Support Div and FloorDiv functor in elementwise system (#33053)

paddle/fluid/operators/elementwise/elementwise_div_op.cu

@@ -12,8 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 #include "paddle/fluid/operators/elementwise/elementwise_div_op.h"
-#include "paddle/fluid/operators/elementwise/elementwise_op_function.cu.h"
-#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h"
 #include "paddle/fluid/platform/complex.h"
 #include "paddle/fluid/platform/float16.h"
 
@@ -23,38 +22,37 @@ namespace plat = paddle::platform;
 namespace paddle {
 namespace operators {
 
+template <typename T, typename Enable = void>
+struct CudaDivFunctor {
+  inline HOSTDEVICE T operator()(const T* args) const {
+    return args[0] / args[1];
+  }
+};
+
 template <typename T>
-struct SameDimsElemwiseDiv<platform::CUDADeviceContext, T> {
-  void operator()(const framework::ExecutionContext& ctx,
-                  const framework::Tensor* x, const framework::Tensor* y,
-                  framework::Tensor* z) {
-    DivRangeFunctor<T> functor(x->data<T>(), y->data<T>(), z->data<T>());
-    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
-    platform::ForRange<platform::CUDADeviceContext> for_range(dev_ctx,
-                                                              x->numel());
-    for_range(functor);
+struct CudaDivFunctor<T,
+                      typename std::enable_if_t<std::is_integral<T>::value>> {
+  inline HOSTDEVICE T operator()(const T* args) const {
+    PADDLE_ENFORCE(args[1] != 0,
+                   "Invalid Argument Error: Integer division by zero "
+                   "encountered in divide. Please check the input value.");
+    return args[0] / args[1];
   }
 };
 
-template <>
-struct SameDimsElemwiseDiv<platform::CUDADeviceContext, platform::float16> {
-  void operator()(const framework::ExecutionContext& ctx,
-                  const framework::Tensor* x, const framework::Tensor* y,
-                  framework::Tensor* z) {
-    auto size = x->numel();
-    dim3 grid_size = dim3(((size + 7) / 8 + PADDLE_CUDA_THREAD_SIZE - 1) /
-                              PADDLE_CUDA_THREAD_SIZE,
-                          1);
-    dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
-    const half* x2 =
-        reinterpret_cast<const half*>(x->data<platform::float16>());
-    const half* y2 =
-        reinterpret_cast<const half*>(y->data<platform::float16>());
-    half* z2 = reinterpret_cast<half*>(z->data<platform::float16>());
-    SameDimsElemwiseDivCUDAKernel<<<
-        grid_size, block_size, 0,
-        ctx.template device_context<platform::CUDADeviceContext>().stream()>>>(
-        x2, y2, z2, size);
+template <typename T>
+class ElementwiseDivKernel<platform::CUDADeviceContext, T>
+    : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    std::vector<const framework::Tensor*> ins;
+    std::vector<framework::Tensor*> outs;
+    const auto& cuda_ctx =
+        ctx.template device_context<platform::CUDADeviceContext>();
+
+    int axis = PackTensorsIntoVector<T>(ctx, &ins, &outs);
+    LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
+        cuda_ctx, ins, &outs, axis, CudaDivFunctor<T>());
   }
 };
 

paddle/fluid/operators/elementwise/elementwise_floordiv_op.cu

@@ -12,11 +12,43 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 #include "paddle/fluid/operators/elementwise/elementwise_floordiv_op.h"
-#include "paddle/fluid/platform/float16.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h"
 
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
+namespace paddle {
+namespace operators {
+
+template <typename T>
+struct CudaFloorDivFunctor {
+  inline HOSTDEVICE T operator()(const T argv[]) const {
+    PADDLE_ENFORCE(argv[1] != 0,
+                   "InvalidArgument: divide by zero "
+                   "encountered in floor-divide ops, please check.\n");
+    return static_cast<T>(std::trunc(argv[0] / argv[1]));
+  }
+};
+
+template <typename T>
+class ElementwiseFloorDivKernel<platform::CUDADeviceContext, T>
+    : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    std::vector<const framework::Tensor*> ins;
+    std::vector<framework::Tensor*> outs;
+    const auto& cuda_ctx =
+        ctx.template device_context<platform::CUDADeviceContext>();
+
+    int axis = PackTensorsIntoVector<T>(ctx, &ins, &outs);
+    LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
+        cuda_ctx, ins, &outs, axis, CudaFloorDivFunctor<T>());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
 REGISTER_OP_CUDA_KERNEL(
     elementwise_floordiv,
     ops::ElementwiseFloorDivKernel<plat::CUDADeviceContext, int>,

paddle/fluid/operators/elementwise/elementwise_floordiv_op.h

@@ -16,7 +16,6 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
-#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/blas.h"
 
 namespace paddle {

paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h

@@ -14,7 +14,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/fluid/framework/tensor.h"
-#include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/platform/cuda_device_function.h"
 #include "paddle/fluid/platform/fast_divmod.h"
 
 #ifdef __HIPCC__
@@ -28,19 +28,62 @@ namespace operators {
 
 enum ElementwiseType { kUnary = 1, kBinary = 2 };
 
+/*
+* According to NVIDIA, if number of threads per block is 64/128/256/512,
+* cuda performs better. And number of blocks should be greater (at least
+* 2x~4x) than number of SMs. Hence, SM count is took into account within
+* this function to determine the right number of threads per block.
+*/
+inline int GetThreadsConfig(const platform::CUDADeviceContext &ctx,
+                            int64_t numel, int vec_size) {
+  int threads = ELEMENTWISE_BLOCK_SIZE;
+  int sm_count = ctx.GetSMCount();
+  int active_threads_num = numel / vec_size;
+  if (active_threads_num / (sm_count << 1) < ELEMENTWISE_BLOCK_SIZE) {
+    // Round up threads number into an exponential multiple of 2, while number
+    // of acitve blocks is about twice of SM, to acquire better performance.
+    threads = platform::RoundToPowerOfTwo(active_threads_num / (sm_count << 1));
+  } else if (active_threads_num / (sm_count << 2) < ELEMENTWISE_BLOCK_SIZE) {
+    // Round up threads number into an exponential multiple of 2, while number
+    // of acitve blocks is about 4 times of SM, to acquire better performance.
+    threads = platform::RoundToPowerOfTwo(active_threads_num / (sm_count << 2));
+  }
+  // Number of threads per block shall be larger than 64.
+  return std::max(64, threads);
+}
+
+/*
+* Only the address of input data is the multiplier of 1,2,4, vectorized load
+* with corresponding multiplier-value is possible. Moreover, the maximum length
+* of vectorized load is 128 bits once. Hence, valid length of vectorized load
+* shall be determined under both former constraints.
+*/
 template <typename T>
 int GetVectorizedSizeImpl(const T *pointer) {
+  constexpr int max_load_bits = 128;
+  int valid_vec_size = max_load_bits / CHAR_BIT / sizeof(T);
   uint64_t address = reinterpret_cast<uint64_t>(pointer);
+  constexpr int vec8 =
+      std::alignment_of<CudaAlignedVector<T, 8>>::value;  // NOLINT
   constexpr int vec4 =
       std::alignment_of<CudaAlignedVector<T, 4>>::value;  // NOLINT
   constexpr int vec2 =
       std::alignment_of<CudaAlignedVector<T, 2>>::value;  // NOLINT
-  if (address % vec4 == 0) {
-    return 4;
+  if (address % vec8 == 0) {
+    /*
+    * Currently, decide to deal with no more than 4 data once while adopting
+    * vectorization load/store, if performance test shows that dealing with
+    * 8 data once in vectorization load/store does get optimized, return code
+    * below can be changed into " return std::min(8, valid_vec_size); " .
+    */
+    return std::min(4, valid_vec_size);
+  } else if (address % vec4 == 0) {
+    return std::min(4, valid_vec_size);
   } else if (address % vec2 == 0) {
-    return 2;
+    return std::min(2, valid_vec_size);
+  } else {
+    return 1;
   }
-  return 1;
 }
 
 template <typename InT, typename OutT>
@@ -96,7 +139,7 @@ struct ElementwiseDataWrapper {
 
 template <ElementwiseType ET, int VecSize, typename InT, typename OutT,
           typename Functor>
-__device__ void VectorizedKernelImpl(
+__device__ inline void VectorizedKernelImpl(
     ElementwiseDataWrapper<ET, VecSize, InT, OutT> data, Functor func,
     int tid) {
   using InVecType = CudaAlignedVector<InT, VecSize>;
@@ -104,34 +147,30 @@ __device__ void VectorizedKernelImpl(
   InVecType ins_vec[ET];
   OutVecType out_vec;
   InT *ins_ptr[ET];
-  OutT *out_ptr;
+  InT ins[ET];
 #pragma unroll
   for (int i = 0; i < ET; ++i) {
     ins_ptr[i] = reinterpret_cast<InT *>(&(ins_vec[i]));
   }
-  out_ptr = reinterpret_cast<OutT *>(&out_vec);
-
   // load
   data.load_vector(ins_vec, tid);
 
 // compute
 #pragma unroll
   for (int i = 0; i < VecSize; ++i) {
-    InT ins[ET];
 #pragma unroll
     for (int j = 0; j < ET; ++j) {
       ins[j] = ins_ptr[j][i];
     }
-    out_ptr[i] = func(ins);
+    out_vec.val[i] = func(ins);
   }
-
   // store
   data.store_vector(out_vec, tid);
 }
 
 template <ElementwiseType ET, int VecSize, typename InT, typename OutT,
           typename Functor>
-__device__ void ScalarKernelImpl(
+__device__ inline void ScalarKernelImpl(
     ElementwiseDataWrapper<ET, VecSize, InT, OutT> data, Functor func,
     int start, int remain) {
   InT ins[ET];
@@ -182,7 +221,7 @@ void LaunchSameDimsElementwiseCudaKernel(
   // calculate the max vec_size for all ins and outs
   auto size = ins[0]->numel();
   int vec_size = GetVectorizedSize<InT, OutT>(ins, *outs);
-  int block_size = ELEMENTWISE_BLOCK_SIZE;
+  int block_size = GetThreadsConfig(ctx, size, vec_size);
   int grid_size =
       ((size + vec_size - 1) / vec_size + block_size - 1) / block_size;
   const InT *in0 = ins[0]->data<InT>();