].reorder(i, j, k, j0, __v_j0)
# CHECK-NEXT: sch[O].unroll(j0)
# CHECK-NEXT: sch[O].vectorize(__v_j0)
# CHECK-NEXT:
# CHECK-NEXT: # from tvm.script import ir as I
# CHECK-NEXT: # from tvm.script import tir as T
# CHECK-NEXT:
# CHECK-NEXT: @I.ir_module
# CHECK-NEXT: class Module:
# CHECK-NEXT: @T.prim_func
# CHECK-NEXT: def main(_0: T.Buffer((4, 256), "float32"), _1: T.Buffer((256, 32), "float32"), C: T.Buffer((4, 32), "float32")):
# CHECK-NEXT: T.func_attr({"from_legacy_te_schedule": T.bool(True), "tir.noalias": T.bool(True)})
# CHECK-NEXT: for i, j_outer in T.grid(4, 2):
# CHECK-NEXT: C_1 = T.Buffer((128,), data=C.data)
# CHECK-NEXT: C_1[i * 32 + j_outer * 24:i * 32 + j_outer * 24 + 8] = T.Broadcast(T.float32(0.0), 8)
# CHECK-NEXT: if T.likely(j_outer < 1):
# CHECK-NEXT: C_1[i * 32 + j_outer * 24 + 8:i * 32 + j_outer * 24 + 8 + 8] = T.Broadcast(T.float32(0.0), 8)
# CHECK-NEXT: if T.likely(j_outer < 1):
# CHECK-NEXT: C_1[i * 32 + j_outer * 24 + 16:i * 32 + j_outer * 24 + 16 + 8] = T.Broadcast(T.float32(0.0), 8)
# CHECK-NEXT: for k in range(256):
# CHECK-NEXT: cse_var_2: T.int32 = j_outer * 24
# CHECK-NEXT: cse_var_1: T.int32 = i * 32 + cse_var_2
# CHECK-NEXT: _0_1 = T.Buffer((1024,), data=_0.data)
# CHECK-NEXT: _1_1 = T.Buffer((8192,), data=_1.data)
# CHECK-NEXT: C_1[cse_var_1:cse_var_1 + 8] = C_1[cse_var_1:cse_var_1 + 8] + T.Broadcast(_0_1[i * 256 + k], 8) * _1_1[k * 32 + cse_var_2:k * 32 + cse_var_2 + 8]
# CHECK-NEXT: if T.likely(j_outer < 1):
# CHECK-NEXT: cse_var_3: T.int32 = cse_var_1 + 8
# CHECK-NEXT: C_1[cse_var_3:cse_var_3 + 8] = C_1[cse_var_3:cse_var_3 + 8] + T.Broadcast(_0_1[i * 256 + k], 8) * _1_1[k * 32 + cse_var_2 + 8:k * 32 + cse_var_2 + 8 + 8]
# CHECK-NEXT: if T.likely(j_outer < 1):
# CHECK-NEXT: cse_var_4: T.int32 = cse_var_1 + 16
# CHECK-NEXT: C_1[cse_var_4:cse_var_4 + 8] = C_1[cse_var_4:cse_var_4 + 8] + T.Broadcast(_0_1[i * 256 + k], 8) * _1_1[k * 32 + cse_var_2 + 16:k * 32 + cse_var_2 + 16 + 8]
# CHECK-NEXT: CODE: 0
Loading