Introduce --dimParams option to set onnx.dim_params attrs of dynamic model inputs, and extend dynamic backend tests on NNPA with this option to utilze NNPA #2781

negiyas · 2024-04-02T09:55:22Z

===== 2024/04/19 updated according to local discussions =====
This PR support the following two features.

(1) This PR introduces --dimParams option for onnx-mlir to set onnx.dim_params attrs of dynamic model inputs, which specifies relationships among dimensions of model inputs. For example, onnx-mlir cannot utilize NNPA for the following dynamic model, because broadcasting of %arg0 or %arg1 may be required and broadcasting is not supported by NNPA.

module {
func.func @main_graph(%arg0: tensor<?x?x?xf32> {onnx.name = "x"}, %arg1: tensor<?x?x?xf32> {onnx.name = "y"})
-> (tensor<3x4x5xf32> {onnx.name = "sum"}) {
%0 = "onnx.Add"(%arg0, %arg1) : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32>
onnx.Return %0 : tensor<?x?x?xf32>
}
"onnx.EntryPoint"() {func = @main_graph} : () -> ()
}

We can use the --dimParams option to specify that all corresponding dimensions between two dynamic inputs %arg0 and %arg1 are the same by --dimParams="0:0=a,1=b,2=c|1:0=a,1=b,2=c". Accordingly compiler can know no broadcasting is required and can utilize NNPA in this case.

module {
func.func @main_graph(%arg0: tensor<?x?x?xf32> {onnx.dim_params = "0:a,1:b,2:c", onnx.name = "x"}, %arg1: tensor<?x?x?xf32> {onnx.dim_params = "0:a,1:b,2:c", onnx.name = "y"}) -> (tensor<?x?x?xf32> {onnx.name = "sum"}) {
%0 = "onnx.Add"(%arg0, %arg1) : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32>
onnx.Return %0 : tensor<?x?x?xf32>
}
"onnx.EntryPoint"() {func = @main_graph} : () -> ()
}

(2) This PR also extends functions of dynamic backend tests for NNPA by using the --dimParams option
In backend tests on NNPA, support new notations for setting arguments of the option in the NNPA_TEST_LIST variable in the test/accelerators/NNPA/backend/CMakeLists.txt. The following examples mean as follows.

    test_add_cpu,zdnn_add,"0:0=a,1=b,2=c|1:0=a,1=b,2=c"
    test_averagepool_2d_default_cpu,zdnn_avgpool2d,NO_DYNAMIC_SHAPE_TEST

In the "test_add_cpu" test, "zdnn_add" should be called and the first, second and third dimensions of the first and second inputs are the same.
In the "test_averagepool_2d_default_cpu" test, dynamic shape test on NNPA should not be executed.

(e.g. "0:a,1:b,2:c|0:a,1:b,2:c") Signed-off-by: Yasushi Negishi <[email protected]>

Signed-off-by: Yasushi Negishi <[email protected]>

…_FORCE_DYNAMIC". Signed-off-by: Yasushi Negishi <[email protected]>

Signed-off-by: Yasushi Negishi <[email protected]>

tungld

Some first comments.

tungld · 2024-04-03T00:36:25Z

docs/Testing.md

 ```
 IMPORTER_FORCE_DYNAMIC='-1:-1' all dimensions of all the inputs will be changed
 IMPORTER_FORCE_DYNAMIC='0:-1' all dimensions of the first input will be changed
 IMPORTER_FORCE_DYNAMIC='0:-1|1:0,1' all dimensions of the first input and the 1st and 2nd dimensions of the second input will be changed
+IMPORTER_FORCE_DYNAMIC='0:0:a,1:b,2:c|1:0:a,1:b,2:c' the first three dimensions of the first of and the second inputs are changed. And assume that the first dimensions of the first and second arguments are the same, and same for the the second and third dimensions.


I know by using : we can pass dimIndex directly to an argument attribute, but I recommend to use a different symbol rather than : for dimIndex, because : has already been used to separate inputIndex and dimString. It confuses the parsers. How about - or =?

This example is quite trivial. Can we shuffle the symbols a bit, e.g. 0:0=a,1=b,2=c|1:0,1=c,2=b, and we can have a mix of <index> and <index> '=' <symbol>

@tungld I changed the notation according to this comment. Thanks for the suggestions!

tungld · 2024-04-03T00:40:18Z

docs/Testing.md

                 <index> ::= -1 | <number>
                <number> ::= <digit> | <digit><number>
                 <digit> ::= 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9
+                <symbol> ::= 'a', 'b', 'c', ...


I would add 'z' at the end to finish the list though it is still no perfect, e.g.
<symbol> ::= 'a', 'b', 'c', ..., 'z'

Fixed. Thanks.

tungld · 2024-04-03T00:47:01Z

docs/Testing.md

@@ -76,15 +80,18 @@ func @main_graph(%arg0: tensor<3x4x5xf32>, %arg1: tensor<3x4x5xf32>) -> tensor<3
 ```
 with `IMPORTER_FORCE_DYNAMIC='-1:-1'`, the result is:
 ```
-func @main_graph(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32>
+func @main_graph(%arg0: tensor<?x?x?xf32>{onnx.name = "x“, onnx.dim_params = "0:a,1:b,2:c"}, %arg1: tensor<?x?x?xf32>{onnx.name = "y“, onnx.dim_params = "0:a,1:b,2:c"}) -> tensor<?x?x?xf32>


-1 for dimIndex to say all arguments have the same dynamic dims does not make sense. It is too restricted.
Let's consider the dynamic tests for CPU where it supports broadcasting, e.g add(tensor<3x5xf32>, tensor<3x1xf32>, using -1:-1 will force all dynamic dimensions being the same and the lowering will generate different code, not taking care of broadcasting.

In summary, I recommend to not support this, and let users specify the same dynamic dimensions manually for the cases they need.

I removed the assumption about default relationships among input data ranks, and specified the relationship manually in the test/accelerators/NNPA/backend/CMakeLists.txt file for each dynamic backend tests on NNPA.
Thanks for the suggestions.

Signed-off-by: Yasushi Negishi <[email protected]>

tungld

Just some comments besides the main thing about introducing --dimParams as we discussed locally.

tungld · 2024-04-16T07:41:12Z

src/Builder/FrontendDialectTransformer.cpp

+        }
+      } else {
+        getParamStr(envStr, paramStrVec[idx]);
+      }


It looks like we can simplify this to avoid calling getParamStr, e.g.

envStr = envStr.substr(pos + 1); std::replace(envStr.begin(), envStr.end(), '=', ':'); if (idx >= 0) { paramStrVec[idx] = envStr; return; } // Idx < 0 means setting all arguments with the same value. for (size_t i = 0; i < numOfArgs; i++) paramStrVec[i] = envStr; return;

Fixed. Thanks!

tungld · 2024-04-16T07:42:31Z

src/Builder/FrontendDialectTransformer.cpp

+
+  // Get named attributesfrom IMPORTER_FORCE_DYNAMIC environment.
+  SmallVector<NamedAttribute> getNamedAttrFromImporterForceDynamic(
+      func::FuncOp funcOp, size_t numOfArgs) {


To avoid copying the result, it's better to pass argNamedAttrs as an argument, e.g.SmallVectorImp<NamedAttribute> &argNamedAttrs.

This code was cleaned up, Thanks.

tungld · 2024-04-16T07:43:44Z

src/Builder/FrontendDialectTransformer.cpp

+    getInputDimParamStrVec(
+        envInputString, funcOp, numOfArgs, inputDimParamStrVec);
+    for (size_t i = 0; i < numOfArgs; ++i) {
+      if (inputDimParamStrVec[i].length() != 0) {


Is it really happened? e.g. input dim param is an empty string.

This code was cleaned up. Thanks.

… tests on NNPA to utilze NNPA. Signed-off-by: Yasushi Negishi <[email protected]>

Signed-off-by: Yasushi Negishi <[email protected]>

tungld

Thank you for the update! The PR is in a good shape. I just put some comments to make the explanation easier to follow.

tungld · 2024-04-22T04:46:15Z

src/Builder/FrontendDialectTransformer.cpp

+  void getInputDimParamsVecFromOption(std::string optionStr, size_t numOfArgs,
+      SmallVector<std::string> &paramStrVec) {
+    std::stringstream paramStrStream;
+    paramStrStream << optionStr;


For brevity: std::stringstream paramStrStream(optionStr);

Fixed. Thanks!

tungld · 2024-04-22T04:59:03Z

src/Builder/FrontendDialectTransformer.cpp

@@ -490,6 +513,10 @@ class FrontendGenImpl {
    // See https://github.com/onnx/onnx/blob/main/docs/IR.md for more
    // information about dim_param.
    llvm::SmallVector<std::string, 4> inputDimParams, outputDimParams;
+    size_t numOfArgs = graph.input().size();
+    llvm::SmallVector<std::string> inputDimParamsFromOption(numOfArgs);


numOfArgs here may contain initializers also, so it does not count exactly the number of actual arguments to the main function. Please see inputIndex below that correctly represents the number of actual arguments.

Changed the code not to use numOfArgs in inputDimParamsFromOption . Thanks!

tungld · 2024-04-22T05:00:29Z

src/Builder/FrontendDialectTransformer.hpp

@@ -55,6 +55,16 @@ struct ImportOptions {
  //   - (arg0: tensor<3x4x5xf32>, arg1: tensor<10x5xf32>)
  //
  std::string shapeInformation = "";
+  // Custom onnx.dim_params attributes for the graph inputs for specifying
+  // relationship among their ranks.


their ranks looks confusing. To be precise, it would be their dynamic dimensions, given that dim in dim_params means dimension.

Fixed. Thanks!

tungld · 2024-04-22T05:03:19Z

src/Builder/FrontendDialectTransformer.hpp

+  // Its format is 'input_id:dim=sym,dim=sym,dim=sym|input_id:dim=sym,dim=sym..'
+  // E.g. An ONNX model has two dynamic inputs
+  //   - (arg0: tensor<?x5xf32>, arg1: tensor<?x5xf32>)
+  // If we want to specify that the first ranks of the first and second


If we want to specify that the first ranks of the first and second arguments are the same => If we want to specify that the first unknown dimension of arg0 and the first unknown dimension of arg1 are the same, we can assign the two dimensions to the same symbol "batch" as follows.

Fixed. Thanks.

tungld · 2024-04-22T05:07:35Z

src/Compiler/CompilerOptions.cpp

+static llvm::cl::opt<std::string, true> dimParamsOpt("dimParams",
+    llvm::cl::desc(
+        "Custom onnx.dim_params attributes for the inputs of the ONNX model for"
+        "specifying relationship among ranks of the inputs.\n"


among ranks of => among dynamic dimensions of

Fixed. Thanks.

tungld · 2024-04-22T05:21:58Z

test/accelerators/NNPA/backend/CMakeLists.txt

+# "zdnn_add" to check the function is called, (3) when dimParams is
+# "NO_DYNAMIC_SHAPE_TEST", backend test is skipped, otherwise the string is
+# passed as --dimParams option. "0:0=a,1=b,2=c|1:0=a,1=b,2=c" means that the
+# first ranks of the first, second and third ranks of the first and second


Could you make this explanation easier to follow? Don't use rank but dimension. Ranks in MLIR means the number of dimensions (in this case rank = 3).

Thanks for the comments. I fixed. I confirmed that all "rank" were removed modifications by this PR.

tungld · 2024-04-22T05:30:09Z

test/backend/common.py

@@ -132,6 +139,7 @@ def compile_model(model, emit):
            "--constants-to-file-total-threshold="
            + str(args.constants_to_file_total_threshold)
        )
+    command_list += get_compile_option(name + "_cpu")


I would have a if directly here, it's easy to follow and faster:

if args.dynamic and test_name in variables.test_to_enable_dimparams_dict: command_list.append("--dimParams=" + variables.test_to_enable_dimparams_dict[test_name])

Fixed. Thanks!

tungld · 2024-04-22T05:32:51Z

test/mlir/onnx/parse/test_add_dimparams.onnxtext

@@ -0,0 +1,23 @@
+// RUN: onnx-mlir --EmitONNXBasic --dimParams="0:0=a,1=b,2=c|1:0=a,1=b,2=c" --printIR %s | FileCheck %s


Thank you for adding this test! Really appreciate it.

tungld · 2024-04-22T05:34:04Z

src/Builder/FrontendDialectTransformer.hpp

@@ -55,6 +55,16 @@ struct ImportOptions {
  //   - (arg0: tensor<3x4x5xf32>, arg1: tensor<10x5xf32>)
  //
  std::string shapeInformation = "";
+  // Custom onnx.dim_params attributes for the graph inputs for specifying
+  // relationship among their ranks.
+  // Its format is 'input_id:dim=sym,dim=sym,dim=sym|input_id:dim=sym,dim=sym..'


dim_id would be better than dim

.. at the end would be three dots ...? Also that ... is for | meaning multiple input_id:dim=sym,dim=sym,dim=sym, or for , meaning multiple dim=sym,dim=sym?

Thanks. Fixed.

tungld · 2024-04-22T05:40:52Z

src/Compiler/CompilerOptions.cpp

+        "Custom onnx.dim_params attributes for the inputs of the ONNX model for"
+        "specifying relationship among ranks of the inputs.\n"
+        "\"value\" is in the format of "
+        "\"INPUT_ID1:D1=S1,D2=S2,...,Dn=Sn|INPUT_ID2:D1=T1,D2=T2,...Dn=Tn\""


missing a , before that last Dn.

It's better to add | ... at the end to denote multiple INPUT_ID, e.g.

INPUT_ID1:D1=S1,D2=S2,...,Dn=Sn | INPUT_ID2:D1=T1,D2=T2,...,Dn=Tn | ...

Fixed. Thanks.

Signed-off-by: Yasushi Negishi <[email protected]>

tungld · 2024-04-24T02:09:43Z

src/Compiler/CompilerOptions.cpp

+        "All dimensions of onnx.dim_params for a specified input index in "
+        "the original onnx model are cleared and repalced by this option. "
+        "onnx.dim_params for other input indices in the original onnx model "
+        "are not cleared"),


Thank you for adding the explanation!

tungld · 2024-04-24T02:19:16Z

src/Builder/FrontendDialectTransformer.cpp

+        paramStrForAllArgs = dimParamStr;
+      else {
+        while ((int)paramStrVec.size() <= idx) // Expand paramStrVec
+          paramStrVec.emplace_back("");


In the option definition you wrote onnx.dim_params for other input indices in the original onnx model are not cleared, but It looks to me only inputs whose index is larger than the max idx here are not cleared. And, the inputs whose index <= the max idx will be cleared.

For example, we have 4 input tensors

arg0: tensor<?x?xf32> {onnx.dim_param=0:x}, arg1: tensor<?x?xf32> {onnx.dim_param=0:y}, arg2: tensor<?x?xf32> {onnx.dim_param=0:z} arg3: tensor<?x?xf32> {onnx.dim_param=0:c,1:d}

with --dimParams="0:0=b,1=a|2:0=a,1=b", it becomes

arg0: tensor<?x?xf32> {onnx.dim_param=0:b,1:a}, arg1: tensor<?x?xf32>, arg2: tensor<?x?xf32> {onnx.dim_param=0:a,1:b} arg3: tensor<?x?xf32> {onnx.dim_param=0:c,1:d}

Is it correct?

@tungld Thanks for the comment.
It is not correct. onnx.dim_param of arg1 is not cleared with --dimParams="0:0=b,1=a|2:0=a,1=b".

// For each input tensor, use either all dimensions by the compiler // option OR all dimensions in the original onnx model. Dimensions // from the option and the model in a single input tensor are not // merged. if (inputIndex < (int)inputDimParamsFromOption.size() && !inputDimParamsFromOption[inputIndex].empty()) inputDimParams.emplace_back(inputDimParamsFromOption[inputIndex]); else if (!inputDimParamsFromOptionForAllArgs.empty()) inputDimParams.emplace_back(inputDimParamsFromOptionForAllArgs); else if (!dimParams.empty()) inputDimParams.emplace_back(dimParams);

The code checks if inputDimParamsFromOption[inputIndex] is empty string, and if it's empty, it uses dimParams (=onnx.dim_param in the original model), so that onnx.dim_param attr of arg1 in the original model is not cleared.

I added test/mlir/onnx/parse/dim_param_and_option.json to verify this situation.
In the test, With --dimParams="1:1=b", onnx.dim_params="0:A,1:B" of the first input argument in the original model, is not cleared with the option.

Original model

module { func.func @main_graph(%arg0: tensor<?x?xf32> {onnx.dim_params = "0:A,1:B", onnx.name = "X"}, %arg1: tensor<?x?xf32> {onnx.dim_params = "0:A,1:B", onnx.name = "Y"}) -> (tensor<?x?xf32> {onnx.dim_params = "0:A,1:B", onnx.name = "Z"}) { %0 = "onnx.Add"(%arg0, %arg1) : (tensor<?x?xf32>, tensor<?x?xf32>) -> tensor<?x?xf32> onnx.Return %0 : tensor<?x?xf32> } "onnx.EntryPoint"() {func = @main_graph} : () -> () }

Imported model

module { func.func @main_graph(%arg0: tensor<?x?xf32> {onnx.dim_params = "0:A,1:B", onnx.name = "X"}, %arg1: tensor<?x?xf32> {onnx.dim_params = "1:b", onnx.name = "Y"}) -> (tensor<?x?xf32> {onnx.dim_params = "0:A,1:B", onnx.name = "Z"}) { %0 = "onnx.Add"(%arg0, %arg1) : (tensor<?x?xf32>, tensor<?x?xf32>) -> tensor<?x?xf32> onnx.Return %0 : tensor<?x?xf32> } "onnx.EntryPoint"() {func = @main_graph} : () -> () }

I see. Thank you for the confirmation! Just a suggestion: if you use a map instead of vector for inputDimParamsFromOption, the code is easier to read and you don't need to expand the vector and check if (inputIndex < (int)inputDimParamsFromOption.size() && !inputDimParamsFromOption[inputIndex].empty()). Vector data structure looks not ideal here. It's a bit tricky to use an empty string.

I changed the code to use std::map<int, std::string> instead of SmallVector<std::string>
Thanks for the suggestion. It makes the code simpler.

Signed-off-by: Yasushi Negishi <[email protected]>

tungld

LGTM.

…rc/Builder/FrontendDialectTransformer.cpp Signed-off-by: Yasushi Negishi <[email protected]>

Signed-off-by: Yasushi Negishi <[email protected]>

jenkins-droid · 2024-04-30T02:36:52Z

Jenkins Linux ppc64le Build #13764 [push] Introduce --dimParams op... started at 22:48

jenkins-droid · 2024-04-30T02:36:52Z

Jenkins Linux amd64 Build #14739 [push] Introduce --dimParams op... started at 21:36

jenkins-droid · 2024-04-30T02:36:52Z

Jenkins Linux s390x Build #14769 [push] Introduce --dimParams op... started at 22:36

jenkins-droid · 2024-04-30T03:42:08Z

Jenkins Linux amd64 Build #14739 [push] Introduce --dimParams op... passed after 1 hr 5 min

jenkins-droid · 2024-04-30T04:06:46Z

Jenkins Linux s390x Build #14769 [push] Introduce --dimParams op... passed after 1 hr 29 min

jenkins-droid · 2024-04-30T04:39:54Z

Jenkins Linux ppc64le Build #13764 [push] Introduce --dimParams op... passed after 2 hr 3 min

AlexandreEichenberger · 2024-10-03T15:24:33Z

@hamptonm1 @Sunny-Anand Suggestion is to create issues so that the root cause of the model features will be evaluated with the goal of increasing our share of model zoo models successfully running with / without dynamic bounds.

negiyas added 4 commits March 28, 2024 09:18

Support new notation for IMPORTER_FORCE_DYNAMICIMPORTER_FORCE_DYNAMIC

12e640f

(e.g. "0:a,1:b,2:c|0:a,1:b,2:c") Signed-off-by: Yasushi Negishi <[email protected]>

Support argument number handling in "IMPORTER_FORCE_DYNAMIC".

dfeb993

Signed-off-by: Yasushi Negishi <[email protected]>

Fix backend test errors.

a4899f7

Signed-off-by: Yasushi Negishi <[email protected]>

Support symbols to specify relationships among arguments in "IMPORTER…

31a7d38

…_FORCE_DYNAMIC". Signed-off-by: Yasushi Negishi <[email protected]>

negiyas marked this pull request as draft April 2, 2024 09:55

negiyas added 2 commits April 2, 2024 09:58

Temporally disable failed dynamic backend tests on NNPA.

b0d38ec

Signed-off-by: Yasushi Negishi <[email protected]>

Merge branch 'main' into pr_extend_dynamic_backend_test

a579103

tungld reviewed Apr 3, 2024

View reviewed changes

negiyas added 6 commits April 3, 2024 07:21

Merge branch 'main' into pr_extend_dynamic_backend_test

4bd4fb3

Applying Tung-san's comments.

27f3892

Signed-off-by: Yasushi Negishi <[email protected]>

Merge branch 'main' into pr_extend_dynamic_backend_test

46c8985

Changing description for dynamicend tests.

bcda1b1

Signed-off-by: Yasushi Negishi <[email protected]>

Fix clang-format black format issues.

729b7e5

Signed-off-by: Yasushi Negishi <[email protected]>

Fix dynamic backend test errors on NNPA

1c6730f

Signed-off-by: Yasushi Negishi <[email protected]>

negiyas marked this pull request as ready for review April 15, 2024 09:18

negiyas changed the title ~~[WIP] Extend models for dynamic backend test with "onnx.dim_params" attributes in input arguments.~~ Extend models for dynamic backend test with "onnx.dim_params" attributes in input arguments. Apr 15, 2024

Add comments to write test/accelerators/NNPA/backend/CMakeLists.txt .

a95dd30

Signed-off-by: Yasushi Negishi <[email protected]>

tungld reviewed Apr 16, 2024

View reviewed changes

negiyas added 4 commits April 16, 2024 23:48

Merge branch 'main' into pr_extend_dynamic_backend_test

b0db459

Introduce --dimParams compiler option, and use it for dynamic backend…

6c5cec1

… tests on NNPA to utilze NNPA. Signed-off-by: Yasushi Negishi <[email protected]>

Fix format error in test/backend/common.py

49e7a6f

Signed-off-by: Yasushi Negishi <[email protected]>

Fix lit errors on test/mlir/onnx/parse/test_add_dimparams.onnxtext.

48a3b06

Signed-off-by: Yasushi Negishi <[email protected]>

negiyas changed the title ~~Extend models for dynamic backend test with "onnx.dim_params" attributes in input arguments.~~ Introduce --dimParams option to set onnx.dim_params attrs of dynamic model inputs, and extend dynamic backend tests on NNPA with this option to utilze NNPA Apr 19, 2024

tungld reviewed Apr 22, 2024

View reviewed changes

negiyas added 3 commits April 23, 2024 02:43

Code updated according to the PR comments.

0813d63

Signed-off-by: Yasushi Negishi <[email protected]>

Fix help compiler help messages.

67067d8

Signed-off-by: Yasushi Negishi <[email protected]>

Merge branch 'main' into pr_extend_dynamic_backend_test

2fe6881

tungld reviewed Apr 24, 2024

View reviewed changes

negiyas added 2 commits April 25, 2024 23:50

Merge branch 'main' into pr_extend_dynamic_backend_test

2f8924a

Add test/mlir/onnx/parse/dim_param_and_option.json test.

25a63f7

Signed-off-by: Yasushi Negishi <[email protected]>

tungld approved these changes Apr 26, 2024

View reviewed changes

negiyas added 2 commits April 26, 2024 08:47

Use std::map instead of SmallVector for inputDimParamsFromOption in s…

1fcca0b

…rc/Builder/FrontendDialectTransformer.cpp Signed-off-by: Yasushi Negishi <[email protected]>

Fix clang-format errors.

7e702a7

Signed-off-by: Yasushi Negishi <[email protected]>

negiyas merged commit 733dfac into onnx:main Apr 30, 2024
6 of 7 checks passed

negiyas mentioned this pull request May 8, 2024

Dynamic backend tests on NNPA cause errors #2817

Closed

		@@ -0,0 +1,23 @@
		// RUN: onnx-mlir --EmitONNXBasic --dimParams="0:0=a,1=b,2=c\|1:0=a,1=b,2=c" --printIR %s \| FileCheck %s

Introduce --dimParams option to set onnx.dim_params attrs of dynamic model inputs, and extend dynamic backend tests on NNPA with this option to utilze NNPA #2781

Introduce --dimParams option to set onnx.dim_params attrs of dynamic model inputs, and extend dynamic backend tests on NNPA with this option to utilze NNPA #2781

Conversation

negiyas commented Apr 2, 2024 • edited Loading

tungld left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tungld left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tungld left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tungld Apr 26, 2024 • edited Loading

Choose a reason for hiding this comment

negiyas Apr 26, 2024 • edited Loading

Choose a reason for hiding this comment

tungld left a comment

Choose a reason for hiding this comment

jenkins-droid commented Apr 30, 2024

jenkins-droid commented Apr 30, 2024

jenkins-droid commented Apr 30, 2024

jenkins-droid commented Apr 30, 2024

jenkins-droid commented Apr 30, 2024

jenkins-droid commented Apr 30, 2024

AlexandreEichenberger commented Oct 3, 2024

negiyas commented Apr 2, 2024 •

edited

Loading

tungld Apr 26, 2024 •

edited

Loading

negiyas Apr 26, 2024 •

edited

Loading