From f04f1d8aedf3441c6bb3fffb1b965872d2936797 Mon Sep 17 00:00:00 2001
From: vatsalas <vatsala.shanubhag@intel.com>
Date: Tue, 18 Mar 2025 14:06:05 +0530
Subject: [PATCH 1/2] implement qlinear softmax
---
.../op/com.microsoft/qlinear_activation.cpp | 26 +++++++++++++++++++
1 file changed, 26 insertions(+)
diff --git a/src/frontends/onnx/frontend/src/op/com.microsoft/qlinear_activation.cpp b/src/frontends/onnx/frontend/src/op/com.microsoft/qlinear_activation.cpp
index 850c24fab2ccc6..ae926e52d1cc94 100644
--- a/src/frontends/onnx/frontend/src/op/com.microsoft/qlinear_activation.cpp
+++ b/src/frontends/onnx/frontend/src/op/com.microsoft/qlinear_activation.cpp
@@ -12,9 +12,13 @@
#include "openvino/op/divide.hpp"
#include "openvino/op/maximum.hpp"
#include "openvino/op/multiply.hpp"
+#include "openvino/op/reshape.hpp"
+#include "openvino/op/shape_of.hpp"
#include "openvino/op/sigmoid.hpp"
+#include "openvino/op/softmax.hpp"
#include "openvino/op/subtract.hpp"
#include "utils/common.hpp"
+#include "utils/reshape.hpp"
using namespace ov::op;
@@ -102,10 +106,32 @@ ov::OutputVector qlinear_avg_pool(const ov::frontend::onnx::Node& node) {
});
}
+ov::OutputVector qlinear_softmax(const ov::frontend::onnx::Node& node) {
+ return qlinear_activation(node, [&](const std::shared_ptr<ov::Node>& input_dequantized) {
+ auto axis = node.get_attribute_value<int64_t>("axis", -1);
+ auto opset = node.get_attribute_value<int64_t>("opset");
+
+ auto shape = std::make_shared<v3::ShapeOf>(input_dequantized);
+
+ std::shared_ptr<ov::Node> softmax_result;
+ if (opset <= 12) {
+ const auto coerced_data = ov::op::util::flatten(input_dequantized, static_cast<int>(axis));
+ softmax_result = std::make_shared<v8::Softmax>(coerced_data, 1);
+ softmax_result = std::make_shared<v1::Reshape>(softmax_result, shape, false);
+ } else {
+ softmax_result = std::make_shared<v8::Softmax>(input_dequantized, axis);
+ }
+
+ return softmax_result;
+ });
+}
+
bool register_multiple_operators(void) {
ONNX_OP_M("QLinearSigmoid", OPSET_SINCE(1), com_microsoft::opset_1::qlinear_sigmoid, MICROSOFT_DOMAIN);
ONNX_OP_M("QLinearLeakyRelu", OPSET_SINCE(1), com_microsoft::opset_1::qlinear_leaky_relu, MICROSOFT_DOMAIN);
ONNX_OP_M("QLinearAveragePool", OPSET_SINCE(1), com_microsoft::opset_1::qlinear_avg_pool, MICROSOFT_DOMAIN);
+ ONNX_OP_M("QLinearSoftmax", OPSET_SINCE(1), com_microsoft::opset_1::qlinear_softmax, MICROSOFT_DOMAIN);
+
return true;
}
From aa22bfa15e564083bec090e3b9c5c0eaa48d396c Mon Sep 17 00:00:00 2001
From: vatsalas <vatsala.shanubhag@intel.com>
Date: Tue, 18 Mar 2025 14:07:20 +0530
Subject: [PATCH 2/2] test qlinear softmax
---
.../qlinear_softmax_opset12.prototxt | 101 +++++++++++++++++
.../qlinear_softmax_opset13.prototxt | 106 ++++++++++++++++++
.../tests/onnx_import_com_microsoft.in.cpp | 45 ++++++++
3 files changed, 252 insertions(+)
create mode 100644 src/frontends/onnx/tests/models/com.microsoft/qlinear_softmax_opset12.prototxt
create mode 100644 src/frontends/onnx/tests/models/com.microsoft/qlinear_softmax_opset13.prototxt
diff --git a/src/frontends/onnx/tests/models/com.microsoft/qlinear_softmax_opset12.prototxt b/src/frontends/onnx/tests/models/com.microsoft/qlinear_softmax_opset12.prototxt
new file mode 100644
index 00000000000000..466746708ea937
--- /dev/null
+++ b/src/frontends/onnx/tests/models/com.microsoft/qlinear_softmax_opset12.prototxt
@@ -0,0 +1,101 @@
+ir_version: 3
+producer_name: "OpenVINO ONNX Frontend"
+producer_version: ""
+model_version: 0
+graph {
+ name: "test_qlinear_softmax_opset12"
+
+ node {
+ input: "X"
+ input: "X_scale"
+ input: "X_zero_point"
+ input: "Y_scale"
+ input: "Y_zero_point"
+ output: "Y"
+ op_type: "QLinearSoftmax"
+ attribute {
+ name: "opset"
+ i: 12
+ type: INT
+ }
+ domain: "com.microsoft"
+ }
+
+ input {
+ name: "X"
+ type {
+ tensor_type {
+ elem_type: 3
+ shape {
+ dim { dim_value: 4 }
+ dim { dim_value: 5 }
+ }
+ }
+ }
+ }
+
+ input {
+ name: "X_scale"
+ type {
+ tensor_type {
+ elem_type: 1
+ shape {
+ dim { dim_value: 1 }
+ }
+ }
+ }
+ }
+
+ input {
+ name: "X_zero_point"
+ type {
+ tensor_type {
+ elem_type: 3
+ shape {
+ dim { dim_value: 1 }
+ }
+ }
+ }
+ }
+
+ input {
+ name: "Y_scale"
+ type {
+ tensor_type {
+ elem_type: 1
+ shape {
+ dim { dim_value: 1 }
+ }
+ }
+ }
+ }
+
+ input {
+ name: "Y_zero_point"
+ type {
+ tensor_type {
+ elem_type: 3
+ shape {
+ dim { dim_value: 1 }
+ }
+ }
+ }
+ }
+
+ output {
+ name: "Y"
+ type {
+ tensor_type {
+ elem_type: 3
+ shape {
+ dim { dim_value: 4 }
+ dim { dim_value: 5 }
+ }
+ }
+ }
+ }
+}
+
+opset_import {
+ version: 1
+}
diff --git a/src/frontends/onnx/tests/models/com.microsoft/qlinear_softmax_opset13.prototxt b/src/frontends/onnx/tests/models/com.microsoft/qlinear_softmax_opset13.prototxt
new file mode 100644
index 00000000000000..64b0e3ed4eb07d
--- /dev/null
+++ b/src/frontends/onnx/tests/models/com.microsoft/qlinear_softmax_opset13.prototxt
@@ -0,0 +1,106 @@
+ir_version: 3
+producer_name: "OpenVINO ONNX Frontend"
+producer_version: ""
+model_version: 0
+graph {
+ name: "test_qlinear_softmax"
+
+ node {
+ input: "X"
+ input: "X_scale"
+ input: "X_zero_point"
+ input: "Y_scale"
+ input: "Y_zero_point"
+ output: "Y"
+ op_type: "QLinearSoftmax"
+ attribute {
+ name: "axis"
+ i: -1
+ type: INT
+ }
+ attribute {
+ name: "opset"
+ i: 12
+ type: INT
+ }
+ domain: "com.microsoft"
+ }
+
+ input {
+ name: "X"
+ type {
+ tensor_type {
+ elem_type: 3
+ shape {
+ dim { dim_value: 2 }
+ dim { dim_value: 3 }
+ }
+ }
+ }
+ }
+
+ input {
+ name: "X_scale"
+ type {
+ tensor_type {
+ elem_type: 1
+ shape {
+ dim { dim_value: 1 }
+ }
+ }
+ }
+ }
+
+ input {
+ name: "X_zero_point"
+ type {
+ tensor_type {
+ elem_type: 3
+ shape {
+ dim { dim_value: 1 }
+ }
+ }
+ }
+ }
+
+ input {
+ name: "Y_scale"
+ type {
+ tensor_type {
+ elem_type: 1
+ shape {
+ dim { dim_value: 1 }
+ }
+ }
+ }
+ }
+
+ input {
+ name: "Y_zero_point"
+ type {
+ tensor_type {
+ elem_type: 3
+ shape {
+ dim { dim_value: 1 }
+ }
+ }
+ }
+ }
+
+ output {
+ name: "Y"
+ type {
+ tensor_type {
+ elem_type: 3
+ shape {
+ dim { dim_value: 2 }
+ dim { dim_value: 3 }
+ }
+ }
+ }
+ }
+}
+
+opset_import {
+ version: 1
+}
diff --git a/src/frontends/onnx/tests/onnx_import_com_microsoft.in.cpp b/src/frontends/onnx/tests/onnx_import_com_microsoft.in.cpp
index 170476aae05dd3..ee89750871481f 100644
--- a/src/frontends/onnx/tests/onnx_import_com_microsoft.in.cpp
+++ b/src/frontends/onnx/tests/onnx_import_com_microsoft.in.cpp
@@ -1740,3 +1740,48 @@ OPENVINO_TEST(${BACKEND_NAME}, onnx_com_microsoft_bias_add) {
test_case.run();
}
+
+OPENVINO_TEST(${BACKEND_NAME}, onnx_com_microsoft_qlinearsoftmax) {
+ const auto model = convert_model("com.microsoft/qlinear_softmax_opset13.onnx");
+ auto test_case = ov::test::TestCase(model, s_device);
+
+ const std::vector<int8_t> data_X{-50, -25, 0, 25, 50, 75};
+ const std::vector<float> x_scale{0.1f};
+ const std::vector<int8_t> x_zero_point{0};
+ const std::vector<float> y_scale{0.03f};
+ const std::vector<int8_t> y_zero_point{5};
+
+ const std::vector<int8_t> expected_output{5, 7, 35, 5, 7, 35};
+
+ test_case.add_input<int8_t>(Shape{2, 3}, data_X);
+ test_case.add_input<float>(Shape{1}, x_scale);
+ test_case.add_input<int8_t>(Shape{1}, x_zero_point);
+ test_case.add_input<float>(Shape{1}, y_scale);
+ test_case.add_input<int8_t>(Shape{1}, y_zero_point);
+
+ test_case.add_expected_output<int8_t>(Shape{2, 3}, expected_output);
+ test_case.run();
+}
+
+OPENVINO_TEST(${BACKEND_NAME}, onnx_com_microsoft_qlinearsoftmax_opset12) {
+ const auto model = convert_model("com.microsoft/qlinear_softmax_opset12.onnx");
+ auto test_case = ov::test::TestCase(model, s_device);
+
+ const std::vector<int8_t> data_X{-60, -40, -20, 0, 20, 40, 60, 80, 100, 120,
+ -10, -5, 0, 5, 10, -80, -60, -40, -20, 0};
+ const std::vector<float> x_scale{0.15f};
+ const std::vector<int8_t> x_zero_point{0};
+ const std::vector<float> y_scale{0.05f};
+ const std::vector<int8_t> y_zero_point{3};
+
+ const std::vector<int8_t> expected_output{3, 3, 3, 3, 22, 3, 3, 3, 3, 22, 3, 4, 5, 8, 13, 3, 3, 3, 3, 22};
+
+ test_case.add_input<int8_t>(Shape{4, 5}, data_X);
+ test_case.add_input<float>(Shape{1}, x_scale);
+ test_case.add_input<int8_t>(Shape{1}, x_zero_point);
+ test_case.add_input<float>(Shape{1}, y_scale);
+ test_case.add_input<int8_t>(Shape{1}, y_zero_point);
+
+ test_case.add_expected_output<int8_t>(Shape{4, 5}, expected_output);
+ test_case.run();
+}