m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add jit kernel hsum, hmax and softmax refer code

Browse Source 
test=develop
revert-15296-async_double_buffered_py_reader
tensor-tang 6 years ago
parent 67e4450c34
commit 8117725852
8 changed files with 269 additions and 121 deletions
Show Stats Download Patch File Download Diff File

101
paddle/fluid/operators/jit/benchmark.cc
Unescape View File

@ -158,7 +158,7 @@ void BenchAllImpls(const typename KernelTuples::attr_type& attr, Args... args) {
using Tensor = paddle::framework::Tensor;
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
template <jit::KernelType KT, typename T, typename PlaceType>
void BenchXYZNKernel() {
for (int d : TestSizes()) {
Tensor x, y, z;
@ -175,7 +175,7 @@ void BenchXYZNKernel() {
}
}
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
template <jit::KernelType KT, typename T, typename PlaceType>
void BenchAXYNKernel() {
for (int d : TestSizes()) {
const T a = static_cast<T>(3);
@ -190,7 +190,17 @@ void BenchAXYNKernel() {
}
}
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
template <jit::KernelType KT, typename T, typename PlaceType>
void BenchXRNKernel() {
for (int d : TestSizes()) {
Tensor x;
RandomVec<T>(d, x.mutable_data<T>({d}, PlaceType()));
T res;
BenchAllImpls<KT, jit::XRNTuples<T>, PlaceType>(d, x.data<T>(), &res, d);
}
}
template <jit::KernelType KT, typename T, typename PlaceType>
void BenchXYNKernel() {
for (int d : TestSizes()) {
Tensor x, y;
@ -203,7 +213,7 @@ void BenchXYNKernel() {
}
}
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
template <jit::KernelType KT, typename T, typename PlaceType>
void BenchLSTMKernel() {
for (bool use_peephole : {true, false}) {
for (int d : TestSizes()) {
@ -240,7 +250,7 @@ void BenchLSTMKernel() {
}
}
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
template <jit::KernelType KT, typename T, typename PlaceType>
void BenchGRUKernel() {
for (int d : TestSizes()) {
const jit::gru_attr_t attr(d, jit::kVSigmoid, jit::kVTanh);
@ -262,7 +272,7 @@ void BenchGRUKernel() {
}
}
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
template <jit::KernelType KT, typename T, typename PlaceType>
void BenchSeqPoolKernel() {
std::vector<jit::SeqPoolType> pool_types = {
jit::SeqPoolType::kSum, jit::SeqPoolType::kAvg, jit::SeqPoolType::kSqrt};
@ -284,7 +294,7 @@ void BenchSeqPoolKernel() {
}
}
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
template <jit::KernelType KT, typename T, typename PlaceType>
void BenchMatMulKernel() {
for (int m : {1, 2, 3, 4}) {
for (int n : TestSizes()) {
@ -305,57 +315,64 @@ void BenchMatMulKernel() {
}
}
template <jit::KernelType KT, typename T, typename PlaceType>
void BenchSoftmaxKernel() {
for (int bs : {1, 2, 10}) {
for (int n : TestSizes()) {
Tensor x, y;
x.Resize({bs, n});
y.Resize({bs, n});
RandomVec<T>(bs * n, x.mutable_data<T>(PlaceType()), -2.f, 2.f);
const T* x_data = x.data<T>();
T* y_data = y.mutable_data<T>(PlaceType());
BenchAllImpls<KT, jit::SoftmaxTuples<T>, PlaceType>(n, x_data, y_data, n,
bs);
}
}
}
using T = float;
using PlaceType = paddle::platform::CPUPlace;
using CPUPlace = paddle::platform::CPUPlace;
// xyzn
BENCH_FP32_CPU(kVMul) { BenchXYZNKernel<jit::kVMul, T, PlaceType>(); }
BENCH_FP32_CPU(kVAdd) { BenchXYZNKernel<jit::kVAdd, T, PlaceType>(); }
BENCH_FP32_CPU(kVAddRelu) { BenchXYZNKernel<jit::kVAddRelu, T, PlaceType>(); }
BENCH_FP32_CPU(kVSub) { BenchXYZNKernel<jit::kVSub, T, PlaceType>(); }
BENCH_FP32_CPU(kVMul) { BenchXYZNKernel<jit::kVMul, T, CPUPlace>(); }
BENCH_FP32_CPU(kVAdd) { BenchXYZNKernel<jit::kVAdd, T, CPUPlace>(); }
BENCH_FP32_CPU(kVAddRelu) { BenchXYZNKernel<jit::kVAddRelu, T, CPUPlace>(); }
BENCH_FP32_CPU(kVSub) { BenchXYZNKernel<jit::kVSub, T, CPUPlace>(); }
// axyn
BENCH_FP32_CPU(kVScal) { BenchAXYNKernel<jit::kVScal, T, PlaceType>(); }
BENCH_FP32_CPU(kVScal) { BenchAXYNKernel<jit::kVScal, T, CPUPlace>(); }
BENCH_FP32_CPU(kVAddBias) { BenchAXYNKernel<jit::kVAddBias, T, CPUPlace>(); }
BENCH_FP32_CPU(kVAddBias) { BenchAXYNKernel<jit::kVAddBias, T, PlaceType>(); }
// xrn
BENCH_FP32_CPU(kHSum) { BenchXRNKernel<jit::kHSum, T, CPUPlace>(); }
BENCH_FP32_CPU(kHMax) { BenchXRNKernel<jit::kHMax, T, CPUPlace>(); }
// xyn
BENCH_FP32_CPU(kVRelu) { BenchXYNKernel<jit::kVRelu, T, PlaceType>(); }
BENCH_FP32_CPU(kVIdentity) { BenchXYNKernel<jit::kVIdentity, T, PlaceType>(); }
BENCH_FP32_CPU(kVSquare) { BenchXYNKernel<jit::kVSquare, T, PlaceType>(); }
BENCH_FP32_CPU(kVExp) { BenchXYNKernel<jit::kVExp, T, PlaceType>(); }
BENCH_FP32_CPU(kVSigmoid) { BenchXYNKernel<jit::kVSigmoid, T, PlaceType>(); }
BENCH_FP32_CPU(kVTanh) { BenchXYNKernel<jit::kVTanh, T, PlaceType>(); }
BENCH_FP32_CPU(kVRelu) { BenchXYNKernel<jit::kVRelu, T, CPUPlace>(); }
BENCH_FP32_CPU(kVIdentity) { BenchXYNKernel<jit::kVIdentity, T, CPUPlace>(); }
BENCH_FP32_CPU(kVSquare) { BenchXYNKernel<jit::kVSquare, T, CPUPlace>(); }
BENCH_FP32_CPU(kVExp) { BenchXYNKernel<jit::kVExp, T, CPUPlace>(); }
BENCH_FP32_CPU(kVSigmoid) { BenchXYNKernel<jit::kVSigmoid, T, CPUPlace>(); }
BENCH_FP32_CPU(kVTanh) { BenchXYNKernel<jit::kVTanh, T, CPUPlace>(); }
// lstm and peephole
BENCH_FP32_CPU(kLSTMCtHt) { BenchLSTMKernel<jit::kLSTMCtHt, T, PlaceType>(); }
BENCH_FP32_CPU(kLSTMC1H1) { BenchLSTMKernel<jit::kLSTMC1H1, T, PlaceType>(); }
BENCH_FP32_CPU(kLSTMCtHt) { BenchLSTMKernel<jit::kLSTMCtHt, T, CPUPlace>(); }
BENCH_FP32_CPU(kLSTMC1H1) { BenchLSTMKernel<jit::kLSTMC1H1, T, CPUPlace>(); }
// gru functions
BENCH_FP32_CPU(kGRUH1) { BenchGRUKernel<jit::kGRUH1, T, PlaceType>(); }
BENCH_FP32_CPU(kGRUHtPart1) {
BenchGRUKernel<jit::kGRUHtPart1, T, PlaceType>();
}
BENCH_FP32_CPU(kGRUHtPart2) {
BenchGRUKernel<jit::kGRUHtPart2, T, PlaceType>();
}
BENCH_FP32_CPU(kGRUH1) { BenchGRUKernel<jit::kGRUH1, T, CPUPlace>(); }
BENCH_FP32_CPU(kGRUHtPart1) { BenchGRUKernel<jit::kGRUHtPart1, T, CPUPlace>(); }
BENCH_FP32_CPU(kGRUHtPart2) { BenchGRUKernel<jit::kGRUHtPart2, T, CPUPlace>(); }
// seq pool function
BENCH_FP32_CPU(kSeqPool) { BenchSeqPoolKernel<jit::kSeqPool, T, PlaceType>(); }
BENCH_FP32_CPU(kSeqPool) { BenchSeqPoolKernel<jit::kSeqPool, T, CPUPlace>(); }
// matmul
BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel<jit::kMatMul, T, PlaceType>(); }
BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel<jit::kMatMul, T, CPUPlace>(); }
// softmax
BENCH_FP32_CPU(kSoftmax) { BenchSoftmaxKernel<jit::kSoftmax, T, CPUPlace>(); }
// Benchmark all jit kernels including jitcode, mkl and refer.
// To use this tool, run command: ./benchmark [options...]

3
paddle/fluid/operators/jit/helper.cc
Unescape View File

@ -49,6 +49,9 @@ const char* to_string(KernelType kt) {
ONE_CASE(kNCHW16CMulNC);
ONE_CASE(kSeqPool);
ONE_CASE(kMatMul);
ONE_CASE(kHMax);
ONE_CASE(kHSum);
ONE_CASE(kSoftmax);
default:
PADDLE_THROW("Not support type: %d, or forget to add it.", kt);
return "NOT JITKernel";

15
paddle/fluid/operators/jit/kernel_base.h
Unescape View File

@ -20,6 +20,7 @@ namespace paddle {
namespace operators {
namespace jit {
// TODO(TJ): reorder by alphabet
typedef enum {
kNone = 0,
kVMul = 1,
@ -44,6 +45,9 @@ typedef enum {
kNCHW16CMulNC,
kSeqPool,
kMatMul,
kHSum, // horizontal max
kHMax, // horizontal sum
kSoftmax,
} KernelType;
typedef enum {
@ -70,6 +74,10 @@ struct XYNTuples {
typedef void (*func_type)(const T*, T*, int);
};
// x, return and int
template <typename T>
struct XRNTuples : public XYNTuples<T> {};
typedef struct {
void* gates; // gates: x_ch, x_ih, x_fh, x_oh
const void* ct_1;
@ -159,6 +167,13 @@ struct LayerNormTuples {
const float, int);
};
template <typename T>
struct SoftmaxTuples {
typedef T data_type;
typedef int attr_type;
typedef void (*func_type)(const T*, T*, int, int);
};
// nChw16c = nChw16c .* NC
template <typename T>
struct NCHW16CMulNCTuples {

3
paddle/fluid/operators/jit/refer/CMakeLists.txt
Unescape View File

@ -29,3 +29,6 @@ USE_JITKERNEL_REFER(kNCHW16CMulNC)
USE_JITKERNEL_REFER(kSeqPool)
USE_JITKERNEL_REFER(kMatMul)
USE_JITKERNEL_REFER(kVSquare)
USE_JITKERNEL_REFER(kHSum)
USE_JITKERNEL_REFER(kHMax)
USE_JITKERNEL_REFER(kSoftmax)

5
paddle/fluid/operators/jit/refer/refer.cc
Unescape View File

@ -52,4 +52,9 @@ REGISTER_REFER_KERNEL(kSeqPool, SeqPool);
REGISTER_REFER_KERNEL(kMatMul, MatMul);
REGISTER_REFER_KERNEL(kHMax, HMax);
REGISTER_REFER_KERNEL(kHSum, HSum);
REGISTER_REFER_KERNEL(kSoftmax, Softmax);
#undef REGISTER_REFER_KERNEL

39
paddle/fluid/operators/jit/refer/refer.h
Unescape View File

@ -378,6 +378,40 @@ void MatMul(const T* A, const T* B, T* C, int M, int N, int K) {
}
}
template <typename T>
void HMax(const T* x, T* res, int n) {
res[0] = x[0];
for (int i = 1; i < n; ++i) {
res[0] = res[0] < x[i] ? x[i] : res[0];
}
}
template <typename T>
void HSum(const T* x, T* res, int n) {
res[0] = x[0];
for (int i = 1; i < n; ++i) {
res[0] += x[i];
}
}
// y = e^(x - max(x))
// y = y / sum(y)
template <typename T>
void Softmax(const T* x, T* y, int n, int bs = 1) {
for (int i = 0; i < bs; ++i) {
T scalar;
HMax(x, &scalar, n);
scalar = static_cast<T>(0) - scalar;
VAddBias(&scalar, x, y, n); // x - max
VExp(y, y, n);
HSum(y, &scalar, n);
scalar = static_cast<T>(1) / scalar;
VScal(&scalar, y, y, n);
x += n;
y += n;
}
}
#define DECLARE_REFER_KERNEL(name, tuples) \
template <typename T> \
class name##Kernel : public ReferKernel<tuples<T>> { \
@ -421,6 +455,11 @@ DECLARE_REFER_KERNEL(SeqPool, SeqPoolTuples);
DECLARE_REFER_KERNEL(MatMul, MatMulTuples);
DECLARE_REFER_KERNEL(HMax, XRNTuples);
DECLARE_REFER_KERNEL(HSum, XRNTuples);
DECLARE_REFER_KERNEL(Softmax, SoftmaxTuples);
#undef DECLARE_REFER_KERNEL
} // namespace refer

222
paddle/fluid/operators/jit/test.cc
View File

File diff suppressed because it is too large Load Diff

2
paddle/fluid/platform/dynload/mklml.h
Unescape View File

@ -70,6 +70,8 @@ extern void* mklml_dso_handle;
__macro(cblas_ddot); \
__macro(cblas_sasum); \
__macro(cblas_dasum); \
__macro(cblas_isamax); \
__macro(cblas_idamax); \
__macro(cblas_sscal); \
__macro(cblas_dscal); \
__macro(vsAdd); \

Write Preview
Loading…
Cancel
Save