Skip to content

Commit

Permalink
Support 5-input concat in XNNPACK delegate (pytorch#7401)
Browse files Browse the repository at this point in the history 
Summary:

I noticed that support for 5-input concatenate ops was added to the XNNPACK library subgraph layer. We can support this in the delegate.

Reviewed By: mcr229

Differential Revision: D67439458
  • Loading branch information
@GregoryComer @facebook-github-bot
GregoryComer authored and facebook-github-bot committed Dec 21, 2024
1 parent 6c3a792 commit 0f8f7c3
Show file tree
Hide file tree
Showing 7 changed files with 94 additions and 38 deletions.
15 changes: 15 additions & 0 deletions backends/xnnpack/operators/op_cat.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@
XNNConcatenate2,
XNNConcatenate3,
XNNConcatenate4,
XNNConcatenate5,
XNNGraph,
XNode,
)
Expand Down Expand Up @@ -71,6 +72,7 @@ def define_node(
input2_id=vals_to_ids[list_of_tensors[1]],
input3_id=XNN_INVALID_VALUE_ID,
input4_id=XNN_INVALID_VALUE_ID,
input5_id=XNN_INVALID_VALUE_ID,
output_id=vals_to_ids[node],
flags=0,
)
Expand All @@ -81,6 +83,7 @@ def define_node(
input2_id=vals_to_ids[list_of_tensors[1]],
input3_id=vals_to_ids[list_of_tensors[2]],
input4_id=XNN_INVALID_VALUE_ID,
input5_id=XNN_INVALID_VALUE_ID,
output_id=vals_to_ids[node],
flags=0,
)
Expand All @@ -91,6 +94,18 @@ def define_node(
input2_id=vals_to_ids[list_of_tensors[1]],
input3_id=vals_to_ids[list_of_tensors[2]],
input4_id=vals_to_ids[list_of_tensors[3]],
input5_id=XNN_INVALID_VALUE_ID,
output_id=vals_to_ids[node],
flags=0,
)
elif num_tensors_to_cat == 5:
xnode = XNNConcatenate5(
axis=axis,
input1_id=vals_to_ids[list_of_tensors[0]],
input2_id=vals_to_ids[list_of_tensors[1]],
input3_id=vals_to_ids[list_of_tensors[2]],
input4_id=vals_to_ids[list_of_tensors[3]],
input5_id=vals_to_ids[list_of_tensors[4]],
output_id=vals_to_ids[node],
flags=0,
)
Expand Down
6 changes: 3 additions & 3 deletions backends/xnnpack/partition/config/generic_node_configs.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -172,17 +172,17 @@ class CatConfig(GenericNodePartitionerConfig):

def check_constraints(self, node: torch.fx.Node, ep: ExportedProgram) -> bool:
"""
Only support concatenation of 2 - 4 tensors
Only support concatenation of 2 - 5 tensors
"""
if not self.check_common_constraints(node, ep):
return False

num_tensors = len(node.all_input_nodes)

if not (num_tensors >= 2 and num_tensors <= 4):
if not (num_tensors >= 2 and num_tensors <= 5):
why(
node,
reason=f"only support concatenation of 2 - 4 tensors, got {num_tensors} tensors",
reason=f"only support concatenation of 2 - 5 tensors, got {num_tensors} tensors",
)
return False

Expand Down
40 changes: 38 additions & 2 deletions backends/xnnpack/runtime/XNNCompiler.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1600,7 +1600,7 @@ Error defineConcatenate2Node(
}

/*
Defines serialized concatenate2 node into the subgraph,
Defines serialized concatenate3 node into the subgraph,
using the remapped ids to map the serialized ids,
to the new ids generated when defining the tensor value
*/
Expand Down Expand Up @@ -1633,7 +1633,7 @@ Error defineConcatenate3Node(
}

/*
Defines serialized concatenate2 node into the subgraph,
Defines serialized concatenate4 node into the subgraph,
using the remapped ids to map the serialized ids,
to the new ids generated when defining the tensor value
*/
Expand Down Expand Up @@ -1666,6 +1666,41 @@ Error defineConcatenate4Node(
return Error::Ok;
}

/*
Defines serialized concatenate5 node into the subgraph,
using the remapped ids to map the serialized ids,
to the new ids generated when defining the tensor value
*/
Error defineConcatenate5Node(
xnn_subgraph_t subgraph_ptr,
const std::unordered_map<uint32_t, uint32_t>& remapped_ids,
const NodePtr node,
const fb_xnnpack::XNNGraph* graph) noexcept {
MAYBE_UNUSED(graph);

auto graph_node = node->xnode_union_as_XNNConcatenate5();

xnn_status status = xnn_define_concatenate5(
subgraph_ptr,
graph_node->axis(),
remapped_ids.at(graph_node->input1_id()),
remapped_ids.at(graph_node->input2_id()),
remapped_ids.at(graph_node->input3_id()),
remapped_ids.at(graph_node->input4_id()),
remapped_ids.at(graph_node->input5_id()),
remapped_ids.at(graph_node->output_id()),
graph_node->flags());

ET_CHECK_OR_RETURN_ERROR(
status == xnn_status_success,
Internal,
"Failed to create cat5 node %i with code: %s",
node->debug_handle(),
xnn_status_to_string(status));

return Error::Ok;
}

/*
Defines serialized static_slice node into the subgraph,
using the remapped ids to map the serialized ids,
Expand Down Expand Up @@ -1832,6 +1867,7 @@ DefineNodeFunc getDefineNodeFunc(fb_xnnpack::XNodeUnion nodeType) {
_DEFINE(Concatenate2)
_DEFINE(Concatenate3)
_DEFINE(Concatenate4)
_DEFINE(Concatenate5)
_DEFINE(StaticSlice)
_DEFINE(ScaledDotProductAttention)
_DEFINE(BatchMatrixMultiply)
Expand Down
2 changes: 2 additions & 0 deletions backends/xnnpack/serialization/runtime_schema.fbs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -133,6 +133,7 @@ union XNodeUnion {
XNNConcatenate2: _XNNCat,
XNNConcatenate3: _XNNCat,
XNNConcatenate4: _XNNCat,
XNNConcatenate5: _XNNCat,
XNNStaticSlice,
XNNScaledDotProductAttention,
XNNBatchMatrixMultiply: _XNNNode2x1,
Expand Down Expand Up @@ -209,6 +210,7 @@ table _XNNCat {
input4_id: uint;
output_id: uint;
flags: uint;
input5_id: uint;
}

table XNNELU {
Expand Down
2 changes: 2 additions & 0 deletions backends/xnnpack/serialization/schema.fbs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -129,6 +129,7 @@ union XNodeUnion {
XNNConcatenate2: _XNNCat,
XNNConcatenate3: _XNNCat,
XNNConcatenate4: _XNNCat,
XNNConcatenate5: _XNNCat,
XNNStaticSlice,
XNNScaledDotProductAttention,
XNNBatchMatrixMultiply: _XNNNode2x1,
Expand Down Expand Up @@ -205,6 +206,7 @@ table _XNNCat {
input4_id: uint;
output_id: uint;
flags: uint;
input5_id: uint;
}

table XNNELU {
Expand Down
7 changes: 7 additions & 0 deletions backends/xnnpack/serialization/xnnpack_graph_schema.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,7 @@ class XNNCat:
input4_id: int
output_id: int
flags: int
input5_id: int


# Generic node data class for convolution type nodes
Expand Down Expand Up @@ -177,6 +178,11 @@ class XNNConcatenate4(XNNCat):
pass


@dataclass
class XNNConcatenate5(XNNCat):
pass


@dataclass
class XNNBatchMatrixMultiply(XNNNode2x1):
pass
Expand Down Expand Up @@ -357,6 +363,7 @@ class XNNScaledDotProductAttention:
XNNConcatenate2,
XNNConcatenate3,
XNNConcatenate4,
XNNConcatenate5,
XNNStaticSlice,
XNNScaledDotProductAttention,
XNNBatchMatrixMultiply,
Expand Down
60 changes: 27 additions & 33 deletions backends/xnnpack/test/ops/test_cat.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,27 +11,9 @@


class TestCat(unittest.TestCase):
class Cat2(torch.nn.Module):
def forward(self, arg1, arg2):
xs = [arg1, arg2]
x = torch.cat(xs)
return x + x # Quantize by propagation.

class Cat3(torch.nn.Module):
def forward(self, arg1, arg2, arg3):
xs = [arg1, arg2, arg3]
x = torch.cat(xs)
return x + x # Quantize by propagation.

class Cat4(torch.nn.Module):
def forward(self, arg1, arg2, arg3, arg4):
xs = [arg1, arg2, arg3, arg4]
x = torch.cat(xs)
return x + x # Quantize by propagation.

class Cat5(torch.nn.Module):
def forward(self, arg1, arg2, arg3, arg4, arg5):
xs = [arg1, arg2, arg3, arg4, arg5]
class Cat(torch.nn.Module):
def forward(self, *args):
xs = [*args]
x = torch.cat(xs)
return x + x # Quantize by propagation.

Expand Down Expand Up @@ -84,7 +66,7 @@ def test_fp16_cat2(self):
torch.randn(1, 2, 3).to(torch.float16),
torch.randn(3, 2, 3).to(torch.float16),
)
self._test_cat(self.Cat2(), inputs)
self._test_cat(self.Cat(), inputs)

def test_fp16_cat3(self):
"""
Expand All @@ -95,7 +77,7 @@ def test_fp16_cat3(self):
torch.randn(3, 2, 3).to(torch.float16),
torch.randn(2, 2, 3).to(torch.float16),
)
self._test_cat(self.Cat3(), inputs)
self._test_cat(self.Cat(), inputs)

def test_fp16_cat4(self):
"""
Expand All @@ -107,15 +89,15 @@ def test_fp16_cat4(self):
torch.randn(2, 2, 3).to(torch.float16),
torch.randn(5, 2, 3).to(torch.float16),
)
self._test_cat(self.Cat4(), inputs)
self._test_cat(self.Cat(), inputs)

def test_fp32_cat2(self):
inputs = (torch.randn(1, 2, 3), torch.randn(3, 2, 3))
self._test_cat(self.Cat2(), inputs)
self._test_cat(self.Cat(), inputs)

def test_fp32_cat3(self):
inputs = (torch.randn(1, 2, 3), torch.randn(3, 2, 3), torch.randn(2, 2, 3))
self._test_cat(self.Cat3(), inputs)
self._test_cat(self.Cat(), inputs)

def test_fp32_cat4(self):
inputs = (
Expand All @@ -124,15 +106,25 @@ def test_fp32_cat4(self):
torch.randn(2, 2, 3),
torch.randn(5, 2, 3),
)
self._test_cat(self.Cat4(), inputs)
self._test_cat(self.Cat(), inputs)

def test_fp32_cat5(self):
inputs = (
torch.randn(1, 2, 3),
torch.randn(3, 2, 3),
torch.randn(2, 2, 3),
torch.randn(5, 2, 3),
torch.randn(1, 2, 3),
)
self._test_cat(self.Cat(), inputs)

def test_qs8_cat2(self):
inputs = (torch.randn(1, 2, 3), torch.randn(3, 2, 3))
self._test_cat(self.Cat2(), inputs, cat_num=2, quant=True)
self._test_cat(self.Cat(), inputs, cat_num=2, quant=True)

def test_qs8_cat3(self):
inputs = (torch.randn(1, 2, 3), torch.randn(3, 2, 3), torch.randn(2, 2, 3))
self._test_cat(self.Cat3(), inputs, cat_num=3, quant=True)
self._test_cat(self.Cat(), inputs, cat_num=3, quant=True)

def test_qs8_cat4(self):
inputs = (
Expand All @@ -141,7 +133,7 @@ def test_qs8_cat4(self):
torch.randn(2, 2, 3),
torch.randn(5, 2, 3),
)
self._test_cat(self.Cat4(), inputs, cat_num=4, quant=True)
self._test_cat(self.Cat(), inputs, cat_num=4, quant=True)

def test_fp32_cat_unsupported(self):
"""
Expand All @@ -153,9 +145,10 @@ def test_fp32_cat_unsupported(self):
torch.randn(2, 2, 3),
torch.randn(5, 2, 3),
torch.randn(1, 2, 3),
torch.randn(2, 2, 3),
)
(
Tester(self.Cat5(), inputs)
Tester(self.Cat(), inputs)
.export()
.check_count({"torch.ops.aten.cat": 1})
.to_edge_transform_and_lower()
Expand All @@ -164,17 +157,18 @@ def test_fp32_cat_unsupported(self):

def test_fp32_cat_unsupported_legacy_mode(self):
"""
XNNPACK only supports concatenating up to 4 values, so it should not delegate here.
XNNPACK only supports concatenating up to 5 values, so it should not delegate here.
"""
inputs = (
torch.randn(1, 2, 3),
torch.randn(3, 2, 3),
torch.randn(2, 2, 3),
torch.randn(5, 2, 3),
torch.randn(1, 2, 3),
torch.randn(6, 2, 3),
)
(
Tester(self.Cat5(), inputs)
Tester(self.Cat(), inputs)
.export()
.check_count({"torch.ops.aten.cat": 1})
.to_edge()
Expand Down

0 comments on commit 0f8f7c3

Please sign in to comment.