Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 2513b2cc 编写于 作者: T tensor-tang
浏览文件
操作

fix bug vtanh

上级 cf8c8e72
显示空白变更内容
内联 并排
Showing with 167 addition and 129 deletion
paddle/fluid/operators/math/jit_kernel.h
浏览文件 @ 2513b2cc
......@@ -29,7 +29,6 @@ namespace jitkernel {
#define SIGMOID_THRESHOLD_MIN -40.0
#define SIGMOID_THRESHOLD_MAX 13.0
#define EXP_MAX_INPUT 40.0
#define AVX_FLOAT_BLOCK 8
#define AVX2_FLOAT_BLOCK 8
#define AVX512_FLOAT_BLOCK 16
......@@ -40,8 +39,9 @@ class Kernel {
public:
Kernel() = default;
virtual ~Kernel() = default;
private:
int num_{0};
int end_{0};
int rest_{0};
DISABLE_COPY_AND_ASSIGN(Kernel);
};
......@@ -95,13 +95,13 @@ class VExpKernel : public Kernel {
template <typename T>
class VSigmoidKernel : public Kernel {
public:
virtual void Compute(const int n, const T *x, T *y) const = 0;
virtual void Compute(const T *x, T *y) const = 0;
};
template <typename T>
class VTanhKernel : public Kernel {
public:
virtual void Compute(const int n, const T *x, T *y) const = 0;
virtual void Compute(const T *x, T *y) const = 0;
};
template <typename T>
......
paddle/fluid/operators/math/jit_kernel_exp.cc
浏览文件 @ 2513b2cc
......@@ -113,17 +113,18 @@ template <typename T, jit::cpu_isa_t isa, jit_block>
class VSigmoidKernelImpl : public VSigmoidKernel<T> {
public:
explicit VSigmoidKernelImpl(int d) : VSigmoidKernel<T>() {
this->num_ = d;
vexp_ = KernelPool::Instance().template Get<VExpKernel<T>>(d);
}
void Compute(const int n, const T* x, T* y) const override {
void Compute(const T* x, T* y) const override {
const T min = SIGMOID_THRESHOLD_MIN;
const T max = SIGMOID_THRESHOLD_MAX;
for (int i = 0; i < n; ++i) {
for (int i = 0; i < this->num_; ++i) {
y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
y[i] = static_cast<T>(0) - y[i];
}
vexp_->Compute(n, y, y);
for (int i = 0; i < n; ++i) {
vexp_->Compute(this->num_, y, y);
for (int i = 0; i < this->num_; ++i) {
y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
}
}
......@@ -142,8 +143,8 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
#define INTRI8_FLOAT(isa) \
template <> \
void VSigmoidKernelImpl<float, isa, kEQ8>::Compute( \
const int n, const float* x, float* y) const { \
void VSigmoidKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
const { \
__m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \
__m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \
__m256 tmp = _mm256_loadu_ps(x); \
......@@ -153,8 +154,8 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
#define INTRI16_FLOAT(isa) \
template <> \
void VSigmoidKernelImpl<float, isa, kEQ16>::Compute( \
const int n, const float* x, float* y) const { \
void VSigmoidKernelImpl<float, isa, kEQ16>::Compute(const float* x, \
float* y) const { \
__m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \
__m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \
__m256 tmp0 = _mm256_loadu_ps(x); \
......@@ -167,8 +168,16 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
#define INTRI_GT8LT16_FLOAT(isa) \
template <> \
void VSigmoidKernelImpl<float, isa, kGT8LT16>::Compute( \
const int n, const float* x, float* y) const { \
VSigmoidKernelImpl<float, isa, kGT8LT16>::VSigmoidKernelImpl(int d) \
: VSigmoidKernel<float>() { \
this->num_ = d; \
this->end_ = AVX_FLOAT_BLOCK; \
this->rest_ = d - this->end_; \
vexp_ = KernelPool::Instance().template Get<VExpKernel<float>>(d); \
} \
template <> \
void VSigmoidKernelImpl<float, isa, kGT8LT16>::Compute(const float* x, \
float* y) const { \
__m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \
__m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \
__m256 tmp = _mm256_loadu_ps(x); \
......@@ -176,38 +185,43 @@ class VSigmoidKernelImpl : public VSigmoidKernel<T> {
_mm256_storeu_ps(y, tmp); \
const float min_ = SIGMOID_THRESHOLD_MIN; \
const float max_ = SIGMOID_THRESHOLD_MAX; \
for (int i = AVX_FLOAT_BLOCK; i < n; ++i) { \
for (int i = this->end_; i < this->num_; ++i) { \
y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]); \
y[i] = 0.f - y[i]; \
} \
vexp_->Compute(n - AVX_FLOAT_BLOCK, y + AVX_FLOAT_BLOCK, \
y + AVX_FLOAT_BLOCK); \
for (int i = AVX_FLOAT_BLOCK; i < n; ++i) { \
vexp_->Compute(this->rest_, y + this->end_, y + this->end_); \
for (int i = this->end_; i < this->num_; ++i) { \
y[i] = 1.f / (1.f + y[i]); \
} \
}
#define INTRI_GT16_FLOAT(isa) \
template <> \
void VSigmoidKernelImpl<float, isa, kGT16>::Compute( \
const int n, const float* x, float* y) const { \
VSigmoidKernelImpl<float, isa, kGT16>::VSigmoidKernelImpl(int d) \
: VSigmoidKernel<float>() { \
this->num_ = d; \
this->rest_ = d % AVX_FLOAT_BLOCK; \
this->end_ = d - this->rest_; \
vexp_ = KernelPool::Instance().template Get<VExpKernel<float>>(d); \
} \
template <> \
void VSigmoidKernelImpl<float, isa, kGT16>::Compute(const float* x, \
float* y) const { \
__m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \
__m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \
const int rest = n % AVX_FLOAT_BLOCK; \
const int end = n - rest; \
for (int i = 0; i < end; i += AVX_FLOAT_BLOCK) { \
for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \
__m256 tmp = _mm256_loadu_ps(x + i); \
INTRI_SIGMOID(tmp, min, max); \
_mm256_storeu_ps(y + i, tmp); \
} \
const float min_ = SIGMOID_THRESHOLD_MIN; \
const float max_ = SIGMOID_THRESHOLD_MAX; \
for (int i = end; i < n; ++i) { \
for (int i = this->end_; i < this->num_; ++i) { \
y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]); \
y[i] = 0.f - y[i]; \
} \
vexp_->Compute(rest, y + end, y + end); \
for (int i = end; i < n; ++i) { \
vexp_->Compute(this->rest_, y + this->end_, y + this->end_); \
for (int i = this->end_; i < this->num_; ++i) { \
y[i] = 1.f / (1.f + y[i]); \
} \
}
......@@ -249,15 +263,16 @@ template <typename T, jit::cpu_isa_t isa, jit_block>
class VTanhKernelImpl : public VTanhKernel<T> {
public:
explicit VTanhKernelImpl(int d) : VTanhKernel<T>() {
this->num_ = d;
vscal_ = KernelPool::Instance().template Get<VScalKernel<T>>(d);
vsigmoid_ = KernelPool::Instance().template Get<VSigmoidKernel<T>>(d);
vaddbias_ = KernelPool::Instance().template Get<VAddBiasKernel<T>>(d);
}
void Compute(const int n, const T* x, T* y) const override {
vscal_->Compute(n, static_cast<T>(2), x, y);
vsigmoid_->Compute(n, y, y);
vscal_->Compute(n, static_cast<T>(2), y);
vaddbias_->Compute(n, static_cast<T>(-1), y, y);
void Compute(const T* x, T* y) const override {
vscal_->Compute(this->num_, static_cast<T>(2), x, y);
vsigmoid_->Compute(y, y);
vscal_->Compute(this->num_, static_cast<T>(2), y);
vaddbias_->Compute(this->num_, static_cast<T>(-1), y, y);
}
private:
......@@ -276,8 +291,8 @@ class VTanhKernelImpl : public VTanhKernel<T> {
#define INTRI8_FLOAT(isa) \
template <> \
void VTanhKernelImpl<float, isa, kEQ8>::Compute(const int n, const float* x, \
float* y) const { \
void VTanhKernelImpl<float, isa, kEQ8>::Compute(const float* x, float* y) \
const { \
__m256 tmp = _mm256_loadu_ps(x); \
INTRI_VTANH(tmp); \
_mm256_storeu_ps(y, tmp); \
......@@ -285,8 +300,8 @@ class VTanhKernelImpl : public VTanhKernel<T> {
#define INTRI16_FLOAT(isa) \
template <> \
void VTanhKernelImpl<float, isa, kEQ16>::Compute( \
const int n, const float* x, float* y) const { \
void VTanhKernelImpl<float, isa, kEQ16>::Compute(const float* x, float* y) \
const { \
__m256 tmp0 = _mm256_loadu_ps(x); \
__m256 tmp1 = _mm256_loadu_ps(x + 8); \
INTRI_VTANH(tmp0); \
......@@ -297,37 +312,60 @@ class VTanhKernelImpl : public VTanhKernel<T> {
#define INTRI_GT8LT16_FLOAT(isa) \
template <> \
void VTanhKernelImpl<float, isa, kGT8LT16>::Compute( \
const int n, const float* x, float* y) const { \
VTanhKernelImpl<float, isa, kGT8LT16>::VTanhKernelImpl(int d) \
: VTanhKernel<float>() { \
this->num_ = d; \
this->end_ = AVX_FLOAT_BLOCK; \
this->rest_ = d - this->end_; \
vscal_ = \
KernelPool::Instance().template Get<VScalKernel<float>>(this->rest_); \
vsigmoid_ = KernelPool::Instance().template Get<VSigmoidKernel<float>>( \
this->rest_); \
vaddbias_ = KernelPool::Instance().template Get<VAddBiasKernel<float>>( \
this->rest_); \
} \
template <> \
void VTanhKernelImpl<float, isa, kGT8LT16>::Compute(const float* x, \
float* y) const { \
__m256 tmp = _mm256_loadu_ps(x); \
INTRI_VTANH(tmp); \
_mm256_storeu_ps(y, tmp); \
x += AVX_FLOAT_BLOCK; \
y += AVX_FLOAT_BLOCK; \
const int rest = n - AVX_FLOAT_BLOCK; \
vscal_->Compute(rest, 2.f, x, y); \
vsigmoid_->Compute(rest, y, y); \
vscal_->Compute(rest, 2.f, y); \
vaddbias_->Compute(rest, -1.f, y, y); \
vscal_->Compute(this->rest_, 2.f, x, y); \
vsigmoid_->Compute(y, y); \
vscal_->Compute(this->rest_, 2.f, y); \
vaddbias_->Compute(this->rest_, -1.f, y, y); \
}
#define INTRI_GT16_FLOAT(isa) \
template <> \
void VTanhKernelImpl<float, isa, kGT16>::Compute( \
const int n, const float* x, float* y) const { \
const int rest = n % AVX_FLOAT_BLOCK; \
const int end = n - rest; \
for (int i = 0; i < end; i += AVX_FLOAT_BLOCK) { \
VTanhKernelImpl<float, isa, kGT16>::VTanhKernelImpl(int d) \
: VTanhKernel<float>() { \
this->num_ = d; \
this->rest_ = d % AVX_FLOAT_BLOCK; \
this->end_ = d - this->rest_; \
vscal_ = \
KernelPool::Instance().template Get<VScalKernel<float>>(this->rest_); \
vsigmoid_ = KernelPool::Instance().template Get<VSigmoidKernel<float>>( \
this->rest_); \
vaddbias_ = KernelPool::Instance().template Get<VAddBiasKernel<float>>( \
this->rest_); \
} \
template <> \
void VTanhKernelImpl<float, isa, kGT16>::Compute(const float* x, float* y) \
const { \
for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \
__m256 tmp = _mm256_loadu_ps(x + i); \
INTRI_VTANH(tmp); \
_mm256_storeu_ps(y + i, tmp); \
} \
x += end; \
y += end; \
vscal_->Compute(rest, 2.f, x, y); \
vsigmoid_->Compute(rest, y, y); \
vscal_->Compute(rest, 2.f, y); \
vaddbias_->Compute(rest, -1.f, y, y); \
x += this->end_; \
y += this->end_; \
vscal_->Compute(this->rest_, 2.f, x, y); \
vsigmoid_->Compute(y, y); \
vscal_->Compute(this->rest_, 2.f, y); \
vaddbias_->Compute(this->rest_, -1.f, y, y); \
}
#ifdef __AVX__
......
paddle/fluid/operators/math/jit_kernel_test.cc
浏览文件 @ 2513b2cc
......@@ -195,7 +195,7 @@ TEST(JitKernel, vsigmoid) {
auto trefe = GetCurrentUS();
auto ttgts = GetCurrentUS();
for (int i = 0; i < repeat; ++i) {
ker->Compute(d, x_data, ztgt_data);
ker->Compute(x_data, ztgt_data);
}
auto ttgte = GetCurrentUS();
......@@ -227,7 +227,7 @@ void vtanh_better(
vaddbias,
const int n, const float* x, float* y) {
vscal->Compute(n, 2.f, x, y);
vsigmoid->Compute(n, y, y);
vsigmoid->Compute(y, y);
vscal->Compute(n, 2.f, y);
vaddbias->Compute(n, -1.f, y, y);
}
......@@ -261,7 +261,7 @@ TEST(JitKernel, vtanh) {
auto trefe = GetCurrentUS();
auto ttgts = GetCurrentUS();
for (int i = 0; i < repeat; ++i) {
ker->Compute(d, x_data, ztgt_data);
ker->Compute(x_data, ztgt_data);
}
auto ttgte = GetCurrentUS();
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或