Support different dtypes of inputs for broadcast for dropout optimization (#52093)

* change judgement for DropoutGradGPUKernelDriver

* add UnrollerWithoutVecSize and after this Loaddata to be refined

* pass unittest

* use same unroller with XPU

* BroadcastWithInt64Index

* BroadcastDataLoader template partial specialization

* fix compile errs in ROCms

* PR comment

paddle/phi/kernels/funcs/dropout_impl.cu.h

@@ -459,41 +459,43 @@ void DropoutGradGPUKernelDriver(const phi::GPUContext& dev_ctx,
     // y = factor * x
     ScaleByDropoutFactor<T, MT>(dev_ctx, grad_y, grad_x, factor);
   } else {
-    phi::DenseTensor broadcasted_mask;
-    if (is_dropout_nd) {
-      broadcasted_mask.Resize(grad_y.dims());
-      dev_ctx.template Alloc<uint8_t>(&broadcasted_mask);
-
-      std::vector<const phi::DenseTensor*> broadcast_ins = {&mask};
-      std::vector<phi::DenseTensor*> broadcast_outs = {&broadcasted_mask};
-      phi::funcs::BroadcastKernel<phi::ElementwiseType::kUnary,
-                                  uint8_t,
-                                  uint8_t>(dev_ctx,
-                                           broadcast_ins,
-                                           &broadcast_outs,
-                                           -1,
-                                           kps::IdentityFunctor<uint8_t>());
-    }
-
-    std::vector<const phi::DenseTensor*> ins = {
-        &grad_y, is_dropout_nd ? &broadcasted_mask : &mask};
-    std::vector<phi::DenseTensor*> outs = {grad_x};
-    if (upscale_in_train) {
-      if (dropout_prob == 1.0f) {
+    if (upscale_in_train && dropout_prob == 1.0f) {
 #ifdef PADDLE_WITH_HIP
-        hipMemset(grad_x->data<T>(), 0, grad_x->numel() * sizeof(T));
+      hipMemset(grad_x->data<T>(), 0, grad_x->numel() * sizeof(T));
 #else
-        cudaMemset(grad_x->data<T>(), 0, grad_x->numel() * sizeof(T));
+      cudaMemset(grad_x->data<T>(), 0, grad_x->numel() * sizeof(T));
 #endif
+    } else {
+      MT factor = upscale_in_train
+                      ? static_cast<MT>(1.0f / (1.0f - dropout_prob))
+                      : static_cast<MT>(1.0f);
+      if (is_dropout_nd) {
+        phi::DenseTensor broadcasted_mask;
+
+        broadcasted_mask.Resize(grad_y.dims());
+        dev_ctx.template Alloc<uint8_t>(&broadcasted_mask);
+
+        std::vector<const phi::DenseTensor*> broadcast_ins = {&mask};
+        std::vector<phi::DenseTensor*> broadcast_outs = {&broadcasted_mask};
+        phi::funcs::BroadcastKernel<phi::ElementwiseType::kUnary,
+                                    uint8_t,
+                                    uint8_t>(dev_ctx,
+                                             broadcast_ins,
+                                             &broadcast_outs,
+                                             -1,
+                                             kps::IdentityFunctor<uint8_t>());
+
+        std::vector<const phi::DenseTensor*> ins = {&grad_y, &broadcasted_mask};
+        std::vector<phi::DenseTensor*> outs = {grad_x};
+        phi::funcs::ElementwiseKernel<T>(
+            dev_ctx, ins, &outs, CudaDropoutGradFunctor<T>(factor));
+
       } else {
-        MT factor = static_cast<MT>(1.0f / (1.0f - dropout_prob));
+        std::vector<const phi::DenseTensor*> ins = {&grad_y, &mask};
+        std::vector<phi::DenseTensor*> outs = {grad_x};
         phi::funcs::ElementwiseKernel<T>(
             dev_ctx, ins, &outs, CudaDropoutGradFunctor<T>(factor));
       }
-    } else {
-      MT factor = static_cast<MT>(1.0f);
-      phi::funcs::ElementwiseKernel<T>(
-          dev_ctx, ins, &outs, CudaDropoutGradFunctor<T>(factor));
     }
   }
 }

paddle/phi/kernels/funcs/elementwise_base.h

@@ -35,7 +35,7 @@ namespace kps = phi::kps;
 
 namespace phi {
 
-enum ElementwiseType { kUnary = 1, kBinary = 2, kTernary = 3, kAny = -1 };
+enum ElementwiseType { kUnary = 1, kBinary = 2, kTernary = 3 };
 /* Packing scalar type T(float, int etc.) into Array<T, NumOuts> type
    for supporting multiple-output feature in elementwise system.*/
 template <class T, int Num>
@@ -508,15 +508,31 @@ struct Unroller<Func, VecSize, End, End> {
   static HOSTDEVICE inline void step(Args &&...args) {}
 };
 
+// static unroller without VecSize for broadcast
+template <template <int Index> typename Func, int End, int Begin = 0>
+struct UnrollerWithoutVecSize {
+  template <typename... Args>
+  static HOSTDEVICE inline void step(Args &&...args) {
+    Func<Begin>::Apply(std::forward<Args>(args)...);
+    UnrollerWithoutVecSize<Func, End, Begin + 1>::step(args...);
+  }
+};
+
+template <template <int Index> typename Func, int End>
+struct UnrollerWithoutVecSize<Func, End, End> {
+  template <typename... Args>
+  static HOSTDEVICE inline void step(Args &&...args) {}
+};
+
 template <int Index, int VecSize>
 struct Loader {
   template <typename Array, typename ArgsT>
-  static __device__ void Apply(const Array &in,
-                               ArgsT *args,
-                               kps::IndexType offset,
-                               int num,
-                               int read_lens,
-                               bool is_boundary) {
+  static __device__ __forceinline__ void Apply(const Array &in,
+                                               ArgsT *args,
+                                               kps::IndexType offset,
+                                               int num,
+                                               int read_lens,
+                                               bool is_boundary) {
     using Type = std::tuple_element_t<Index, ArgsT>;
     kps::Init<Type, ArgsT, Index, VecSize>(
         args, static_cast<Type>(1.0f), read_lens);
@@ -536,7 +552,7 @@ struct Loader {
   }
 };
 
-template <int Index, int VecSize>
+template <int Index>
 struct InputSetter {
   template <typename Array>
   static HOSTDEVICE void Apply(
@@ -545,7 +561,7 @@ struct InputSetter {
   }
 };
 
-template <int Index, int VecSize>
+template <int Index>
 struct VecSizeGetter {
   template <typename ArgsT>
   static HOSTDEVICE void Apply(const std::vector<const DenseTensor *> &ins,
@@ -569,8 +585,7 @@ int GetVectorizedSizeForTensors(const std::vector<const DenseTensor *> &ins,
   int vec_size = 4;
   uint64_t addr = static_cast<uint64_t>(0);
   ArgsT arg;
-  // The Arg VecSize=1 is to match the Unroller template.
-  Unroller<VecSizeGetter, 1, Arity>::step(ins, arg, &vec_size);
+  UnrollerWithoutVecSize<VecSizeGetter, Arity>::step(ins, arg, &vec_size);
   for (auto iter = outs.begin(); iter != outs.end(); ++iter) {
     addr = (addr | reinterpret_cast<uint64_t>((*iter)->data<OutT>()));
   }
@@ -580,73 +595,6 @@ int GetVectorizedSizeForTensors(const std::vector<const DenseTensor *> &ins,
   return vec_size;
 }
 
-template <typename InT,
-          typename OutT,
-          int VecSize,
-          typename Functor,
-          int Arity,
-          bool CallElementwiseAny = false>
-struct ElementwisePrimitiveCaller {
-  __device__ inline void operator()(Functor func,
-                                    InT (*args)[VecSize],
-                                    OutT *result,
-                                    int read_lens);
-};
-
-template <typename InT, typename OutT, int VecSize, typename Functor, int Arity>
-struct ElementwisePrimitiveCaller<InT, OutT, VecSize, Functor, Arity, true> {
-  __device__ inline void operator()(Functor func,
-                                    InT (*args)[VecSize],
-                                    OutT *result,
-                                    int read_lens) {
-    kps::ElementwiseAny<InT, OutT, VecSize, 1, Arity, Functor>(
-        result, args, func);
-  }
-};
-
-template <typename InT, typename OutT, int VecSize, typename Functor>
-struct ElementwisePrimitiveCaller<InT, OutT, VecSize, Functor, 0, false> {
-  __device__ inline void operator()(Functor func,
-                                    InT (*args)[VecSize],
-                                    OutT *result,
-                                    int read_lens) {
-    kps::ElementwiseConstant<InT, OutT, VecSize, 1, Functor>(result, func);
-  }
-};
-
-template <typename InT, typename OutT, int VecSize, typename Functor>
-struct ElementwisePrimitiveCaller<InT, OutT, VecSize, Functor, 1, false> {
-  __device__ inline void operator()(Functor func,
-                                    InT (*args)[VecSize],
-                                    OutT *result,
-                                    int read_lens) {
-    kps::ElementwiseUnary<InT, OutT, VecSize, 1, Functor>(
-        result, args[0], func);
-  }
-};
-
-template <typename InT, typename OutT, int VecSize, typename Functor>
-struct ElementwisePrimitiveCaller<InT, OutT, VecSize, Functor, 2, false> {
-  __device__ inline void operator()(Functor func,
-                                    InT (*args)[VecSize],
-                                    OutT *result,
-                                    int read_lens) {
-    kps::ElementwiseBinary<InT, OutT, VecSize, 1, Functor>(
-        result, args[0], args[1], func, read_lens);
-  }
-};
-
-template <typename InT, typename OutT, int VecSize, typename Functor>
-struct ElementwisePrimitiveCaller<InT, OutT, VecSize, Functor, 3, false> {
-  __device__ inline void operator()(Functor func,
-                                    InT (*args)[VecSize],
-                                    OutT *result,
-                                    int read_lens) {
-    kps::ElementwiseTernary<InT, OutT, VecSize, 1, Functor>(
-        result, args[0], args[1], args[2], func);
-  }
-};
-
 namespace detail {
 template <class F, class Tuple, std::size_t... Index>
 // GCC/Clang need the decltype() return type
@@ -802,7 +750,7 @@ void LaunchElementwiseCudaKernel(const KPDevice &ctx,
   phi::Array<const _ptr_ char *__restrict__, Arity> ins_data;
   phi::Array<_ptr_ OutT *, NumOuts> outs_data;
 
-  Unroller<InputSetter, VecSize, Arity>::step(ins, &ins_data);
+  UnrollerWithoutVecSize<InputSetter, Arity>::step(ins, &ins_data);
   for (int i = 0; i < NumOuts; ++i) {
     outs_data[i] = (_ptr_ OutT *)(ctx.Alloc<OutT>((*outs)[i]));
   }

paddle/phi/kernels/primitive/datamover_primitives.h

@@ -255,6 +255,18 @@ __device__ __forceinline__ void Init(ArgsT* dst, T init_data, int read_lens) {
   }
 }
 
+/**
+ * The difference from the above function is that
+ * it supports different data types of inputs.
+ */
+template <typename T, typename ArgsT, int Index, int NX>
+__device__ __forceinline__ void Init(ArgsT* dst, T init_data) {
+#pragma unroll
+  for (int i = 0; i < NX; i++) {
+    std::get<Index>(dst[i]) = init_data;
+  }
+}
+
 /**
  * @brief Read 1D data from global memory to register. When IsBoundary = true
  * and (NX % 4 == 0 or Nx % 2 == 0), vectorized load data will be used to
@@ -307,6 +319,23 @@ __device__ __forceinline__ void ReadData(T* dst,
   }
 }
 
+/**
+ * @brief Read 1D data from global memory to register.
+ * @template paraments
+ * T: The type of data.
+ * NX: Each thread load NX data from global memory continuously.
+ * NY: Each thread need to load NY rows, only NY = 1 was supported.
+ * IsBoundary: Whether to make an out-of-bounds judgment on access to memory.
+ * When the number of data processed by this block is less than
+ * NX x NY x blockDim.x, boundary judgment is required to avoid memory access
+ * crossing the boundary.
+ *
+ * @param：
+ * dst: The register pointer of the thread, the size is NX * NY.
+ * src: The data pointer of the current block.
+ * size: The current block needs to load size data continuously.
+ */
+
 template <typename T, int NX, int NY, bool IsBoundary = false>
 __device__ __forceinline__ void ReadData(T* dst,
                                          const T* __restrict__ src,
@@ -347,9 +376,8 @@ __device__ __forceinline__ void ReadData(T* dst,
  * T: The type of data.
  * NX: Each thread load NX data from global memory continuously.
  * NY: Each thread need to load NY rows, only NY = 1 was supported.
- * ArgsT: The Type if dst, ArgsT can be std::tuple<T> or std::tuple<Args>
+ * ArgsT: The Type of dst, ArgsT can be std::tuple<T> or std::tuple<Args>
  * Index: The index of data stored in dst.
- * threadIdx.x is used as the thread index. Currently only GPU was supported.
  * IsBoundary: Whether to make an out-of-bounds judgment on access to memory.
  * When the number of data processed by this block is less than
  * NX x NY x blockDim.x, boundary judgment is required to avoid memory access
@@ -369,7 +397,7 @@ template <typename T,
 __device__ __forceinline__ void ReadData(ArgsT* dst,
                                          const T* __restrict__ src,
                                          int num,
-                                         int read_lens) {
+                                         int read_lens = 0) {
   if (IsBoundary) {  // blockDim.x * NX > num
     int thread_offset = threadIdx.x * NX;
 #pragma unroll
@@ -743,7 +771,6 @@ __device__ __forceinline__ void Init(T* dst, T* init_data, int num) {
  * NX: The number of data continuously loaded by each thread.
  * NY: The number of data rows loaded by each thread, only NY = 1 was supported.
  * threadIdx.x is used as the thread index. Currently only GPU was supported.
- * Rank: The shape size of out. eg in[1, 35], out[32, 35] then shape size is 2.
  * IsBoundary: Indicates whether to perform block access storage out-of-bounds
  * judgment. When the number of data processed by the block is less than
  * NX x NY x blockDim.x, boundary judgment is required to avoid memory access
@@ -788,6 +815,67 @@ __device__ __forceinline__ void ReadDataBc(
   }
 }
 
+/**
+ * @brief Read 1D data from global memory to register with broadcast form.
+ * The difference from the above function is that it supports different data
+ * types of inputs.
+ *
+ * @template paraments
+ * T: The type of data stored in the global memory.
+ * NX: The number of data continuously loaded by each thread.
+ * NY: The number of data rows loaded by each thread, only NY = 1 was supported.
+ * ArgsT: The Type of dst, ArgsT can be std::tuple<T> or std::tuple<Args>
+ * Index: The index of data stored in dst.
+ * IsBoundary: Indicates whether to perform block access storage out-of-bounds
+ * judgment. When the number of data processed by the block is less than
+ * NX x NY x blockDim.x, boundary judgment is required to avoid memory access
+ * crossing the boundary.
+ *
+ * @param：
+ * dst: The register pointer of the thread, the size is NX * NY.
+ * src: The original input data pointer of kernel.
+ * block_offset: The data offset of this block, blockDim.x * blockIdx.x * NX;
+ * config: Calculation configuration of broadcast. It is used to calculate the
+ * coordinate mapping relationship between output data and input data.
+ * total_num_output: Total number of original output.
+ */
+
+template <typename T,
+          int NX,
+          int NY,
+          typename ArgsT,
+          int Index,
+          bool IsBoundary = false>
+__device__ __forceinline__ void ReadDataBc(
+    ArgsT* dst,
+    const T* __restrict__ src,
+    uint32_t block_offset,
+    const details::BroadcastConfig& config,
+    int total_num_output,
+    int read_lens = NX) {
+  uint32_t thread_offset = block_offset + threadIdx.x * NX;
+  uint32_t index_src = 0;
+
+#pragma unroll
+  for (uint32_t nx = 0; nx < NX; ++nx) {
+    uint32_t index_output = thread_offset + nx;
+    index_src = 0;
+    if (IsBoundary) {
+      if (index_output >= total_num_output) {
+        break;
+      }
+    }
+#pragma unroll
+    for (int i = 0; i < phi::DDim::kMaxRank; ++i) {
+      if (i >= config.rank) break;
+      auto fast_divmoder = config.divmoders[i].Divmod(index_output);
+      index_output = fast_divmoder.val[0];
+      index_src += fast_divmoder.val[1] * config.strides[i];
+    }
+    std::get<Index>(dst[nx]) = src[index_src];
+  }
+}
+
 /**
  * @brief Initialize register with data index.
  *