Skip to content

Add CUDA BitwiseXor integer kernels #33645

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

Open
Justin Chu (justinchuby) wants to merge 1 commit into
base: main
Choose a base branch
from
+115 −1
Open

Add CUDA BitwiseXor integer kernels #33645

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
1 change: 1 addition & 0 deletions docs/OperatorKernels.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -703,6 +703,7 @@ The **OpSet Version** column uses the following notation:
|||14|**T** = tensor(double), tensor(float), tensor(float16)<br/> **U** = tensor(double), tensor(float), tensor(float16)|
|||[9, 13]|**T** = tensor(double), tensor(float), tensor(float16)|
|||[7, 8]|**T** = tensor(double), tensor(float), tensor(float16)|
|BitwiseXor|*in* A:**T**<br> *in* B:**T**<br> *out* C:**T**|18+|**T** = tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|
|Cast|*in* input:**T1**<br> *out* output:**T2**|28+|**T1** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float4e2m1), tensor(float8e4m3fn), tensor(float8e5m2), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)<br/> **T2** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float4e2m1), tensor(float8e4m3fn), tensor(float8e5m2), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|
|||[25, 27]|**T1** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float4e2m1), tensor(float8e4m3fn), tensor(float8e5m2), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)<br/> **T2** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float4e2m1), tensor(float8e4m3fn), tensor(float8e5m2), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|
|||[23, 24]|**T1** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float4e2m1), tensor(float8e4m3fn), tensor(float8e5m2), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)<br/> **T2** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float4e2m1), tensor(float8e4m3fn), tensor(float8e5m2), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|
Expand Down
16 changes: 16 additions & 0 deletions onnxruntime/core/providers/cuda/cuda_execution_provider.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -721,6 +721,14 @@ class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kO
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, bool, And);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, bool, Or);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, bool, Xor);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, uint8_t, BitwiseXor);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, uint16_t, BitwiseXor);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, uint32_t, BitwiseXor);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, uint64_t, BitwiseXor);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int8_t, BitwiseXor);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int16_t, BitwiseXor);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int32_t, BitwiseXor);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int64_t, BitwiseXor);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 7, Sum);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 8, 12, Sum);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 11, Max);
Expand Down Expand Up @@ -2050,6 +2058,14 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, bool, And)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, bool, Or)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, bool, Xor)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, uint8_t, BitwiseXor)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, uint16_t, BitwiseXor)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, uint32_t, BitwiseXor)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, uint64_t, BitwiseXor)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int8_t, BitwiseXor)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int16_t, BitwiseXor)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int32_t, BitwiseXor)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 18, int64_t, BitwiseXor)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 7, Sum)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 8, 12, Sum)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 6, 11, Max)>,
Expand Down
8 changes: 8 additions & 0 deletions onnxruntime/core/providers/cuda/math/binary_elementwise_ops.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -292,6 +292,14 @@ BINARY_OP_REGISTER_VERSIONED_CLASS_HFD(Pow, Pow_7, 7, 11)
BINARY_LOGICALOP_TYPED(And, 7, bool)
BINARY_LOGICALOP_TYPED(Or, 7, bool)
BINARY_LOGICALOP_TYPED(Xor, 7, bool)
BINARY_OP_TYPED(BitwiseXor, 18, uint8_t)
BINARY_OP_TYPED(BitwiseXor, 18, uint16_t)
BINARY_OP_TYPED(BitwiseXor, 18, uint32_t)
BINARY_OP_TYPED(BitwiseXor, 18, uint64_t)
BINARY_OP_TYPED(BitwiseXor, 18, int8_t)
BINARY_OP_TYPED(BitwiseXor, 18, int16_t)
BINARY_OP_TYPED(BitwiseXor, 18, int32_t)
BINARY_OP_TYPED(BitwiseXor, 18, int64_t)
BINARY_OP_VERSIONED_HFD(PRelu, 7, 8)
BINARY_OP_VERSIONED_HFD(PRelu, 9, 15)
// Opset-16 adds BFloat16 to allowed types for the PRelu operator
Expand Down
7 changes: 7 additions & 0 deletions onnxruntime/core/providers/cuda/math/binary_elementwise_ops.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -200,6 +200,13 @@ class Xor final : public BinaryElementwise<ShouldBroadcast> {
Status ComputeInternal(OpKernelContext* context) const override;
};

template <typename T>
class BitwiseXor final : public BinaryElementwise<ShouldBroadcast> {
public:
BitwiseXor(const OpKernelInfo& info) : BinaryElementwise(info) {}
Status ComputeInternal(OpKernelContext* context) const override;
};

// PRelu is activation function, but it's closer to binary elementwise ops in implementation
template <typename T>
class PRelu final : public BinaryElementwise<ShouldBroadcast> {
Expand Down
8 changes: 8 additions & 0 deletions onnxruntime/core/providers/cuda/math/binary_elementwise_ops_impl.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -159,6 +159,14 @@ SPECIALIZED_BINARY_ELEMENTWISE_IMPL_HFD(Pow_7)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(And, bool)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(Or, bool)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(Xor, bool)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(BitwiseXor, uint8_t)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(BitwiseXor, uint16_t)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(BitwiseXor, uint32_t)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(BitwiseXor, uint64_t)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(BitwiseXor, int8_t)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(BitwiseXor, int16_t)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(BitwiseXor, int32_t)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL(BitwiseXor, int64_t)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL_HFD(PRelu)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL_UZILHFD(Max)
SPECIALIZED_BINARY_ELEMENTWISE_IMPL_UZILHFD(Min)
Expand Down
1 change: 1 addition & 0 deletions onnxruntime/core/providers/cuda/math/binary_elementwise_ops_impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,7 @@ namespace cuda {
BINARY_OP_NAME_EXPR(And, (a & b)) \
BINARY_OP_NAME_EXPR(Or, (a | b)) \
BINARY_OP_NAME_EXPR(Xor, (a ^ b)) \
BINARY_OP_NAME_EXPR(BitwiseXor, (a ^ b)) \
BINARY_OP_NAME_EXPR(PRelu, (a > (T)0 ? a : a * b)) \
BINARY_OP_NAME_EXPR(Max, _Max(a, b)) \
BINARY_OP_NAME_EXPR(Min, _Min(a, b)) \
Expand Down
75 changes: 74 additions & 1 deletion onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,14 +12,14 @@
#include "test/common/trt_op_test_utils.h"
#include "test/common/random_generator.h"
#include "core/util/math.h"
#include "core/session/onnxruntime_session_options_config_keys.h"
#include <algorithm>
#include <limits>
#include <math.h>
#include <type_traits>

#ifdef USE_DML
#include "core/providers/dml/dml_session_options_config_keys.h"
#include "core/session/onnxruntime_session_options_config_keys.h"
#endif

namespace onnxruntime {
Expand Down Expand Up @@ -5321,6 +5321,79 @@ TEST(MathOpTest, BitwiseXor_multidirectional_broadcastBA) {
test.Run();
}

#ifdef USE_CUDA
template <typename T>
class CudaBitwiseXorTest : public ::testing::Test {};

using BitwiseXorIntegerTypes = ::testing::Types<int8_t, uint8_t, int16_t, uint16_t,
int32_t, uint32_t, int64_t, uint64_t>;
TYPED_TEST_SUITE(CudaBitwiseXorTest, BitwiseXorIntegerTypes);

template <typename T>
void RunCudaBitwiseXor(const std::vector<int64_t>& lhs_shape, const std::vector<T>& lhs,
const std::vector<int64_t>& rhs_shape, const std::vector<T>& rhs,
const std::vector<int64_t>& output_shape, const std::vector<T>& expected) {
for (int opset : {18, 24}) {
auto cuda_ep = DefaultCudaExecutionProvider();
if (!cuda_ep) {
GTEST_SKIP() << "CUDA EP is unavailable";
}
OpTester test("BitwiseXor", opset);
test.AddInput<T>("A", lhs_shape, lhs);
test.AddInput<T>("B", rhs_shape, rhs);
test.AddOutput<T>("C", output_shape, expected);
SessionOptions options;
// A correct CPU fallback must not hide a missing CUDA registration.
ASSERT_STATUS_OK(options.config_options.AddConfigEntry(kOrtSessionOptionsDisableCPUEPFallback, "1"));
test.Config(options).ConfigEp(std::move(cuda_ep)).RunWithConfig();
}
}

TYPED_TEST(CudaBitwiseXorTest, IntegerBoundaries) {
using T = TypeParam;
const T low = std::numeric_limits<T>::min();
const T high = std::numeric_limits<T>::max();
const T all_bits = static_cast<T>(~T{0});
RunCudaBitwiseXor<T>({7}, {low, high, all_bits, low, high, T{0}, T{1}},
{7}, {high, high, T{0}, low, T{0}, all_bits, all_bits},
{7}, {all_bits, T{0}, all_bits, T{0}, high, all_bits, static_cast<T>(all_bits ^ T{1})});
}

TYPED_TEST(CudaBitwiseXorTest, ScalarBroadcastAndTail) {
using T = TypeParam;
std::vector<T> input(259);
std::vector<T> expected(259);
const T mask = std::numeric_limits<T>::max();
for (size_t i = 0; i < input.size(); ++i) {
input[i] = static_cast<T>(i % 127);
expected[i] = static_cast<T>(input[i] ^ mask);
}
RunCudaBitwiseXor<T>({}, {mask}, {259}, input, {259}, expected);
RunCudaBitwiseXor<T>({259}, input, {}, {mask}, {259}, expected);
RunCudaBitwiseXor<T>({}, {T{7}}, {}, {T{3}}, {}, {T{4}});
}

TYPED_TEST(CudaBitwiseXorTest, MultidirectionalBroadcast) {
using T = TypeParam;
RunCudaBitwiseXor<T>({2, 1, 2}, {10, 11, 12, 13}, {1, 3, 1}, {15, 7, 3},
{2, 3, 2}, {5, 4, 13, 12, 9, 8, 3, 2, 11, 10, 15, 14});
RunCudaBitwiseXor<T>({1, 3, 1}, {15, 7, 3}, {2, 1, 2}, {10, 11, 12, 13},
{2, 3, 2}, {5, 4, 13, 12, 9, 8, 3, 2, 11, 10, 15, 14});
}

TYPED_TEST(CudaBitwiseXorTest, EmptyBroadcast) {
using T = TypeParam;
RunCudaBitwiseXor<T>({0, 3}, {}, {1, 3}, {1, 2, 3}, {0, 3}, {});
RunCudaBitwiseXor<T>({}, {T{1}}, {2, 0, 3}, {}, {2, 0, 3}, {});
}

TEST(MathOpTest, BitwiseXor_CUDA_Int64RngState) {
const std::vector<int64_t> state{0, 1, 2047, 2048, 2049, 1073741824, 2147483647};
RunCudaBitwiseXor<int64_t>({7}, state, {7}, {0, 0, 0, 1, 1, 524288, 1048575},
{7}, {0, 1, 2047, 2049, 2048, 1074266112, 2146435072});
}
#endif

TEST(MathOpTest, BitwiseNot) {
OpTester test("BitwiseNot", 18);
std::vector<int64_t> dims{3};
Expand Down
Loading