Skip to content

[QNN EP] Support more Einsum equation: bhwc,hkc->bhwk #25518

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
HectorSVC merged 1 commit into from
Jul 28, 2025
Merged

[QNN EP] Support more Einsum equation: bhwc,hkc->bhwk #25518

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
61 changes: 52 additions & 9 deletions onnxruntime/core/providers/qnn/builder/opbuilder/einsum_op_builder.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,13 +45,7 @@
if (term_1.empty() || term_2.empty()) {
return std::nullopt;
}
if (term_1.size() < 2) {
return std::nullopt;
}
if (term_1.size() != term_2.size()) {
return std::nullopt;
}
if (term_1.size() != result.size()) {
if (term_1.size() < 2 || term_2.size() < 2 || result.size() < 2) {
return std::nullopt;
}
if (!std::all_of(term_1.begin(), term_1.end(), [](unsigned char c) { return std::islower(c); })) {
Expand Down Expand Up @@ -154,6 +148,50 @@
return true;
}

bool IsEquationMatMulBroadcastTransposeY(const Equation& equation) {
// E.g., bhwc,hkc->bhwk
const auto& [term_1, term_2, result] = equation;
const size_t term1_dims = term_1.size();
if (term1_dims != 4) {
return false;
}
const size_t term2_dims = term_2.size();
if (term2_dims != 3) {
return false;
}
const size_t result_dims = result.size();
if (result_dims != 4) {
return false;
}
// Check matrix multiplication dimensions
char term_1_m = term_1[term1_dims - 2];
char term_1_k = term_1[term1_dims - 1];
char term_2_k = term_2[term2_dims - 1];
char term_2_n = term_2[term2_dims - 2];
char result_m = result[result_dims - 2];
char result_n = result[result_dims - 1];
if (term_1_m != result_m) {
return false;
}
if (term_1_k != term_2_k) {
return false;
}
if (term_2_n != result_n) {
return false;
}
// Check batch dimensions
if (term_1[0] != result[0]) {
return false;
}
if (term_1[1] != result[1]) {
return false;
}
if (term_2[0] != result[1]) {
return false;
}
return true;
}

/**
* @brief Sets the parameter tensor names for a MatMul op.
*
Expand Down Expand Up @@ -317,6 +355,7 @@
}
if (!IsEquationMatMul(parsed_equation.value()) &&
!IsEquationMatMulTransposeY(parsed_equation.value()) &&
!IsEquationMatMulBroadcastTransposeY(parsed_equation.value()) &&
!IsEquationMatMulTransposeAll(parsed_equation.value())) {
return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, node_unit.OpType() + " unsupported equation: " + equation);
}
Expand Down Expand Up @@ -353,7 +392,8 @@
/*logger=*/logger,
/*do_op_validation=*/do_op_validation,
/*qnn_op_type=*/QNN_OP_MAT_MUL));
} else if (IsEquationMatMulTransposeY(parsed_equation.value())) {
} else if (IsEquationMatMulTransposeY(parsed_equation.value()) ||
IsEquationMatMulBroadcastTransposeY(parsed_equation.value())) {
std::vector<std::string> param_tensor_names = SetMatMulParamTensorNames(
&qnn_model_wrapper, node_unit, /*transpose_in0=*/false, /*transpose_in1=*/true);
ORT_RETURN_IF_ERROR(ProcessOutputs(/*qnn_model_wrapper=*/qnn_model_wrapper,
Expand All @@ -364,7 +404,10 @@
/*do_op_validation=*/do_op_validation,
/*qnn_op_type=*/QNN_OP_MAT_MUL));
} else if (IsEquationMatMulTransposeAll(parsed_equation.value())) {
ORT_RETURN_IF_ERROR(CreateMatMulTransposeAll(&qnn_model_wrapper, node_unit, std::move(input_names), do_op_validation));
ORT_RETURN_IF_ERROR(CreateMatMulTransposeAll(/*qnn_model_wrapper=*/&qnn_model_wrapper,
/*node_unit=*/node_unit,
/*input_names=*/std::move(input_names),

Check warning on line 409 in onnxruntime/core/providers/qnn/builder/opbuilder/einsum_op_builder.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Add #include <utility> for move  [build/include_what_you_use] [4]

Raw Output:
onnxruntime/core/providers/qnn/builder/opbuilder/einsum_op_builder.cc:409:  Add #include <utility> for move  [build/include_what_you_use] [4]
/*do_op_validation=*/do_op_validation));
} else {
return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, node_unit.OpType() + " unsupported equation: " + equation);
}
Expand Down
52 changes: 52 additions & 0 deletions onnxruntime/test/providers/qnn/einsum_op_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,6 +189,19 @@ TEST_F(QnnCPUBackendTests, EinsumRank4MatMulTransposeAll1) {
/*tolerance=*/1e-4f);
}

TEST_F(QnnCPUBackendTests, EinsumMatMulBroadcastTransposeY) {
const std::vector<int64_t> shape0{2, 3, 3, 4};
const std::vector<int64_t> shape1{3, 3, 4};
const std::vector<float> data0 = GetSequentialFloatData(shape0, /*start=*/-0.1f, /*step=*/0.05f);
const std::vector<float> data1 = GetSequentialFloatData(shape1, /*start=*/-0.1f, /*step=*/0.05f);
RunQnnEinsum<float>(
/*backend=*/kQnnBackendTypeCpu,
/*in0=*/TestInputDef<float>(shape0, /*is_initializer=*/false, std::move(data0)),
/*in1=*/TestInputDef<float>(shape1, /*is_initializer=*/false, std::move(data1)),
/*equation=*/"bhwc,hkc->bhwk",
/*tolerance=*/1e-4f);
}

TEST_F(QnnCPUBackendTests, EinsumRank4MatMulTransposeAll2) {
const std::vector<int64_t> shape0{1, 7, 1, 7};
const std::vector<int64_t> shape1{1, 9, 1, 7};
Expand Down Expand Up @@ -273,6 +286,19 @@ TEST_F(QnnHTPBackendTests, EinsumF16Rank4MatMulTransposeAll2) {
/*tolerance=*/1e-2f);
}

TEST_F(QnnHTPBackendTests, EinsumF16MatMulBroadcastTransposeY) {
const std::vector<int64_t> shape0{2, 3, 3, 4};
const std::vector<int64_t> shape1{3, 3, 4};
const std::vector<float> data0 = GetSequentialFloatData(shape0, /*start=*/-0.1f, /*step=*/0.05f);
const std::vector<float> data1 = GetSequentialFloatData(shape1, /*start=*/-0.1f, /*step=*/0.05f);
RunQnnEinsum<float>(
/*backend=*/kQnnBackendTypeHtp,
/*in0=*/TestInputDef<float>(shape0, /*is_initializer=*/false, std::move(data0)),
/*in1=*/TestInputDef<float>(shape1, /*is_initializer=*/false, std::move(data1)),
/*equation=*/"bhwc,hkc->bhwk",
/*tolerance=*/1e-2f);
}

//
// QNN HTP QDQ
//
Expand Down Expand Up @@ -337,6 +363,18 @@ TEST_F(QnnHTPBackendTests, EinsumQdqRank4MatMulTransposeAll2) {
/*tolerance=*/QDQTolerance());
}

TEST_F(QnnHTPBackendTests, EinsumQdqMatMulBroadcastTransposeY) {
const std::vector<int64_t> shape0{2, 3, 3, 4};
const std::vector<int64_t> shape1{3, 3, 4};
const std::vector<float> data0 = GetSequentialFloatData(shape0, /*start=*/-0.1f, /*step=*/0.05f);
const std::vector<float> data1 = GetSequentialFloatData(shape1, /*start=*/-0.1f, /*step=*/0.05f);
RunQnnHtpQdqEinsum<uint8_t, uint8_t>(
/*in0=*/TestInputDef<float>(shape0, /*is_initializer=*/false, std::move(data0)),
/*in1=*/TestInputDef<float>(shape1, /*is_initializer=*/false, std::move(data1)),
/*equation=*/"bhwc,hkc->bhwk",
/*tolerance=*/QDQTolerance());
}

#endif // defined(__aarch64__) || defined(_M_ARM64) || defined(__linux__)

#if defined(_M_ARM64)
Expand Down Expand Up @@ -422,6 +460,20 @@ TEST_F(QnnGPUBackendTests, EinsumRank4MatMulTransposeAll2) {
/*tolerance=*/1e-4f);
}

// Numeric instability in GPU backend, see also MatMul tests.
TEST_F(QnnGPUBackendTests, DISABLED_EinsumMatMulBroadcastTransposeY) {
const std::vector<int64_t> shape0{2, 3, 3, 4};
const std::vector<int64_t> shape1{3, 3, 4};
const std::vector<float> data0 = GetSequentialFloatData(shape0, /*start=*/-0.1f, /*step=*/0.05f);
const std::vector<float> data1 = GetSequentialFloatData(shape1, /*start=*/-0.1f, /*step=*/0.05f);
RunQnnEinsum<float>(
/*backend=*/kQnnBackendTypeGpu,
/*in0=*/TestInputDef<float>(shape0, /*is_initializer=*/false, std::move(data0)),
/*in1=*/TestInputDef<float>(shape1, /*is_initializer=*/false, std::move(data1)),
/*equation=*/"bhwc,hkc->bhwk",
/*tolerance=*/1e-4f);
}

#endif // defined(_M_ARM64) GPU tests

} // namespace test
Expand Down
Loading