Skip to content

P
PaddleDetection

s920243400 / PaddleDetection
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleDetection

通知 2
Star 0
Fork 0
P
PaddleDetection
提交 dee5d35c 编写于 作者: T tensor-tang
浏览文件
操作

refine vmul

上级 92031968
隐藏空白更改
内联 并排
Showing with 100 addition and 72 deletion
paddle/fluid/operators/math/cpu_vec.h
浏览文件 @ dee5d35c
......@@ -125,10 +125,8 @@ inline void vec_scal<float, platform::jit::avx2>(const int n, const float a,
}
template <>
inline void vec_scal<float, platform::jit::avx512_common>(const int n,
const float a,
const float* x,
float* y) {
inline void vec_scal<float, platform::jit::avx512f>(const int n, const float a,
const float* x, float* y) {
// TODO(TJ): enable me
vec_scal<float, platform::jit::avx2>(n, a, x, y);
}
......@@ -181,10 +179,10 @@ inline void vec_bias_sub<float, platform::jit::avx2>(const int n, const float a,
}
template <>
inline void vec_bias_sub<float, platform::jit::avx512_common>(const int n,
const float a,
const float* x,
float* y) {
inline void vec_bias_sub<float, platform::jit::avx512f>(const int n,
const float a,
const float* x,
float* y) {
// TODO(TJ): enable me
vec_bias_sub<float, platform::jit::avx2>(n, a, x, y);
}
......@@ -242,7 +240,7 @@ inline void vec_cross<float, platform::jit::avx2>(const int n, const float* x,
}
template <>
inline void vec_cross<float, platform::jit::avx512_common>(
inline void vec_cross<float, platform::jit::avx512f>(
const int n, const float* x, const float* y, const float* z, float* out) {
// TODO(TJ): enable me
vec_cross<float, platform::jit::avx>(n, x, y, z, out);
......@@ -296,10 +294,10 @@ inline void vec_add_bias<float, platform::jit::avx2>(const int n, const float a,
}
template <>
inline void vec_add_bias<float, platform::jit::avx512_common>(const int n,
const float a,
const float* x,
float* y) {
inline void vec_add_bias<float, platform::jit::avx512f>(const int n,
const float a,
const float* x,
float* y) {
// TODO(TJ): enable me
vec_add_bias<float, platform::jit::avx2>(n, a, x, y);
}
......@@ -390,9 +388,9 @@ inline void vec_sigmoid<float, platform::jit::avx2>(const int n, const float* x,
}
template <>
inline void vec_sigmoid<float, platform::jit::avx512_common>(const int n,
const float* x,
float* y) {
inline void vec_sigmoid<float, platform::jit::avx512f>(const int n,
const float* x,
float* y) {
// TODO(TJ): enable me
vec_sigmoid<float, platform::jit::avx2>(n, x, y);
}
......@@ -454,9 +452,8 @@ inline void vec_relu<float, platform::jit::avx2>(const int n, const float* x,
}
template <>
inline void vec_relu<float, platform::jit::avx512_common>(const int n,
const float* x,
float* y) {
inline void vec_relu<float, platform::jit::avx512f>(const int n, const float* x,
float* y) {
// TODO(TJ): enable me
vec_relu<float, platform::jit::avx2>(n, x, y);
}
......
paddle/fluid/operators/math/cpu_vec_test.cc
浏览文件 @ dee5d35c
......@@ -110,7 +110,7 @@ TEST(CpuVecTest, sigmoid) {
TestAndBench<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
TestAndBench<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
TestAndBench<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
TestAndBench<float>(sz, vec_sigmoid<float, jit::avx512_common>,
TestAndBench<float>(sz, vec_sigmoid<float, jit::avx512f>,
ref_sigmoid<float>);
}
TestAndBench<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
......@@ -123,8 +123,7 @@ TEST(CpuVecTest, tanh) {
TestAndBench<float>(sz, vec_tanh<float>, ref_tanh<float>);
TestAndBench<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
TestAndBench<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
TestAndBench<float>(sz, vec_tanh<float, jit::avx512_common>,
ref_tanh<float>);
TestAndBench<float>(sz, vec_tanh<float, jit::avx512f>, ref_tanh<float>);
}
TestAndBench<double>(30, vec_tanh<double>, ref_tanh<double>);
}
......@@ -136,8 +135,7 @@ TEST(CpuVecTest, relu) {
TestAndBench<float>(sz, vec_relu<float>, ref_relu<float>);
TestAndBench<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
TestAndBench<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
TestAndBench<float>(sz, vec_relu<float, jit::avx512_common>,
ref_relu<float>);
TestAndBench<float>(sz, vec_relu<float, jit::avx512f>, ref_relu<float>);
}
TestAndBench<double>(30, vec_relu<double>, ref_relu<double>);
}
......@@ -170,7 +168,7 @@ TEST(CpuVecTest, inplace_sigmoid) {
TestInplace<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
TestInplace<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
TestInplace<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
TestInplace<float>(sz, vec_sigmoid<float, jit::avx512_common>,
TestInplace<float>(sz, vec_sigmoid<float, jit::avx512f>,
ref_sigmoid<float>);
}
TestInplace<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
......@@ -183,8 +181,7 @@ TEST(CpuVecTest, inplace_tanh) {
TestInplace<float>(sz, vec_tanh<float>, ref_tanh<float>);
TestInplace<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
TestInplace<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
TestInplace<float>(sz, vec_tanh<float, jit::avx512_common>,
ref_tanh<float>);
TestInplace<float>(sz, vec_tanh<float, jit::avx512f>, ref_tanh<float>);
}
TestInplace<double>(30, vec_tanh<double>, ref_tanh<double>);
}
......@@ -196,8 +193,7 @@ TEST(CpuVecTest, inplace_relu) {
TestInplace<float>(sz, vec_relu<float>, ref_relu<float>);
TestInplace<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
TestInplace<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
TestInplace<float>(sz, vec_relu<float, jit::avx512_common>,
ref_relu<float>);
TestInplace<float>(sz, vec_relu<float, jit::avx512f>, ref_relu<float>);
}
TestInplace<double>(30, vec_relu<double>, ref_relu<double>);
}
paddle/fluid/operators/math/jit_kernel.cc
浏览文件 @ dee5d35c
......@@ -36,35 +36,38 @@ KernelPool& KernelPool::Instance() {
static KernelPool g_jit_kernels;
return g_jit_kernels;
}
#define SEARCH_BLOCK(src, t, isa) \
if (d < AVX_FLOAT_BLOCK) { \
Compute = src<t, isa, kLT8>; \
} else if (d == AVX_FLOAT_BLOCK) { \
Compute = src<t, isa, kEQ8>; \
} else if (d == AVX512_FLOAT_BLOCK) { \
Compute = src<t, isa, kEQ16>; \
} else { \
Compute = src<t, isa, kGT16>; \
#define SEARCH_BLOCK(src, t, isa) \
if (d < AVX_FLOAT_BLOCK) { \
Compute = src<t, isa, kLT8>; \
} else if (d == AVX_FLOAT_BLOCK) { \
Compute = src<t, isa, kEQ8>; \
} else if (d > AVX_FLOAT_BLOCK && d < AVX512_FLOAT_BLOCK) { \
Compute = src<t, isa, kGT8LT16>; \
} else if (d == AVX512_FLOAT_BLOCK) { \
Compute = src<t, isa, kEQ16>; \
} else { \
Compute = src<t, isa, kGT16>; \
}
#define SEARCH_ISA_BLOCK(src, t) \
if (jit::MayIUse(jit::avx512_common)) { \
SEARCH_BLOCK(src, t, jit::avx512_common); \
} else if (jit::MayIUse(jit::avx2)) { \
SEARCH_BLOCK(src, t, jit::avx2); \
} else if (jit::MayIUse(jit::avx)) { \
SEARCH_BLOCK(src, t, jit::avx); \
} else { \
SEARCH_BLOCK(src, t, jit::isa_any); \
#define SEARCH_ISA_BLOCK(src, t) \
if (jit::MayIUse(jit::avx512f)) { \
SEARCH_BLOCK(src, t, jit::avx512f); \
} else if (jit::MayIUse(jit::avx2)) { \
SEARCH_BLOCK(src, t, jit::avx2); \
} else if (jit::MayIUse(jit::avx)) { \
SEARCH_BLOCK(src, t, jit::avx); \
} else { \
SEARCH_BLOCK(src, t, jit::isa_any); \
}
#define FOR_EACH_BLOCK(macro_, isa) \
macro_(isa, kLT8) macro_(isa, kEQ8) macro_(isa, kEQ16) macro_(isa, kGT16)
// do not include lt8, eq8, eq16
#define FOR_EACH_COMMON_BLOCK(macro_, isa) \
macro_(isa, kGT8LT16) macro_(isa, kGT16)
#define FOR_EACH_ISA_BLOCK(macro_) \
FOR_EACH_BLOCK(macro_, jit::avx512_common) \
FOR_EACH_BLOCK(macro_, jit::avx2) \
FOR_EACH_BLOCK(macro_, jit::avx) \
#define FOR_EACH_ISA_COMMON_BLOCK(macro_) \
FOR_EACH_BLOCK(macro_, jit::avx512f) \
FOR_EACH_BLOCK(macro_, jit::avx2) \
FOR_EACH_BLOCK(macro_, jit::avx) \
FOR_EACH_BLOCK(macro_, jit::any)
#define VMUL_ANY \
......@@ -78,24 +81,56 @@ static void VMulCompute(const int n, const T* x, const T* y, T* z) {
}
#ifdef PADDLE_USE_MKLML
#define DEFINE_VMUL_COMPUTE_FLOAT(isa, block) \
template <> \
static void VMulCompute<float, isa, block>(const int n, const float* x, \
const float* y, float* z) { \
platform::dynload::vsMul(n, x, y, z); \
#define DEFINE_VMUL_COMPUTE_FLOAT(isa, block) \
template <> \
void VMulCompute<float, isa, block>(const int n, const float* x, \
const float* y, float* z) { \
platform::dynload::vsMul(n, x, y, z); \
}
#define DEFINE_VMUL_COMPUTE_DOUBLE(isa, block) \
template <> \
static void VMulCompute<double, isa, block>(const int n, const double* x, \
const double* y, float* z) { \
platform::dynload::vdMul(n, x, y, z); \
#define DEFINE_VMUL_COMPUTE_DOUBLE(isa, block) \
template <> \
void VMulCompute<double, isa, block>(const int n, const double* x, \
const double* y, float* z) { \
platform::dynload::vdMul(n, x, y, z); \
}
FOR_EACH_ISA_BLOCK(DEFINE_VMUL_COMPUTE_FLOAT)
FOR_EACH_ISA_BLOCK(DEFINE_VMUL_COMPUTE_DOUBLE)
// TODO(TJ): add EQ8
FOR_EACH_ISA_COMMON_BLOCK(DEFINE_VMUL_COMPUTE_FLOAT)
FOR_EACH_ISA_COMMON_BLOCK(DEFINE_VMUL_COMPUTE_DOUBLE)
DEFINE_VMUL_COMPUTE_FLOAT(jit::avx, kLT8)
DEFINE_VMUL_COMPUTE_FLOAT(jit::avx, kEQ16)
#endif
// mkl > avx > for, ">" means better
#ifdef PADDLE_USE_MKLML
DEFINE_VMUL_COMPUTE_FLOAT(jit::avx, kEQ8)
#elif defined __AVX__
template <>
void VMulCompute<float, jit::avx, kEQ8>(const int n, const float* x,
const float* y, float* z) {
__m256 tmpx, tmpy;
tmpx = _mm256_loadu_ps(x);
tmpy = _mm256_loadu_ps(y);
tmpx = _mm256_mul_ps(tmpx, tmpy);
_mm256_storeu_ps(z, tmpx);
}
#endif
// avx2 > mkl > for
#ifdef __AVX2__
template <>
void VMulCompute<float, jit::avx2, kEQ8>(const int n, const float* x,
const float* y, float* z) {
__m256 tmpx, tmpy;
tmpx = _mm256_loadu_ps(x);
tmpy = _mm256_loadu_ps(y);
tmpx = _mm256_mul_ps(tmpx, tmpy);
_mm256_storeu_ps(z, tmpx);
}
#elif defined PADDLE_USE_MKLML
DEFINE_VMUL_COMPUTE_FLOAT(jit::avx2, kEQ8)
#endif
// TODO(TJ): test and complete avx512
#undef DEFINE_VMUL_COMPUTE_FLOAT
#undef DEFINE_VMUL_COMPUTE_DOUBLE
......@@ -142,8 +177,8 @@ LSTMKernel<float>::LSTMKernel(int d, const std::string& act_gate_str,
: Kernel(), d_(d) {
d2_ = d * 2;
d3_ = d * 3;
if (platform::jit::MayIUse(platform::jit::avx512_common)) {
math::VecActivations<float, platform::jit::avx512_common> act_functor;
if (platform::jit::MayIUse(platform::jit::avx512f)) {
math::VecActivations<float, platform::jit::avx512f> act_functor;
act_gate_ = act_functor(act_gate_str);
act_cell_ = act_functor(act_cell_str);
act_cand_ = act_functor(act_cand_str);
......
paddle/fluid/operators/math/jit_kernel.h
浏览文件 @ dee5d35c
......@@ -36,7 +36,7 @@ namespace jitkernel {
#define AVX512_FLOAT_BLOCK 16
#define AVX512_DOUBLE_BLOCK 8
typedef enum { kLT8, kEQ8, kEQ16, kGT16 } jit_block;
typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block;
class Kernel {
public:
......
paddle/fluid/platform/cpu_info.cc
浏览文件 @ dee5d35c
......@@ -128,7 +128,7 @@ bool MayIUse(const cpu_isa_t cpu_isa) {
return cpu.has(Cpu::tAVX);
case avx2:
return cpu.has(Cpu::tAVX2);
case avx512_common:
case avx512f:
return cpu.has(Cpu::tAVX512F);
case avx512_core:
return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) &&
......
paddle/fluid/platform/cpu_info.h
浏览文件 @ dee5d35c
......@@ -43,7 +43,7 @@ typedef enum {
sse42,
avx,
avx2,
avx512_common,
avx512f,
avx512_core,
avx512_core_vnni,
avx512_mic,
......
paddle/fluid/platform/init.cc
浏览文件 @ dee5d35c
......@@ -116,7 +116,7 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
platform::SetNumThreads(FLAGS_paddle_num_threads);
#endif
if (platform::jit::MayIUse(platform::jit::avx512_common)) {
if (platform::jit::MayIUse(platform::jit::avx512f)) {
#ifndef __AVX512F__
LOG(WARNING) << "AVX512F is available, Please re-compile on local machine";
#endif
......
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