[NNAPI QDQ] Add QDQ Resize support #10442
Conversation
onnxruntime/core/providers/nnapi/nnapi_builtin/builders/op_support_checker.cc
Show resolved
Hide resolved
| return false; | ||
| if (!HasValidQuantizationZeroPoints(initializers, node_unit, {0}, false /* is_input */)) | ||
| return false; | ||
| } |
There was a problem hiding this comment.
As this is also validating inputs/output maybe it could be called something like IsValidQuantizedOp.
There was a problem hiding this comment.
Or maybe all the validation should be in a separate func given we seem to do some here and some in ResizeOpBuilder::AddToModelBuilderImpl
There was a problem hiding this comment.
Yes, there will be a separated change to merge all these HasValidQuantizationScales and HasValidQuantizationZeroPoints into one single helper function, plus the input and output type check,
Will be added in a later PR.
The validation of quantized IO params are different in SupportChecker and OPbuilder, the one in SupportChecker only checks if the quantized scale and zp are initializers and some other checks such as per-channel setting, the one in OPBuilder will also verify if these quantization params are matching the existing numbers in the NNAPI model, this part cannot be done in SupportChecker.
Will figure out better names for these in a later PR
| }; | ||
|
|
||
| /* static */ bool ResizeOpBuilder::IsQuantizedOp(const NodeUnit& node_unit) { | ||
| static bool is_quant_op_type = |
There was a problem hiding this comment.
static variable? so this function returns the same result each time?
There was a problem hiding this comment.
same comment for ResizeOpSupportChecker::IsQuantizedOp()
There was a problem hiding this comment.
Yep, this is incorrect, typical result of late night coding :(, updated both places
| auto check_matmul_graph = [&](InferenceSessionWrapper& session) { | ||
| auto op_to_count = CountOpsInGraph(session.GetGraph()); | ||
| EXPECT_EQ(op_to_count["Resize"], 1); | ||
| EXPECT_EQ(op_to_count["QuantizeLinear"], 0); | ||
| EXPECT_EQ(op_to_count["DequantizeLinear"], 0); | ||
| }; | ||
|
|
||
| TransformerTester(build_test_case, check_matmul_graph, | ||
| TransformerTester(BuildQDQResizeTestCase(input1_shape, sizes_shape), | ||
| check_matmul_graph, |
There was a problem hiding this comment.
Will update in my next PR
Description: [NNAPI QDQ] Add QDQ Conv support
Motivation and Context
This is part of task 6, to add support of QDQ Resize which does not have
QLinearversion