Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 14949521 编写于 作者: L limingshu 提交者： GitHub
浏览文件
操作

Binary functor envoking of elementwise broadcast (#32928)

上级 6f8de31d
显示空白变更内容
内联 并排
Showing with 193 addition and 170 deletion
paddle/fluid/operators/abs_op.cu
浏览文件 @ 14949521
......@@ -52,8 +52,9 @@ class AbsKernel<platform::CUDADeviceContext, T>
std::vector<const framework::Tensor*> ins = {x};
std::vector<framework::Tensor*> outs = {out};
auto functor = CudaAbsFunctor<T>();
LaunchElementwiseCudaKernel<ElementwiseType::kUnary, T, math::Real<T>>(
dev_ctx, ins, &outs, functor);
LaunchSameDimsElementwiseCudaKernel<ElementwiseType::kUnary, T,
math::Real<T>>(dev_ctx, ins, &outs,
functor);
}
};
......
paddle/fluid/operators/activation_op.cu
浏览文件 @ 14949521
......@@ -1316,8 +1316,8 @@ class ActivationCudaKernel
for (auto& attr : attrs) {
*attr.second = ctx.Attr<float>(attr.first);
}
LaunchElementwiseCudaKernel<ElementwiseType::kUnary, T, T>(dev_ctx, ins,
&outs, functor);
LaunchSameDimsElementwiseCudaKernel<ElementwiseType::kUnary, T, T>(
dev_ctx, ins, &outs, functor);
}
};
......@@ -1346,16 +1346,16 @@ class ActivationGradCudaKernel
if (static_cast<int>(Functor::FwdDeps()) == static_cast<int>(kDepOut)) {
// Only need forward output Out
ins.push_back(out);
LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
LaunchSameDimsElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
dev_ctx, ins, &outs, functor);
} else if (static_cast<int>(Functor::FwdDeps()) ==
static_cast<int>(kDepX)) {
// Only need forward input X
ins.push_back(x);
LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
LaunchSameDimsElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
dev_ctx, ins, &outs, functor);
} else {
LaunchElementwiseCudaKernel<ElementwiseType::kUnary, T, T>(
LaunchSameDimsElementwiseCudaKernel<ElementwiseType::kUnary, T, T>(
dev_ctx, ins, &outs, functor);
}
}
......
paddle/fluid/operators/elementwise/elementwise_add_op.cc
浏览文件 @ 14949521
......@@ -69,15 +69,6 @@ struct SameDimsElemwiseAdd<
}
};
template <typename T>
struct BroadcastElemwiseAdd<platform::CPUDeviceContext, T> {
void operator()(const framework::ExecutionContext &ctx,
const framework::Tensor *x, const framework::Tensor *y,
framework::Tensor *z) {
default_elementwise_add<platform::CPUDeviceContext, T>(ctx, x, y, z);
}
};
class ElementwiseAddOpMaker : public ElementwiseOpMaker {
protected:
std::string GetName() const override { return "Add"; }
......
paddle/fluid/operators/elementwise/elementwise_add_op.cu
浏览文件 @ 14949521
......@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h"
#include "paddle/fluid/platform/complex128.h"
#include "paddle/fluid/platform/complex64.h"
#include "paddle/fluid/platform/float16.h"
......@@ -40,29 +39,24 @@ struct CudaAddFunctor {
};
template <typename T>
struct SameDimsElemwiseAdd<platform::CUDADeviceContext, T> {
void operator()(const framework::ExecutionContext& ctx,
const framework::Tensor* x, const framework::Tensor* y,
framework::Tensor* z) {
class ElementwiseAddKernel<platform::CUDADeviceContext, T>
: public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<framework::LoDTensor>("X");
auto* y = ctx.Input<framework::LoDTensor>("Y");
auto* z = ctx.Output<framework::LoDTensor>("Out");
z->mutable_data<T>(ctx.GetPlace());
int axis = ctx.Attr<int>("axis");
axis = axis == -1 ? std::abs(x->dims().size() - y->dims().size()) : axis;
std::vector<const framework::Tensor*> ins = {x, y};
std::vector<framework::Tensor*> outs = {z};
LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
ctx.template device_context<platform::CUDADeviceContext>(), ins, &outs,
CudaAddFunctor<T>());
}
};
const auto& cuda_ctx =
ctx.template device_context<platform::CUDADeviceContext>();
template <typename T>
struct BroadcastElemwiseAdd<platform::CUDADeviceContext, T> {
void operator()(const framework::ExecutionContext& ctx,
const framework::Tensor* x, const framework::Tensor* y,
framework::Tensor* out) {
std::vector<const framework::Tensor*> ins = {x, y};
int axis = ctx.Attr<int>("axis");
axis = axis == -1 ? std::abs(x->dims().size() - y->dims().size()) : axis;
LaunchBroadcastElementwiseCudaKernel<ElementwiseType::kBinary, T>(
ctx.template device_context<platform::CUDADeviceContext>(), ins, out,
CudaAddFunctor<T>(), axis);
LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
cuda_ctx, ins, &outs, axis, CudaAddFunctor<T>());
}
};
......
paddle/fluid/operators/elementwise/elementwise_add_op.h
浏览文件 @ 14949521
......@@ -26,7 +26,7 @@ limitations under the License. */
#include <cuda.h>
#include <cuda_fp16.h>
#include "cub/cub.cuh"
#include "paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h"
#endif
#ifdef __HIPCC__
#include <hip/hip_fp16.h>
......@@ -40,9 +40,10 @@ namespace paddle {
namespace operators {
template <typename DeviceContext, typename T>
void default_elementwise_add(const framework::ExecutionContext &ctx,
void LaunchBroadcastElementwiseCpuKernel(const framework::ExecutionContext &ctx,
const framework::Tensor *x,
const framework::Tensor *y, framework::Tensor *z) {
const framework::Tensor *y,
framework::Tensor *z) {
int axis = ctx.Attr<int>("axis");
auto x_dims = x->dims();
auto y_dims = y->dims();
......@@ -62,13 +63,6 @@ struct SameDimsElemwiseAdd {
framework::Tensor *z);
};
template <typename DeviceContext, typename T, class Enable = void>
struct BroadcastElemwiseAdd {
void operator()(const framework::ExecutionContext &ctx,
const framework::Tensor *x, const framework::Tensor *y,
framework::Tensor *z);
};
template <typename DeviceContext, typename T>
class ElementwiseAddKernel : public framework::OpKernel<T> {
public:
......@@ -77,13 +71,13 @@ class ElementwiseAddKernel : public framework::OpKernel<T> {
auto *y = ctx.Input<framework::LoDTensor>("Y");
auto *z = ctx.Output<framework::LoDTensor>("Out");
z->mutable_data<T>(ctx.GetPlace());
auto dims_equal = x->dims() == y->dims();
if (dims_equal) {
SameDimsElemwiseAdd<DeviceContext, T> same_dims_add;
same_dims_add(ctx, x, y, z);
if (x->dims() == y->dims()) {
SameDimsElemwiseAdd<platform::CPUDeviceContext, T>
LaunchElementwiseCpuKernel;
LaunchElementwiseCpuKernel(ctx, x, y, z);
} else {
BroadcastElemwiseAdd<DeviceContext, T> broadcast_add;
broadcast_add(ctx, x, y, z);
LaunchBroadcastElementwiseCpuKernel<platform::CPUDeviceContext, T>(ctx, x,
y, z);
}
}
};
......@@ -469,8 +463,8 @@ class ElementwiseAddDoubleGradKernel : public framework::OpKernel<T> {
GetDoubleGradSafeTensor<DeviceContext, T>(ctx, y, ddy, &ddy_safe);
ddout->mutable_data<T>(ctx.GetPlace());
default_elementwise_add<DeviceContext, T>(ctx, &ddx_safe, &ddy_safe,
ddout);
LaunchBroadcastElementwiseCpuKernel<DeviceContext, T>(ctx, &ddx_safe,
&ddy_safe, ddout);
}
}
};
......
paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h
浏览文件 @ 14949521
......@@ -14,7 +14,7 @@
#pragma once
#include "paddle/fluid/operators/elementwise/elementwise_op_broadcast_impl.cu.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h"
namespace paddle {
namespace operators {
......@@ -28,7 +28,8 @@ struct DimensionsTransform {
std::vector<DimVector> in_dims;
private:
// 1. To compensate the lackage of input_tensors` dimension;
// To compensate the lackage of input_tensors` dimension with input variable
// 'axis'
void InputDimensionsExtend(int N, int axis) {
for (auto &in_dim : in_dims) {
int64_t in_idx = 0;
......@@ -70,7 +71,7 @@ struct DimensionsTransform {
}
template <typename MergeFunctor>
__inline__ void DimensionsReorganise(MergeFunctor merge_func, int N) {
__inline__ void MergeDimensions(MergeFunctor merge_func, int N) {
auto VectorReorganise = [](DimVector *vec, int l_idx, int m_idx) {
(*vec)[m_idx - 1] =
std::accumulate(vec->begin() + l_idx, vec->begin() + m_idx, 1,
......@@ -139,7 +140,7 @@ struct DimensionsTransform {
// To Merge the dimensions of input_tensors while the consequtive
// equal-dimensions appears.
MergeFunctor merge_ptr = merge_sequential_dims;
DimensionsReorganise<MergeFunctor>(merge_ptr, N);
MergeDimensions<MergeFunctor>(merge_ptr, N);
int min_idx = 0;
int min_val = std::accumulate(in_dims[0].begin(), in_dims[0].end(), 1,
......@@ -155,12 +156,12 @@ struct DimensionsTransform {
// To Merge the dimension of input_tensors while the consequtive
// 1-value-dimensions appears.
merge_ptr = merge_sequential_one_dims;
DimensionsReorganise<MergeFunctor>(merge_ptr, N);
MergeDimensions<MergeFunctor>(merge_ptr, N);
std::swap(in_dims[min_idx], in_dims[0]);
}
};
struct CalculateInputStrides {
struct StridesCalculation {
std::vector<std::vector<uint32_t>> strides;
std::vector<FastDivMod> divmoders;
......@@ -181,8 +182,8 @@ struct CalculateInputStrides {
}
public:
explicit CalculateInputStrides(
const int64_t &dim_size, const std::vector<std::vector<int64_t>> &in_dims,
explicit StridesCalculation(const int64_t &dim_size,
const std::vector<std::vector<int64_t>> &in_dims,
const std::vector<int64_t> &out_dims) {
const auto N = in_dims.size();
divmoders.resize(dim_size);
......@@ -195,34 +196,40 @@ struct CalculateInputStrides {
}
};
template <typename T, ElementwiseType ET, int VecSize, int kDims>
template <typename T, typename Functor, ElementwiseType ET, int VecSize,
int kDims>
struct BroadcastArgsWarpper {
using DimsVec = CudaAlignedVector<T, VecSize>;
using VecType = CudaAlignedVector<T, VecSize>;
T *out_data;
VecType *vec_out_data;
const T *__restrict__ in_data[ET];
uint32_t strides[ET][framework::DDim::kMaxRank];
const VecType *__restrict__ vec_in_data[ET];
bool no_broadcast[ET];
FastDivMod divmoders[kDims];
uint32_t scalar_offset;
uint32_t strides[ET][framework::DDim::kMaxRank];
uint32_t scalar_cal_offset;
Functor func;
HOSTDEVICE BroadcastArgsWarpper(
const std::vector<const framework::Tensor *> &ins,
const CalculateInputStrides &offset_calculator, framework::Tensor *out,
int scalar_offset)
: scalar_offset(scalar_offset) {
const std::vector<const framework::Tensor *> &ins, framework::Tensor *out,
int scalar_cal_offset, Functor func,
const StridesCalculation &offset_calculator)
: scalar_cal_offset(scalar_cal_offset), func(func) {
for (int j = 0; j < ET; ++j) {
in_data[j] = ins[j]->data<T>();
vec_in_data[j] = reinterpret_cast<const VecType *>(in_data[j]);
no_broadcast[j] = ins[j]->dims() == out->dims() ? true : false;
memcpy(strides[j], offset_calculator.strides[j].data(),
kDims * sizeof(uint32_t));
}
out_data = out->data<T>();
vec_out_data = reinterpret_cast<VecType *>(out_data);
memcpy(divmoders, offset_calculator.divmoders.data(),
kDims * sizeof(FastDivMod));
}
__device__ __forceinline__ uint32_t GetDivmodOffset(int idx, int in_idx) {
__device__ __forceinline__ uint32_t GetOffsetByDivmod(int idx, int in_idx) {
uint32_t offset = 0;
#pragma unroll(kDims)
......@@ -234,120 +241,127 @@ struct BroadcastArgsWarpper {
return offset;
}
__device__ __forceinline__ void CommonVector(DimsVec args[], int tid,
int idx) {
const DimsVec *__restrict__ vec_data =
reinterpret_cast<const DimsVec *__restrict__>(in_data[idx]);
args[idx] = vec_data[tid];
__device__ __forceinline__ void LoadVectorizedDataCommon(VecType *vector_args,
int tid, int idx) {
*vector_args = vec_in_data[idx][tid];
}
__device__ __forceinline__ void DivmodVector(DimsVec args[], int tid,
int idx) {
__device__ __forceinline__ void LoadVectorizedDataByDivmod(T *scalar_args,
int tid, int idx) {
int index = tid * VecSize;
#pragma unroll(VecSize)
for (int i = 0; i < VecSize; ++i) {
uint32_t offset = GetDivmodOffset(index + i, idx);
args[idx].val[i] = in_data[idx][offset];
uint32_t offset = GetOffsetByDivmod(index + i, idx);
scalar_args[i] = in_data[idx][offset];
}
}
__device__ __forceinline__ void CommonScalar(T args[], int tid, int idx) {
args[idx] = in_data[idx][tid + scalar_offset];
__device__ __forceinline__ void LoadScalarizedDataCommon(T args[], int tid,
int idx) {
args[idx] = in_data[idx][tid + scalar_cal_offset];
}
__device__ __forceinline__ void DivmodScalar(T args[], int tid, int idx) {
auto offset = GetDivmodOffset(tid + scalar_offset, idx);
__device__ __forceinline__ void LoadScalarizedDataByDivmod(T args[], int tid,
int idx) {
auto offset = GetOffsetByDivmod(tid + scalar_cal_offset, idx);
args[idx] = in_data[idx][offset];
}
__device__ __forceinline__ void LoadVector(DimsVec args[], int tid) {
__device__ __forceinline__ void LoadVectorizedData(T (*args)[VecSize],
int tid) {
#pragma unroll(ET)
for (int j = 0; j < ET; ++j) {
if (no_broadcast[j]) {
CommonVector(args, tid, j);
VecType *vector_args = reinterpret_cast<VecType *>(args[j]);
LoadVectorizedDataCommon(vector_args, tid, j);
} else {
DivmodVector(args, tid, j);
LoadVectorizedDataByDivmod(args[j], tid, j);
}
}
}
__device__ __forceinline__ void LoadScalar(T args[], int tid) {
__device__ __forceinline__ void LoadScalarizedData(T args[], int tid) {
#pragma unroll(ET)
for (int j = 0; j < ET; ++j) {
if (no_broadcast[j]) {
CommonScalar(args, tid, j);
LoadScalarizedDataCommon(args, tid, j);
} else {
DivmodScalar(args, tid, j);
LoadScalarizedDataByDivmod(args, tid, j);
}
}
}
__device__ __forceinline__ void StoreVector(DimsVec args[], int tid) {
DimsVec *vec_out = reinterpret_cast<DimsVec *>(out_data);
vec_out[tid] = args[0];
__device__ __forceinline__ void StoreVectorizedData(T (*args)[VecSize],
int tid) {
VecType *args_out = reinterpret_cast<VecType *>(args[0]);
vec_out_data[tid] = *args_out;
}
__device__ __forceinline__ void StoreScalar(T args[], int tid) {
out_data[scalar_offset + tid] = args[0];
__device__ __forceinline__ void StoreScalarizedData(T args[], int tid) {
out_data[scalar_cal_offset + tid] = args[0];
}
};
template <typename T, typename BroadcastArgsWarpper, ElementwiseType ET>
__device__ inline void ScalarizedBroadcastKernelImpl(
BroadcastArgsWarpper data_transfer, int tid) {
BroadcastArgsWarpper broadcast_warpper, int tid) {
T args[ET];
data_transfer.LoadScalar(args, tid);
broadcast_warpper.LoadScalarizedData(args, tid);
#pragma unroll(ET)
for (int j = 1; j < ET; ++j) {
args[0] += args[j];
args[0] = broadcast_warpper.func(args);
}
data_transfer.StoreScalar(args, tid);
broadcast_warpper.StoreScalarizedData(args, tid);
}
template <typename T, typename BroadcastArgsWarpper, ElementwiseType ET,
int VecSize>
__device__ inline void VectorizedBroadcastKernelImpl(
BroadcastArgsWarpper data_transfer, int tid) {
using VecT = CudaAlignedVector<T, VecSize>;
VecT args[ET];
data_transfer.LoadVector(args, tid);
BroadcastArgsWarpper broadcast_warpper, int tid) {
T ins[ET];
T args[ET][VecSize];
broadcast_warpper.LoadVectorizedData(args, tid);
#pragma unroll(ET)
for (int j = 1; j < ET; ++j) {
#pragma unroll(VecSize)
for (int i = 0; i < VecSize; ++i) {
args[0].val[i] += args[j].val[i];
#pragma unroll(ET)
for (int j = 0; j < ET; ++j) {
ins[j] = args[j][i];
}
args[0][i] = broadcast_warpper.func(ins);
}
data_transfer.StoreVector(args, tid);
broadcast_warpper.StoreVectorizedData(args, tid);
}
template <typename T, typename BroadcastArgsWarpper, ElementwiseType ET,
int VecSize>
__global__ void ElementwiseBroadcastKernel(BroadcastArgsWarpper data_transfer,
int main_tid, int tail_tid) {
__global__ void ElementwiseBroadcastKernel(
BroadcastArgsWarpper broadcast_warpper, int main_tid, int tail_tid) {
int tid = threadIdx.x + blockIdx.x * blockDim.x;
// Aimming at vectorized calculation of major data whose length is max
// multipler of VecSize.
// Vectorized calculation of major data whose length is the max multipler of
// VecSize,
// eg: Calcualting the front 1024-length data in total 1027 data once VecSize
// is 4.
if (tid < main_tid) {
VectorizedBroadcastKernelImpl<T, BroadcastArgsWarpper, ET, VecSize>(
data_transfer, tid);
broadcast_warpper, tid);
}
// Aimming at scalar calculation of rest data whose lenght cannot fulfill
// VecSize.
// Scalarzed calculation of rest data whose lenght cannot fulfill VecSize.
// eg: Calcualting the rest 3-length data in total 1027 data once VecSize is
// 4.
if (tid < tail_tid) {
ScalarizedBroadcastKernelImpl<T, BroadcastArgsWarpper, ET>(data_transfer,
tid);
ScalarizedBroadcastKernelImpl<T, BroadcastArgsWarpper, ET>(
broadcast_warpper, tid);
}
}
template <typename T, ElementwiseType ET, int VecSize = 1>
template <typename T, ElementwiseType ET, int VecSize, typename Functor>
void LaunchBroadcastKernelForDifferentDimSize(
const platform::CUDADeviceContext &ctx,
const std::vector<const framework::Tensor *> &ins, framework::Tensor *out,
int axis) {
int axis, Functor func) {
int numel = out->numel();
const int threads = 256;
int blocks = ((numel + VecSize - 1) / VecSize + threads - 1) / threads;
......@@ -357,72 +371,72 @@ void LaunchBroadcastKernelForDifferentDimSize(
auto stream = ctx.stream();
const auto merge_dims = DimensionsTransform(ins, out->dims(), axis);
const auto offset_calculator = CalculateInputStrides(
const auto offset_calculator = StridesCalculation(
merge_dims.dim_size, merge_dims.in_dims, merge_dims.out_dims);
switch (merge_dims.dim_size) {
case 1: {
auto data_transfer = BroadcastArgsWarpper<T, ET, VecSize, 1>(
ins, offset_calculator, out, vec_len);
ElementwiseBroadcastKernel<T, decltype(data_transfer), ET,
auto broadcast_warpper = BroadcastArgsWarpper<T, Functor, ET, VecSize, 1>(
ins, out, vec_len, func, offset_calculator);
ElementwiseBroadcastKernel<T, decltype(broadcast_warpper), ET,
VecSize><<<blocks, threads, 0, stream>>>(
data_transfer, main_tid, tail_tid);
broadcast_warpper, main_tid, tail_tid);
break;
}
case 2: {
auto data_transfer = BroadcastArgsWarpper<T, ET, VecSize, 2>(
ins, offset_calculator, out, vec_len);
ElementwiseBroadcastKernel<T, decltype(data_transfer), ET,
auto broadcast_warpper = BroadcastArgsWarpper<T, Functor, ET, VecSize, 2>(
ins, out, vec_len, func, offset_calculator);
ElementwiseBroadcastKernel<T, decltype(broadcast_warpper), ET,
VecSize><<<blocks, threads, 0, stream>>>(
data_transfer, main_tid, tail_tid);
broadcast_warpper, main_tid, tail_tid);
break;
}
case 3: {
auto data_transfer = BroadcastArgsWarpper<T, ET, VecSize, 3>(
ins, offset_calculator, out, vec_len);
ElementwiseBroadcastKernel<T, decltype(data_transfer), ET,
auto broadcast_warpper = BroadcastArgsWarpper<T, Functor, ET, VecSize, 3>(
ins, out, vec_len, func, offset_calculator);
ElementwiseBroadcastKernel<T, decltype(broadcast_warpper), ET,
VecSize><<<blocks, threads, 0, stream>>>(
data_transfer, main_tid, tail_tid);
broadcast_warpper, main_tid, tail_tid);
break;
}
case 4: {
auto data_transfer = BroadcastArgsWarpper<T, ET, VecSize, 4>(
ins, offset_calculator, out, vec_len);
ElementwiseBroadcastKernel<T, decltype(data_transfer), ET,
auto broadcast_warpper = BroadcastArgsWarpper<T, Functor, ET, VecSize, 4>(
ins, out, vec_len, func, offset_calculator);
ElementwiseBroadcastKernel<T, decltype(broadcast_warpper), ET,
VecSize><<<blocks, threads, 0, stream>>>(
data_transfer, main_tid, tail_tid);
broadcast_warpper, main_tid, tail_tid);
break;
}
case 5: {
auto data_transfer = BroadcastArgsWarpper<T, ET, VecSize, 5>(
ins, offset_calculator, out, vec_len);
ElementwiseBroadcastKernel<T, decltype(data_transfer), ET,
auto broadcast_warpper = BroadcastArgsWarpper<T, Functor, ET, VecSize, 5>(
ins, out, vec_len, func, offset_calculator);
ElementwiseBroadcastKernel<T, decltype(broadcast_warpper), ET,
VecSize><<<blocks, threads, 0, stream>>>(
data_transfer, main_tid, tail_tid);
broadcast_warpper, main_tid, tail_tid);
break;
}
case 6: {
auto data_transfer = BroadcastArgsWarpper<T, ET, VecSize, 6>(
ins, offset_calculator, out, vec_len);
ElementwiseBroadcastKernel<T, decltype(data_transfer), ET,
auto broadcast_warpper = BroadcastArgsWarpper<T, Functor, ET, VecSize, 6>(
ins, out, vec_len, func, offset_calculator);
ElementwiseBroadcastKernel<T, decltype(broadcast_warpper), ET,
VecSize><<<blocks, threads, 0, stream>>>(
data_transfer, main_tid, tail_tid);
broadcast_warpper, main_tid, tail_tid);
break;
}
case 7: {
auto data_transfer = BroadcastArgsWarpper<T, ET, VecSize, 7>(
ins, offset_calculator, out, vec_len);
ElementwiseBroadcastKernel<T, decltype(data_transfer), ET,
auto broadcast_warpper = BroadcastArgsWarpper<T, Functor, ET, VecSize, 7>(
ins, out, vec_len, func, offset_calculator);
ElementwiseBroadcastKernel<T, decltype(broadcast_warpper), ET,
VecSize><<<blocks, threads, 0, stream>>>(
data_transfer, main_tid, tail_tid);
broadcast_warpper, main_tid, tail_tid);
break;
}
case 8: {
auto data_transfer = BroadcastArgsWarpper<T, ET, VecSize, 8>(
ins, offset_calculator, out, vec_len);
ElementwiseBroadcastKernel<T, decltype(data_transfer), ET,
auto broadcast_warpper = BroadcastArgsWarpper<T, Functor, ET, VecSize, 8>(
ins, out, vec_len, func, offset_calculator);
ElementwiseBroadcastKernel<T, decltype(broadcast_warpper), ET,
VecSize><<<blocks, threads, 0, stream>>>(
data_transfer, main_tid, tail_tid);
broadcast_warpper, main_tid, tail_tid);
break;
}
default: {
......@@ -437,9 +451,11 @@ void LaunchBroadcastKernelForDifferentDimSize(
template <ElementwiseType ET, typename T, typename Functor>
void LaunchBroadcastElementwiseCudaKernel(
const platform::CUDADeviceContext &ctx,
const std::vector<const framework::Tensor *> &ins, framework::Tensor *out,
Functor func, int axis) {
const std::vector<const framework::Tensor *> &ins,
std::vector<framework::Tensor *> *outs, int axis, Functor func) {
static_assert(ET == (ElementwiseType)2, "Only Support binary calculation.");
int in_vec_size = 4;
framework::Tensor *out = (*outs)[0];
for (auto *in : ins) {
auto temp_size = GetVectorizedSizeImpl<T>(in->data<T>());
in_vec_size = in->dims() == out->dims() ? std::min(temp_size, in_vec_size)
......@@ -450,19 +466,46 @@ void LaunchBroadcastElementwiseCudaKernel(
switch (vec_size) {
case 4: {
LaunchBroadcastKernelForDifferentDimSize<T, ET, 4>(ctx, ins, out, axis);
LaunchBroadcastKernelForDifferentDimSize<T, ET, 4>(ctx, ins, out, axis,
func);
break;
}
case 2: {
LaunchBroadcastKernelForDifferentDimSize<T, ET, 2>(ctx, ins, out, axis);
LaunchBroadcastKernelForDifferentDimSize<T, ET, 2>(ctx, ins, out, axis,
func);
break;
}
case 1: {
LaunchBroadcastKernelForDifferentDimSize<T, ET, 1>(ctx, ins, out, axis,
func);
break;
}
default: {
LaunchBroadcastKernelForDifferentDimSize<T, ET, 1>(ctx, ins, out, axis);
PADDLE_THROW(platform::errors::Unimplemented(
"Unsupported vectorized size: %d !", vec_size));
break;
}
}
}
template <ElementwiseType ET, typename InT, typename OutType, typename Functor>
void LaunchElementwiseCudaKernel(
const platform::CUDADeviceContext &cuda_ctx,
const std::vector<const framework::Tensor *> &ins,
std::vector<framework::Tensor *> *outs, int axis, Functor func) {
bool no_broadcast_flag = true;
for (auto *in : ins) {
no_broadcast_flag = ins[0]->dims() == in->dims();
}
if (no_broadcast_flag) {
LaunchSameDimsElementwiseCudaKernel<ElementwiseType::kBinary, InT, OutType>(
cuda_ctx, ins, outs, func);
} else {
LaunchBroadcastElementwiseCudaKernel<ElementwiseType::kBinary, InT>(
cuda_ctx, ins, outs, axis, func);
}
}
} // namespace operators
} // namespace paddle
paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h
浏览文件 @ 14949521
......@@ -15,8 +15,7 @@ limitations under the License. */
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/float16.h"
#include "paddle/fluid/platform/fast_divmod.h"
#ifdef __HIPCC__
#define ELEMENTWISE_BLOCK_SIZE 256
......@@ -29,11 +28,6 @@ namespace operators {
enum ElementwiseType { kUnary = 1, kBinary = 2 };
template <typename T, int Size>
struct alignas(sizeof(T) * Size) CudaAlignedVector {
T val[Size];
};
template <typename T>
int GetVectorizedSizeImpl(const T *pointer) {
uint64_t address = reinterpret_cast<uint64_t>(pointer);
......@@ -181,7 +175,7 @@ __global__ void ScalarKernel(const InT *__restrict__ in0,
}
template <ElementwiseType ET, typename InT, typename OutT, typename Functor>
void LaunchElementwiseCudaKernel(
void LaunchSameDimsElementwiseCudaKernel(
const platform::CUDADeviceContext &ctx,
const std::vector<const framework::Tensor *> &ins,
std::vector<framework::Tensor *> *outs, Functor func) {
......
paddle/fluid/operators/elementwise/elementwise_op_broadcast_impl.cu.h → paddle/fluid/platform/fast_divmod.h
浏览文件 @ 14949521
......@@ -14,13 +14,19 @@ limitations under the License. */
#pragma once
#include "paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h"
#include <cstdint>
#include "paddle/fluid/platform/hostdevice.h"
#define INT_BITS 32
namespace paddle {
namespace operators {
template <typename T, int Size>
struct alignas(sizeof(T) * Size) CudaAlignedVector {
T val[Size];
};
struct FastDivMod {
// 1st value represents the result of input number divides by recorded divisor
// 2nd value represents the result of input number modulo by recorded divisor
......
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