Explore custom implementation of operator add

_unittests/ut__skl2onnx/test_sklearn_label_encoder_converter.py

@@ -50,7 +50,8 @@ def test_model_label_encoder_int(self):
         model = LabelEncoder()
         data = numpy.array([10, 3, 5, -34, 0], dtype=numpy.int64)
         model.fit(data)
-        for op in sorted(set([9, 10, 11, 12, 13, 14, TARGET_OPSET])):  # opset=13, 14, ...
+        # opset=13, 14, ...
+        for op in sorted(set([9, 10, 11, 12, 13, 14, TARGET_OPSET])):
             if op > TARGET_OPSET:
                 continue
             with self.subTest(opset=op):

_unittests/ut_onnxrt/test_benchmark_replay.py

@@ -17,7 +17,8 @@ def test_benchmark_replay(self):
             enumerate_benchmark_replay(temp, runtime='python')),
             FileNotFoundError)
         res = list(enumerate_validated_operator_opsets(
-            0, fLOG=None, models={"LogisticRegression"}, opset_min=14,  # opset=13, 14, ...
+            # opset=13, 14, ...
+            0, fLOG=None, models={"LogisticRegression"}, opset_min=14,
             opset_max=14, benchmark=False, store_models=True, dump_all=True,
             dump_folder=temp, filter_exp=lambda m, p: (
                 "64" not in p and "b-cl" in p and "dec" not in p)))

_unittests/ut_onnxrt/test_onnxrt_python_runtime_.py

@@ -2874,7 +2874,8 @@ def test_onnxt_runtime_reduce_prod(self):
     def test_onnxt_runtime_reduce_sum(self):
         X = numpy.array([[2, 1], [0, 1]], dtype=float)
 
-        for opset in (10, 11, 12, 13, 14, get_opset_number_from_onnx()):  # opset=13, 14, ...
+        # opset=13, 14, ...
+        for opset in (10, 11, 12, 13, 14, get_opset_number_from_onnx()):
             if onnx_opset_version() < opset:
                 continue
             if opset < 13:
@@ -2977,7 +2978,8 @@ def test_onnxt_runtime_reduce_sum_square(self):
     def test_onnxt_runtime_reduce_sum_noop(self):
         X = numpy.array([], dtype=float).reshape((2, 0))
 
-        for opset in (13, 14, get_opset_number_from_onnx()):  # opset=13, 14, ...
+        # opset=13, 14, ...
+        for opset in (13, 14, get_opset_number_from_onnx()):
             if onnx_opset_version() < opset:
                 continue
             cl = OnnxReduceSum
@@ -3191,7 +3193,8 @@ def test_onnxt_runtime_slice(self):
 
     @wraplog()
     def test_onnxt_runtime_slice_step_none(self):
-        for opset in [13, 14, get_opset_number_from_onnx()]:  # opset=13, 14, ...
+        # opset=13, 14, ...
+        for opset in [13, 14, get_opset_number_from_onnx()]:
             if opset > get_opset_number_from_onnx():
                 continue
             with self.subTest(opset=opset):
@@ -3211,7 +3214,8 @@ def test_onnxt_runtime_slice_step_none(self):
 
     @wraplog()
     def test_onnxt_runtime_split(self):
-        for opset in [10, 11, 12, 13, 14, get_opset_number_from_onnx()]:  # opset=13, 14, ...
+        # opset=13, 14, ...
+        for opset in [10, 11, 12, 13, 14, get_opset_number_from_onnx()]:
             if opset > get_opset_number_from_onnx():
                 continue
             with self.subTest(opset=opset):
@@ -3259,7 +3263,8 @@ def test_onnxt_runtime_sqrt(self):
 
     @wraplog()
     def test_onnxt_runtime_squeeze(self):
-        for opset in [10, 11, 12, 13, 14, get_opset_number_from_onnx()]:  # opset=13, 14, ...
+        # opset=13, 14, ...
+        for opset in [10, 11, 12, 13, 14, get_opset_number_from_onnx()]:
             if opset > get_opset_number_from_onnx():
                 continue
             with self.subTest(opset=opset):
@@ -3444,7 +3449,8 @@ def test_onnxt_runtime_transpose(self):
 
     @wraplog()
     def test_onnxt_runtime_unsqueeze(self):
-        for opset in [10, 11, 12, 13, 14, get_opset_number_from_onnx()]:  # opset=13, 14, ...
+        # opset=13, 14, ...
+        for opset in [10, 11, 12, 13, 14, get_opset_number_from_onnx()]:
             if opset > get_opset_number_from_onnx():
                 continue
             with self.subTest(opset=opset):

_unittests/ut_testing/test_custom_add.py

@@ -0,0 +1,50 @@
+"""
+@brief      test log(time=8s)
+"""
+import unittest
+import numpy
+from pyquickhelper.pycode import ExtTestCase
+from mlprodict.testing.experimental_c import (  # pylint: disable=E0611
+    BroadcastMatrixAddLeftInplaceDouble,
+    BroadcastMatrixAddLeftInplaceFloat,
+    BroadcastMatrixAddLeftInplaceInt64)
+
+
+class TestCustomAdd(ExtTestCase):
+
+    add_dtypes = {
+        numpy.float64: BroadcastMatrixAddLeftInplaceDouble,
+        numpy.float32: BroadcastMatrixAddLeftInplaceFloat,
+        numpy.int64: BroadcastMatrixAddLeftInplaceInt64
+    }
+
+    def _common_broadcast_matrix(self, dt):
+        with self.subTest(dtype=dt):
+            fct = TestCustomAdd.add_dtypes[dt]
+
+            m1 = numpy.array([1, 2, 3, 4, 5, 6], dtype=dt).reshape((-1, 2))
+            m2 = numpy.array([1, 2], dtype=dt).reshape((1, 2))
+            m3 = m1 + m2
+            fct(m1, m2)
+            self.assertEqualArray(m3, m1)
+
+            m1 = numpy.array([1, 2, 3, 4, 5, 6], dtype=dt).reshape((-1, 3))
+            m2 = numpy.array([1, 2], dtype=dt).reshape((2, 1))
+            m3 = m1 + m2
+            fct(m1, m2)
+            self.assertEqualArray(m3, m1)
+
+            m1 = numpy.array([1, 2, 3, 4, 5, 6], dtype=dt).reshape((-1, 3))
+            m2 = numpy.array([1, 2], dtype=dt).reshape((2, ))
+            m3 = m1 + m2.reshape((2, 1))
+            fct(m1, m2)
+            self.assertEqualArray(m3, m1)
+
+    def test_broadcast_matrix(self):
+        for dt in [numpy.float64, numpy.float32, numpy.int64]:
+            self._common_broadcast_matrix(dt)
+
+
+if __name__ == "__main__":
+    # TestEinsum().test_np_test_broadcasting_dot_cases1()
+    unittest.main()

bin/debug/debug.cpp

@@ -5,6 +5,7 @@
 #include "op_qlinear_conv_.hpp"
 #include "op_qlinear_cpp_qgemm_tester_.hpp"
 #include "op_qlinear_cpp_tester_.hpp"
+#include "experimental_c.h"
 
 
 void test_qlinear_conv1() {
@@ -100,6 +101,9 @@ void test_qlinear_conv2(bool random) {
 }
 
 int main() {
+    experimental_ut_add();
+    experimental_ut_einsum();
+    experimental_ut_reduce();
     TestQGemm0();
     TestQGemm1();
     test_qlinear_conv2(false);

bin/debug/debug.vcxproj

@@ -116,7 +116,7 @@
       <SDLCheck>true</SDLCheck>
       <PreprocessorDefinitions>_DEBUG;_CONSOLE;%(PreprocessorDefinitions);SKIP_PYTHON</PreprocessorDefinitions>
       <ConformanceMode>true</ConformanceMode>
-      <AdditionalIncludeDirectories>../../mlprodict/onnxrt/ops_cpu</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>../../mlprodict/onnxrt/ops_cpu;../../mlprodict/testing</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -150,6 +150,16 @@
     <ClCompile Include="../../mlprodict/onnxrt/ops_cpu/op_conv_matrices_.hpp" />
     <ClCompile Include="../../mlprodict/onnxrt/ops_cpu/op_common_.cpp" />
     <ClCompile Include="../../mlprodict/onnxrt/ops_cpu/op_common_.hpp" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c_helper.h" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c_helper.hpp" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c_add.h" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c_add.hpp" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c_einsum.h" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c_einsum.hpp" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c_reduce.h" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c_reduce.hpp" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c.cpp" />
+    <ClCompile Include="../../mlprodict/testing/experimental_c.h" />
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">

mlprodict/onnxrt/ops_cpu/op_conv_matrices_.cpp

@@ -1,3 +1,6 @@
+// Inspired from 
+// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/nn/qlinearconv.cc.
+
 #include "op_conv_matrices_.hpp"
 
 

mlprodict/onnxrt/ops_cpu/op_conv_matrices_.hpp

@@ -1,7 +1,7 @@
 #pragma once
 
 // Inspired from 
-// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/ml/tree_ensemble_classifier.cc.
+// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/nn/qlinearconv.cc.
 
 #if !defined(_CRT_SECURE_NO_WARNINGS)
 #define _CRT_SECURE_NO_WARNINGS

mlprodict/onnxrt/ops_cpu/op_qlinear_conv_.cpp

@@ -1,11 +1,10 @@
 // Inspired from 
-// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/ml/tree_ensemble_classifier.cc.
+// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/nn/qlinearconv.cc.
 
 #if !defined(_CRT_SECURE_NO_WARNINGS)
 #define _CRT_SECURE_NO_WARNINGS
 #endif
 
-
 #include "op_qlinear_conv_.hpp"
 #include "op_qlinear_cpp_tester_.hpp"
 #include "op_qlinear_cpp_qgemm_tester_.hpp"
@@ -20,7 +19,6 @@
 #undef max
 #endif
 
-
 class RuntimeTesterQLinearConv : public RuntimeTester {
 public:
     RuntimeTesterQLinearConv(const char* op_name, int opset = 13) : RuntimeTester(op_name, opset) {}

mlprodict/onnxrt/ops_cpu/op_qlinear_conv_.hpp

@@ -1,7 +1,7 @@
 #pragma once
 
 // Inspired from 
-// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/ml/tree_ensemble_classifier.cc.
+// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/nn/qlinearconv.cc.
 
 #if !defined(_CRT_SECURE_NO_WARNINGS)
 #define _CRT_SECURE_NO_WARNINGS
@@ -49,8 +49,7 @@ inline uint32_t BitsOfFp32(float FloatValue) {
 */
 template <typename T>
 void RequantizeOutput(
-    const int32_t* Input, T* Output, const int32_t* Bias,
-    size_t M, size_t N,
+    const int32_t* Input, T* Output, const int32_t* Bias, size_t M, size_t N,
     const float* Scale, bool PerColumnScale, T ZeroPoint) {
     const float PerMatrixScaleValue = PerColumnScale ? 0.0f : *Scale;
     const float MinimumValue = float(0 - ZeroPoint);

mlprodict/onnxrt/ops_cpu/op_qlinear_cpp_qgemm_tester_.cpp

@@ -1,5 +1,5 @@
 // Inspired from 
-// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/ml/tree_ensemble_classifier.cc.
+// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/test/providers/cpu/nn/qlinearconv_op_test.cc.
 
 #if !defined(_CRT_SECURE_NO_WARNINGS)
 #define _CRT_SECURE_NO_WARNINGS

mlprodict/onnxrt/ops_cpu/op_qlinear_cpp_qgemm_tester_.hpp

@@ -1,7 +1,7 @@
 #pragma once
 
 // Inspired from 
-// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/ml/tree_ensemble_classifier.cc.
+// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/test/providers/cpu/nn/qlinearconv_op_test.cc.
 
 #if !defined(_CRT_SECURE_NO_WARNINGS)
 #define _CRT_SECURE_NO_WARNINGS

mlprodict/onnxrt/ops_cpu/op_qlinear_cpp_tester_.hpp

@@ -1,7 +1,7 @@
 #pragma once
 
 // Inspired from 
-// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/ml/tree_ensemble_classifier.cc.
+// https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/test/providers/cpu/nn/qlinearconv_op_test.cc.
 
 #if !defined(_CRT_SECURE_NO_WARNINGS)
 #define _CRT_SECURE_NO_WARNINGS