1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
|
// See docs in ../ops/math_ops.cc.
#define EIGEN_USE_THREADS
#include "tensorflow/core/kernels/matmul_op.h"
#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/kernels/fill_functor.h"
#if GOOGLE_CUDA
#include "tensorflow/core/common_runtime/gpu_device_context.h"
#include "tensorflow/stream_executor/stream.h"
#endif // GOOGLE_CUDA
namespace tensorflow {
#if GOOGLE_CUDA
namespace {
template <typename T>
perftools::gputools::DeviceMemory<T> AsDeviceMemory(const T* cuda_memory) {
perftools::gputools::DeviceMemoryBase wrapped(const_cast<T*>(cuda_memory));
perftools::gputools::DeviceMemory<T> typed(wrapped);
return typed;
}
} // namespace
#endif // GOOGLE_CUDA
typedef Eigen::ThreadPoolDevice CPUDevice;
typedef Eigen::GpuDevice GPUDevice;
template <typename Device, typename T, bool USE_CUBLAS>
struct LaunchMatMul;
// On CPUs, we ignore USE_CUBLAS
template <typename T>
struct LaunchMatMulCPU {
static void launch(
OpKernelContext* ctx, OpKernel* kernel, const Tensor& a, const Tensor& b,
const Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1>& dim_pair,
Tensor* out) {
functor::MatMulFunctor<CPUDevice, T>()(ctx->eigen_device<CPUDevice>(),
out->matrix<T>(), a.matrix<T>(),
b.matrix<T>(), dim_pair);
}
};
template <typename T, bool USE_CUBLAS>
struct LaunchMatMul<CPUDevice, T, USE_CUBLAS> : public LaunchMatMulCPU<T> {};
#if GOOGLE_CUDA
template <typename T>
struct LaunchMatMul<GPUDevice, T, true /* USE_CUBLAS */> {
static void launch(
OpKernelContext* ctx, OpKernel* kernel, const Tensor& a, const Tensor& b,
const Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1>& dim_pair,
Tensor* out) {
perftools::gputools::blas::Transpose trans[] = {
perftools::gputools::blas::Transpose::kNoTranspose,
perftools::gputools::blas::Transpose::kTranspose};
const uint64 m = a.dim_size(1 - dim_pair[0].first);
const uint64 k = a.dim_size(dim_pair[0].first);
const uint64 n = b.dim_size(1 - dim_pair[0].second);
bool transpose_a = dim_pair[0].first == 0;
bool transpose_b = dim_pair[0].second == 1;
auto blas_transpose_a = trans[transpose_a];
auto blas_transpose_b = trans[transpose_b];
auto* stream = ctx->op_device_context<GPUDeviceContext>()->stream();
OP_REQUIRES(ctx, stream, errors::Internal("No GPU stream available."));
auto a_ptr = AsDeviceMemory(a.template flat<T>().data());
auto b_ptr = AsDeviceMemory(b.template flat<T>().data());
auto c_ptr = AsDeviceMemory(out->template flat<T>().data());
// Cublas does
// C = A x B
// where A, B and C are assumed to be in column major.
// We want the output to be in row-major, so we can compute
// C' = B' x A' (' stands for transpose)
bool blas_launch_status =
stream->ThenBlasGemm(blas_transpose_b, blas_transpose_a, n, m, k, 1.0f,
b_ptr, transpose_b ? k : n, a_ptr,
transpose_a ? m : k, 0.0f, &c_ptr, n)
.ok();
if (!blas_launch_status) {
ctx->SetStatus(errors::Internal(
"Blas SGEMM launch failed : a.shape=(", a.dim_size(0), ", ",
a.dim_size(1), "), b.shape=(", b.dim_size(0), ", ", b.dim_size(1),
"), m=", m, ", n=", n, ", k=", k));
}
}
};
template <typename T>
struct LaunchMatMul<GPUDevice, T, false /* USE_CUBLAS */> {
static void launch(
OpKernelContext* ctx, OpKernel* kernel, const Tensor& a, const Tensor& b,
const Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1>& dim_pair,
Tensor* out) {
functor::MatMulFunctor<GPUDevice, T>()(ctx->eigen_device<GPUDevice>(),
out->matrix<T>(), a.matrix<T>(),
b.matrix<T>(), dim_pair);
}
};
#endif // GOOGLE_CUDA
template <typename Device, typename T, bool USE_CUBLAS>
class MatMulOp : public OpKernel {
public:
explicit MatMulOp(OpKernelConstruction* ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("transpose_a", &transpose_a_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("transpose_b", &transpose_b_));
}
void Compute(OpKernelContext* ctx) override {
const Tensor& a = ctx->input(0);
const Tensor& b = ctx->input(1);
// Check that the dimensions of the two matrices are valid.
OP_REQUIRES(ctx, TensorShapeUtils::IsMatrix(a.shape()),
errors::InvalidArgument("In[0] is not a matrix"));
OP_REQUIRES(ctx, TensorShapeUtils::IsMatrix(b.shape()),
errors::InvalidArgument("In[1] is not a matrix"));
Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair;
dim_pair[0].first = transpose_a_ ? 0 : 1;
dim_pair[0].second = transpose_b_ ? 1 : 0;
OP_REQUIRES(ctx,
a.dim_size(dim_pair[0].first) == b.dim_size(dim_pair[0].second),
errors::InvalidArgument("Matrix size-compatible: In[0]: ",
a.shape().DebugString(), ", In[1]: ",
b.shape().DebugString()));
int a_dim_remaining = 1 - dim_pair[0].first;
int b_dim_remaining = 1 - dim_pair[0].second;
TensorShape out_shape(
{a.dim_size(a_dim_remaining), b.dim_size(b_dim_remaining)});
Tensor* out = nullptr;
OP_REQUIRES_OK(ctx, ctx->allocate_output(0, out_shape, &out));
if (out->NumElements() == 0) {
// If a has shape [0, x] or b has shape [x, 0], the output shape
// is a 0-element matrix, so there is nothing to do.
return;
}
if (a.NumElements() == 0 || b.NumElements() == 0) {
// If a has shape [x, 0] and b has shape [0, y], the
// output shape is [x, y] where x and y are non-zero, so we fill
// the output with zeros.
functor::SetZeroFunctor<Device, T> f;
f(ctx->eigen_device<Device>(), out->flat<T>());
return;
}
LaunchMatMul<Device, T, USE_CUBLAS>::launch(ctx, this, a, b, dim_pair, out);
}
private:
bool transpose_a_;
bool transpose_b_;
};
namespace functor {
// Partial specialization MatMulFunctor<Device=CPUDevice, T>.
template <typename T>
struct MatMulFunctor<CPUDevice, T> {
void operator()(
const CPUDevice& d, typename MatMulTypes<T>::out_type out,
typename MatMulTypes<T>::in_type in0,
typename MatMulTypes<T>::in_type in1,
const Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1>& dim_pair) {
MatMul<CPUDevice>(d, out, in0, in1, dim_pair);
}
};
} // end namespace functor
#define REGISTER_CPU(T) \
REGISTER_KERNEL_BUILDER( \
Name("MatMul").Device(DEVICE_CPU).TypeConstraint<T>("T"), \
MatMulOp<CPUDevice, T, false /* cublas, ignored for CPU */>); \
REGISTER_KERNEL_BUILDER( \
Name("MatMul").Device(DEVICE_CPU).TypeConstraint<T>("T").Label("eigen"), \
MatMulOp<CPUDevice, T, false /* cublas, ignored for CPU */>)
#define REGISTER_GPU(T) \
REGISTER_KERNEL_BUILDER( \
Name("MatMul").Device(DEVICE_GPU).TypeConstraint<T>("T"), \
MatMulOp<GPUDevice, T, true /* cublas, true by default */>); \
REGISTER_KERNEL_BUILDER(Name("MatMul") \
.Device(DEVICE_GPU) \
.TypeConstraint<T>("T") \
.Label("cublas"), \
MatMulOp<GPUDevice, T, true /* cublas */>); \
REGISTER_KERNEL_BUILDER( \
Name("MatMul").Device(DEVICE_GPU).TypeConstraint<T>("T").Label("eigen"), \
MatMulOp<GPUDevice, T, false /* cublas */>)
REGISTER_CPU(float);
REGISTER_CPU(double);
REGISTER_CPU(int32);
REGISTER_CPU(complex64);
#if GOOGLE_CUDA
REGISTER_GPU(float);
// REGISTER_GPU(double);
#endif // GOOGLE_CUDA
} // namespace tensorflow
|
