support sparse tensor dense matmul fot CPU
This commit is contained in:
parent
e8cb23e35e
commit
7df6bfe7dd
|
|
@ -0,0 +1,75 @@
|
||||||
|
/**
|
||||||
|
* Copyright 2020-2021 Huawei Technologies Co., Ltd
|
||||||
|
*
|
||||||
|
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
* you may not use this file except in compliance with the License.
|
||||||
|
* You may obtain a copy of the License at
|
||||||
|
*
|
||||||
|
* http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
*
|
||||||
|
* Unless required by applicable law or agreed to in writing, software
|
||||||
|
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
* See the License for the specific language governing permissions and
|
||||||
|
* limitations under the License.
|
||||||
|
*/
|
||||||
|
|
||||||
|
#include "backend/kernel_compiler/cpu/sparse_tensor_dense_matmul_cpu_kernel.h"
|
||||||
|
|
||||||
|
namespace mindspore {
|
||||||
|
namespace kernel {
|
||||||
|
template <typename I, typename T>
|
||||||
|
void SparseTensorDenseMatmulCPUKernel<I, T>::InitKernel(const CNodePtr &kernel_node) {
|
||||||
|
output_size_ = 1;
|
||||||
|
auto output_shape = AnfAlgo::GetOutputInferShape(kernel_node, 0);
|
||||||
|
for (auto &dim : output_shape) {
|
||||||
|
output_size_ *= dim;
|
||||||
|
}
|
||||||
|
|
||||||
|
aValues_size_ = 1;
|
||||||
|
auto aValues_shape = AnfAlgo::GetInputDeviceShape(kernel_node, 1);
|
||||||
|
for (auto &dim : aValues_shape) {
|
||||||
|
aValues_size_ *= dim;
|
||||||
|
}
|
||||||
|
|
||||||
|
b_shape_ = AnfAlgo::GetInputDeviceShape(kernel_node, 3);
|
||||||
|
output_shape_ = AnfAlgo::GetOutputDeviceShape(kernel_node, 0);
|
||||||
|
}
|
||||||
|
|
||||||
|
template <typename I, typename T>
|
||||||
|
bool SparseTensorDenseMatmulCPUKernel<I, T>::Launch(const std::vector<kernel::AddressPtr> &inputs,
|
||||||
|
const std::vector<kernel::AddressPtr> & /*workspace*/,
|
||||||
|
const std::vector<kernel::AddressPtr> &outputs) {
|
||||||
|
auto a_indices = reinterpret_cast<I *>(inputs[0]->addr);
|
||||||
|
auto a_values = reinterpret_cast<T *>(inputs[1]->addr);
|
||||||
|
auto b = reinterpret_cast<T *>(inputs[3]->addr);
|
||||||
|
auto out = reinterpret_cast<T *>(outputs[0]->addr);
|
||||||
|
|
||||||
|
memset(out, 0, output_size_);
|
||||||
|
|
||||||
|
const size_t nnz = aValues_size_;
|
||||||
|
const size_t rhs_right = b_shape_[1];
|
||||||
|
const size_t lhs_right = b_shape_[0];
|
||||||
|
|
||||||
|
for (size_t i = 0; i < nnz; ++i) {
|
||||||
|
const size_t m = a_indices[i * 2];
|
||||||
|
const size_t k = a_indices[i * 2 + 1];
|
||||||
|
|
||||||
|
if (k > lhs_right) {
|
||||||
|
MS_LOG(ERROR) << "Invalid value: k: " << k << ", lhs_right: " << lhs_right;
|
||||||
|
return false;
|
||||||
|
}
|
||||||
|
if (m > output_shape_[0]) {
|
||||||
|
MS_LOG(ERROR) << "Invalid value: m: " << m << ", output_shape: " << output_shape_[0];
|
||||||
|
return false;
|
||||||
|
}
|
||||||
|
|
||||||
|
for (size_t n = 0; n < rhs_right; ++n) {
|
||||||
|
const float b_value = b[k * lhs_right + n];
|
||||||
|
out[m * output_shape_[0] + n] += a_values[i] * b_value;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
return true;
|
||||||
|
}
|
||||||
|
} // namespace kernel
|
||||||
|
} // namespace mindspore
|
||||||
|
|
@ -0,0 +1,243 @@
|
||||||
|
/**
|
||||||
|
* Copyright 2020-2021 Huawei Technologies Co., Ltd
|
||||||
|
*
|
||||||
|
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
* you may not use this file except in compliance with the License.
|
||||||
|
* You may obtain a copy of the License at
|
||||||
|
*
|
||||||
|
* http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
*
|
||||||
|
* Unless required by applicable law or agreed to in writing, software
|
||||||
|
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
* See the License for the specific language governing permissions and
|
||||||
|
* limitations under the License.
|
||||||
|
*/
|
||||||
|
|
||||||
|
#ifndef MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_SPARSE_TENSOR_DENSE_MATMUL_CPU_KERNEL_H_
|
||||||
|
#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_SPARSE_TENSOR_DENSE_MATMUL_CPU_KERNEL_H_
|
||||||
|
|
||||||
|
#include <vector>
|
||||||
|
#include "backend/kernel_compiler/cpu/cpu_kernel.h"
|
||||||
|
#include "backend/kernel_compiler/cpu/cpu_kernel_factory.h"
|
||||||
|
|
||||||
|
namespace mindspore {
|
||||||
|
namespace kernel {
|
||||||
|
template <typename I, typename T>
|
||||||
|
class SparseTensorDenseMatmulCPUKernel : public CPUKernel {
|
||||||
|
public:
|
||||||
|
SparseTensorDenseMatmulCPUKernel() = default;
|
||||||
|
~SparseTensorDenseMatmulCPUKernel() override = default;
|
||||||
|
|
||||||
|
void InitKernel(const CNodePtr &kernel_node) override;
|
||||||
|
|
||||||
|
bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
|
||||||
|
const std::vector<AddressPtr> &outputs) override;
|
||||||
|
|
||||||
|
private:
|
||||||
|
std::vector<size_t> output_shape_;
|
||||||
|
std::vector<size_t> b_shape_;
|
||||||
|
size_t output_size_;
|
||||||
|
size_t aValues_size_;
|
||||||
|
};
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeBool)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeBool)
|
||||||
|
.AddOutputAttr(kNumberTypeBool),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, bool);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt8)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt8)
|
||||||
|
.AddOutputAttr(kNumberTypeUInt8),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, uint8_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt16)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt16)
|
||||||
|
.AddOutputAttr(kNumberTypeUInt16),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, uint16_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt32)
|
||||||
|
.AddOutputAttr(kNumberTypeUInt32),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, uint32_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt64)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt64)
|
||||||
|
.AddOutputAttr(kNumberTypeUInt64),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, uint64_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt8)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt8)
|
||||||
|
.AddOutputAttr(kNumberTypeInt8),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, int8_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt16)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt16)
|
||||||
|
.AddOutputAttr(kNumberTypeInt16),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, int16_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddOutputAttr(kNumberTypeInt32),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, int32_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddOutputAttr(kNumberTypeInt64),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, int64_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeFloat32)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeFloat32)
|
||||||
|
.AddOutputAttr(kNumberTypeFloat32),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, float);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeFloat64)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeFloat64)
|
||||||
|
.AddOutputAttr(kNumberTypeFloat64),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int32_t, double);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeBool)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeBool)
|
||||||
|
.AddOutputAttr(kNumberTypeBool),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, bool);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeUInt8)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt8)
|
||||||
|
.AddOutputAttr(kNumberTypeUInt8),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, uint8_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeUInt16)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt16)
|
||||||
|
.AddOutputAttr(kNumberTypeUInt16),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, uint16_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeUInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt32)
|
||||||
|
.AddOutputAttr(kNumberTypeUInt32),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, uint32_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeUInt64)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeUInt64)
|
||||||
|
.AddOutputAttr(kNumberTypeUInt64),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, uint64_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeInt8)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt8)
|
||||||
|
.AddOutputAttr(kNumberTypeInt8),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, int8_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeInt16)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt16)
|
||||||
|
.AddOutputAttr(kNumberTypeInt16),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, int16_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddOutputAttr(kNumberTypeInt32),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, int32_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddOutputAttr(kNumberTypeInt64),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, int64_t);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeFloat32)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeFloat32)
|
||||||
|
.AddOutputAttr(kNumberTypeFloat32),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, float);
|
||||||
|
|
||||||
|
MS_REG_CPU_KERNEL_T_S(SparseTensorDenseMatmul,
|
||||||
|
KernelAttr()
|
||||||
|
.AddInputAttr(kNumberTypeInt64)
|
||||||
|
.AddInputAttr(kNumberTypeFloat64)
|
||||||
|
.AddInputAttr(kNumberTypeInt32)
|
||||||
|
.AddInputAttr(kNumberTypeFloat64)
|
||||||
|
.AddOutputAttr(kNumberTypeFloat64),
|
||||||
|
SparseTensorDenseMatmulCPUKernel, int64_t, double);
|
||||||
|
|
||||||
|
} // namespace kernel
|
||||||
|
} // namespace mindspore
|
||||||
|
#endif // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_RMSPROP_CPU_KERNEL_H_
|
||||||
|
|
@ -15,8 +15,9 @@
|
||||||
"""
|
"""
|
||||||
Sparse related transformation.
|
Sparse related transformation.
|
||||||
"""
|
"""
|
||||||
from .sparse import SparseToDense
|
from .sparse import (SparseToDense, SparseTensorDenseMatmul)
|
||||||
|
|
||||||
__all__ = [
|
__all__ = [
|
||||||
"SparseToDense",
|
"SparseToDense",
|
||||||
|
"SparseTensorDenseMatmul",
|
||||||
]
|
]
|
||||||
|
|
|
||||||
|
|
@ -1,4 +1,4 @@
|
||||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
# Copyright 2020-2021 Huawei Technologies Co., Ltd
|
||||||
#
|
#
|
||||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
# you may not use this file except in compliance with the License.
|
# you may not use this file except in compliance with the License.
|
||||||
|
|
@ -52,3 +52,50 @@ class SparseToDense(Cell):
|
||||||
return self.sparse_to_dense(sparse_tensor.indices,
|
return self.sparse_to_dense(sparse_tensor.indices,
|
||||||
sparse_tensor.values,
|
sparse_tensor.values,
|
||||||
sparse_tensor.dense_shape)
|
sparse_tensor.dense_shape)
|
||||||
|
|
||||||
|
class SparseTensorDenseMatmul(Cell):
|
||||||
|
"""
|
||||||
|
Multiply SparseTensor(of rank 2) "A" by dense tensor.
|
||||||
|
The shape of sparse tensor is :math:`(N, C)`, and the shape of dense tensor is :math:`(C, M)`, then the shape of
|
||||||
|
output tensor is :math:`(N, M)`.The output data type is the same as "values".
|
||||||
|
|
||||||
|
Args:
|
||||||
|
- *adjoint_st** (Bool) - If true, SparseTensor is transposed before multiplication. Default: False.
|
||||||
|
- *adjoint_dt** (Bool) - If true, DenseTensor is transposed before multiplication. Default: False.
|
||||||
|
|
||||||
|
Inputs:
|
||||||
|
- **indices** (Tensor) - The indices of sparse representation, support int32/int64.
|
||||||
|
- **values** (Tensor) - Values corresponding to each row of indices.
|
||||||
|
- **dense_shape** (tuple) - An int tuple which specifies the shape of dense tensor. The dense_shape is :
|
||||||
|
math:`(N, C)`. If `adjoint_st` is True, its shape must be :math:`(N, C)` after transpose.
|
||||||
|
- **dense** (Tensor) - Dense Matrix. The shape of the tensor is :math:`(C, M)`. If
|
||||||
|
`adjoint_dt` is True, its shape must be :math:`(C, M)` after transpose.
|
||||||
|
|
||||||
|
Returns:
|
||||||
|
Tensor, the shape of tensor is :math:`(N, M)`.The output data type is the same as "values".
|
||||||
|
|
||||||
|
Examples:
|
||||||
|
>>> class NetSparseDenseMatmul(nn.Cell):
|
||||||
|
... def __init__(self):
|
||||||
|
... super(NetSparseDenseMatmul, self).__init__()
|
||||||
|
... self.matmul = nn.SparseTensorDenseMatmul()
|
||||||
|
...
|
||||||
|
... def construct(self, indices, values, dens_shape, dt):
|
||||||
|
... return self.matmul(indices, values, dens_shape, dt)
|
||||||
|
...
|
||||||
|
>>> indices = Tensor([[0, 1], [1, 2]], dtype=ms.int32)
|
||||||
|
>>> values = Tensor([1, 2], dtype=ms.float32)
|
||||||
|
>>> dense_shape = (3, 4)
|
||||||
|
>>> dsMatrix = Tensor([[1, 1], [2, 2], [3, 3], [4, 4]], dtype=ms.float32)
|
||||||
|
>>> test_SparseDenseMatmul = NetSparseDenseMatmul()
|
||||||
|
>>> out = test_SparseDenseMatmul(indices, values, dens_shape, dsMatrix)
|
||||||
|
"""
|
||||||
|
def __init__(self, adjoint_st=False, adjoint_dt=False):
|
||||||
|
"""Initialize SparseTensorDenseMatmul"""
|
||||||
|
super(SparseTensorDenseMatmul, self).__init__()
|
||||||
|
self.adjst = adjoint_st
|
||||||
|
self.adjdt = adjoint_dt
|
||||||
|
self.matmul = P.SparseTensorDenseMatmul(adjoint_st=self.adjst, adjoint_dt=self.adjdt)
|
||||||
|
|
||||||
|
def construct(self, indices, values, dense_shape, dense):
|
||||||
|
return self.matmul(indices, values, dense_shape, dense)
|
||||||
|
|
|
||||||
|
|
@ -94,7 +94,7 @@ from ._thor_ops import (CusBatchMatMul, CusCholeskyTrsm, CusFusedAbsMax1, CusImg
|
||||||
CusMatMulCubeDenseRight,
|
CusMatMulCubeDenseRight,
|
||||||
CusMatMulCubeFraczLeftCast, Im2Col, UpdateThorGradient, Cholesky, CholeskyTrsm, DetTriangle,
|
CusMatMulCubeFraczLeftCast, Im2Col, UpdateThorGradient, Cholesky, CholeskyTrsm, DetTriangle,
|
||||||
ProdForceSeA)
|
ProdForceSeA)
|
||||||
from .sparse_ops import SparseToDense
|
from .sparse_ops import (SparseToDense, SparseTensorDenseMatmul)
|
||||||
from ._embedding_cache_ops import (CacheSwapHashmap, SearchCacheIdx, CacheSwapTable, UpdateCache, MapCacheIdx,
|
from ._embedding_cache_ops import (CacheSwapHashmap, SearchCacheIdx, CacheSwapTable, UpdateCache, MapCacheIdx,
|
||||||
SubAndFilter,
|
SubAndFilter,
|
||||||
MapUniform, DynamicAssign, PadAndShift)
|
MapUniform, DynamicAssign, PadAndShift)
|
||||||
|
|
@ -428,6 +428,7 @@ __all__ = [
|
||||||
"Pull",
|
"Pull",
|
||||||
"ReLUV2",
|
"ReLUV2",
|
||||||
"SparseToDense",
|
"SparseToDense",
|
||||||
|
"SparseTensorDenseMatmul",
|
||||||
"MatrixInverse",
|
"MatrixInverse",
|
||||||
"Range",
|
"Range",
|
||||||
"IndexAdd",
|
"IndexAdd",
|
||||||
|
|
|
||||||
|
|
@ -1,6 +1,6 @@
|
||||||
# coding: utf-8
|
# coding: utf-8
|
||||||
|
|
||||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
# Copyright 2020-2021 Huawei Technologies Co., Ltd
|
||||||
#
|
#
|
||||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
# you may not use this file except in compliance with the License.
|
# you may not use this file except in compliance with the License.
|
||||||
|
|
@ -53,3 +53,72 @@ class SparseToDense(PrimitiveWithInfer):
|
||||||
'dtype': values['dtype'],
|
'dtype': values['dtype'],
|
||||||
'value': None}
|
'value': None}
|
||||||
return out
|
return out
|
||||||
|
|
||||||
|
class SparseTensorDenseMatmul(PrimitiveWithInfer):
|
||||||
|
"""
|
||||||
|
Multiply SparseTensor(of rank 2) "A" by dense tensor.
|
||||||
|
The shape of sparse tensor is :math:`(N, C)`, and the shape of dense tensor is :math:`(C, M)`, then the shape of
|
||||||
|
output tensor is :math:`(N, M)`.The output data type is the same as "values".
|
||||||
|
tensors.
|
||||||
|
|
||||||
|
Args:
|
||||||
|
- *adjoint_st** (Bool) - If true, SparseTensor is transposed before multiplication. Default: False.
|
||||||
|
- *adjoint_dt** (Bool) - If true, DenseTensor is transposed before multiplication. Default: False.
|
||||||
|
|
||||||
|
Inputs:
|
||||||
|
- **indices** (Tensor) - The indices of sparse representation, support int32/int64.
|
||||||
|
- **values** (Tensor) - Values corresponding to each row of indices.
|
||||||
|
- **dense_shape** (tuple) - An int tuple which specifies the shape of dense tensor. The dense_shape is :
|
||||||
|
math:`(N, C)`. If `adjoint_st` is True, its shape must be :math:`(N, C)` after transpose.
|
||||||
|
- **dense** (Tensor) - Dense Matrix. The shape of the tensor is :math:`(C, M)`. If
|
||||||
|
`adjoint_dt` is True, its shape must be :math:`(C, M)` after transpose.
|
||||||
|
|
||||||
|
Outputs:
|
||||||
|
Tensor, the shape of tensor is :math:`(N, M)`. The output data type is the same as "values".
|
||||||
|
|
||||||
|
Raises:
|
||||||
|
TypeError: If `indices` is neither int32 nor int64.
|
||||||
|
TypeError: If 'values' is not boot, uint8-64, int8-64, float16-64.
|
||||||
|
TypeError: If 'dense' is not boot, uint8-64, int8-64, float16-64.
|
||||||
|
ValueError: If length of shape of `SparseTensor` or `DenseTensor` is not equal to 2
|
||||||
|
|
||||||
|
Supported Platforms:
|
||||||
|
``CPU``
|
||||||
|
|
||||||
|
Examples:
|
||||||
|
>>> indices = Tensor([[0, 1], [1, 2]], dtype=ms.int32)
|
||||||
|
>>> values = Tensor([1, 2], dtype=ms.float32)
|
||||||
|
>>> dense_shape = (3, 4)
|
||||||
|
>>> dsMatrix = Tensor([[1,1], [2,2], [3,3 ], [4, 4]], dtype=ms.float32)
|
||||||
|
>>> out = ops.SparseTensorDenseMatmul(indices, values, dense_shape, dsMatrix)
|
||||||
|
"""
|
||||||
|
@prim_attr_register
|
||||||
|
def __init__(self, adjoint_st=False, adjoint_dt=False):
|
||||||
|
"""Initialize SparseTensorDenseMatmul"""
|
||||||
|
self.adjoint_st = adjoint_st
|
||||||
|
self.adjoint_dt = adjoint_dt
|
||||||
|
self.init_prim_io_names(inputs=['indices', 'values', 'dense_shape', 'dense'],
|
||||||
|
outputs=['output'])
|
||||||
|
self.add_prim_attr('adjoint_st', self.adjoint_st)
|
||||||
|
self.add_prim_attr('adjoint_dt', self.adjoint_dt)
|
||||||
|
validator.check_value_type("adjoint_st", adjoint_st, [bool], self.name)
|
||||||
|
validator.check_value_type("adjoint_dt", adjoint_dt, [bool], self.name)
|
||||||
|
|
||||||
|
def __infer__(self, indices, values, dense_shape, dense):
|
||||||
|
validator.check_tensor_dtype_valid('indices', indices['dtype'], [mstype.int32, mstype.int64], self.name)
|
||||||
|
valid_types = mstype.number_type + (mstype.bool_,)
|
||||||
|
args = {'values': values['dtype'], 'dense': dense['dtype']}
|
||||||
|
validator.check_tensors_dtypes_same_and_valid(args, valid_types, self.name)
|
||||||
|
a_shape = dense_shape['value']
|
||||||
|
b_shape = dense['shape']
|
||||||
|
if len(a_shape) != 2 or len(b_shape) != 2:
|
||||||
|
raise ValueError('SparseTensorDenseMatmul SparseTensor, DenseTensor should have the same dimension size '
|
||||||
|
+ f'and equal to 2, while SparseTensor size is ({len(a_shape)}) and DenseTensor size is '
|
||||||
|
+ f'({len(b_shape)}).')
|
||||||
|
out_shape = []
|
||||||
|
out_shape.append(a_shape[0])
|
||||||
|
out_shape.append(b_shape[1])
|
||||||
|
out = {'shape': tuple(out_shape),
|
||||||
|
'dtype': values['dtype'],
|
||||||
|
'value': None}
|
||||||
|
return out
|
||||||
|
|
|
||||||
|
|
@ -0,0 +1,53 @@
|
||||||
|
# Copyright 2020-2021 Huawei Technologies Co., Ltd
|
||||||
|
#
|
||||||
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
# you may not use this file except in compliance with the License.
|
||||||
|
# You may obtain a copy of the License at
|
||||||
|
#
|
||||||
|
# http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
#
|
||||||
|
# Unless required by applicable law or agreed to in writing, software
|
||||||
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
# See the License for the specific language governing permissions and
|
||||||
|
# limitations under the License.
|
||||||
|
# ============================================================================
|
||||||
|
|
||||||
|
import numpy as np
|
||||||
|
import mindspore as ms
|
||||||
|
import mindspore.context as context
|
||||||
|
import mindspore.nn as nn
|
||||||
|
from mindspore import Tensor
|
||||||
|
from mindspore import SparseTensor
|
||||||
|
|
||||||
|
context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
|
||||||
|
|
||||||
|
class NetSparseDenseMatmul(nn.Cell):
|
||||||
|
def __init__(self):
|
||||||
|
super(NetSparseDenseMatmul, self).__init__()
|
||||||
|
self.matmul = nn.SparseTensorDenseMatmul()
|
||||||
|
|
||||||
|
def construct(self, indices, values, dens_shape, dt):
|
||||||
|
return self.matmul(indices, values, dens_shape, dt)
|
||||||
|
|
||||||
|
class NetSparseTensor(nn.Cell):
|
||||||
|
def __init__(self, dense_shape):
|
||||||
|
super(NetSparseTensor, self).__init__()
|
||||||
|
self.dense_shape = dense_shape
|
||||||
|
def construct(self, indices, values):
|
||||||
|
x = SparseTensor(indices, values, self.dense_shape)
|
||||||
|
return x.values, x.indices, x.dense_shape
|
||||||
|
|
||||||
|
def test_sparse_tensor_dense_matmul():
|
||||||
|
indices = Tensor([[0, 1], [1, 1]])
|
||||||
|
values = Tensor([5, 5], dtype=ms.float32)
|
||||||
|
dens_shape = (3, 3)
|
||||||
|
spMatrix = np.array([[5, 0, 0], [0, 5, 0], [0, 0, 5]], dtype=np.float32)
|
||||||
|
dsMatrix = np.array([[1, 1, 1], [1, 1, 1], [1, 1, 1]], dtype=np.float32)
|
||||||
|
test_SparseDenseMatmul = NetSparseDenseMatmul()
|
||||||
|
|
||||||
|
out_ms = test_SparseDenseMatmul(indices, values, dens_shape, Tensor(dsMatrix))
|
||||||
|
out_np = np.matmul(spMatrix, dsMatrix)
|
||||||
|
error = np.ones(shape=dsMatrix.shape) * 10e-6
|
||||||
|
diff = out_ms.asnumpy() - out_np
|
||||||
|
assert np.all(diff < error)
|
||||||
Loading…
Reference in New Issue