[PHI] Optimize GPUScatterNdAdd (#72666)

zhanghonggeng · web-flow · commit 6428d9cf2532 · 2025-05-14T18:57:16.000+08:00
diff --git a/paddle/phi/kernels/funcs/scatter.cu.h b/paddle/phi/kernels/funcs/scatter.cu.h
@@ -96,19 +96,26 @@ __global__ void ScatterCUDAKernel(const T* params,
   }
 }
 
-template <typename T, typename IndexT = int>
+template <typename T, typename IndexT, int VecSize>
 __global__ void ScatterNdCUDAKernel(const T* update,
                                     const IndexT* indices,
                                     T* output,
                                     const Dim<DDim::kMaxRank> output_dims,
                                     size_t remain_size,
                                     size_t slice_size,
                                     size_t end_size) {
-  CUDA_KERNEL_LOOP_TYPE(i, remain_size * slice_size, int64_t) {
-    int64_t indices_i = i / slice_size;
-    int64_t slice_i = i - indices_i * slice_size;  // offset inside the slice
+  int64_t total_size = remain_size * slice_size;
+  int64_t idx = (blockIdx.x * blockDim.x + threadIdx.x) * VecSize;
+  int64_t stride = blockDim.x * gridDim.x * VecSize;
+
+#pragma unroll
+  for (; idx < total_size; idx += stride) {
+    int64_t indices_i = idx / slice_size;
+    int64_t slice_i = idx % slice_size;
     int64_t gather_i = 0;
     int64_t temp = slice_size;
+
+#pragma unroll
     for (int64_t j = end_size - 1; j >= 0; --j) {
       IndexT index_value = indices[indices_i * end_size + j];
       PADDLE_ENFORCE(
@@ -128,8 +135,17 @@ __global__ void ScatterNdCUDAKernel(const T* update,
       gather_i += (index_value * temp);
       temp *= output_dims[j];
     }
+
     int64_t output_i = gather_i + slice_i;
-    phi::CudaAtomicAdd(output + output_i, *(update + i));
+
+    using VecType = kps::details::VectorType<T, VecSize>;
+    const VecType* src = reinterpret_cast<const VecType*>(&update[idx]);
+    VecType* dst = reinterpret_cast<VecType*>(&output[output_i]);
+
+#pragma unroll
+    for (int k = 0; k < VecSize; ++k) {
+      phi::CudaAtomicAdd(&(dst->val[k]), src->val[k]);
+    }
   }
 }
 
@@ -254,6 +270,72 @@ void GPUScatterGradForX(const phi::GPUContext& ctx,
       p_index, p_output, dst_dims[0], index_size, slice_size);
 }
 
+template <typename T, typename IndexT, int VecSize>
+void DispatchScatterNdKernel(
+    const phi::GPUContext& ctx,
+    const T* p_update,
+    const IndexT* p_index,
+    T* p_output,
+    const Dim<DDim::kMaxRank>& g_output_dims,
+    int64_t remain_numel,
+    int64_t slice_size,
+    int64_t end_size,
+    int vec_size,
+    const phi::backends::gpu::GpuLaunchConfig& config) {
+  if (vec_size == VecSize) {
+    auto stream = ctx.stream();
+    ScatterNdCUDAKernel<T, IndexT, VecSize>
+        <<<config.block_per_grid, config.thread_per_block, 0, stream>>>(
+            p_update,
+            p_index,
+            p_output,
+            g_output_dims,
+            remain_numel,
+            slice_size,
+            end_size);
+  } else {
+    DispatchScatterNdKernel<T, IndexT, VecSize / 2>(ctx,
+                                                    p_update,
+                                                    p_index,
+                                                    p_output,
+                                                    g_output_dims,
+                                                    remain_numel,
+                                                    slice_size,
+                                                    end_size,
+                                                    vec_size,
+                                                    config);
+  }
+}
+
+template <typename T, typename IndexT>
+void DispatchScatterNdKernel(
+    const phi::GPUContext& ctx,
+    const T* p_update,
+    const IndexT* p_index,
+    T* p_output,
+    const Dim<DDim::kMaxRank>& g_output_dims,
+    int64_t remain_numel,
+    int64_t slice_size,
+    int64_t end_size,
+    int vec_size,
+    const phi::backends::gpu::GpuLaunchConfig& config) {
+  if (vec_size == 1) {
+    auto stream = ctx.stream();
+    ScatterNdCUDAKernel<T, IndexT, 1>
+        <<<config.block_per_grid, config.thread_per_block, 0, stream>>>(
+            p_update,
+            p_index,
+            p_output,
+            g_output_dims,
+            remain_numel,
+            slice_size,
+            end_size);
+  } else {
+    PADDLE_THROW(common::errors::Unimplemented(
+        "Unsupported vectorized size: %d", vec_size));
+  }
+}
+
 template <typename T, typename IndexT = int>
 void GPUScatterNdAdd(const phi::GPUContext& ctx,
                      const DenseTensor& update,
@@ -286,19 +368,27 @@ void GPUScatterNdAdd(const phi::GPUContext& ctx,
     g_output_dims[i] = output_dims[i];
   }
 
-  int block = 512;
-  int64_t n = slice_size * remain_numel;
-  dim3 grid = dim3((n + block - 1) / block);
-  phi::backends::gpu::LimitGridDim(ctx, &grid);
+  int vec_size = 4;
+  vec_size = std::min(phi::GetVectorizedSize(p_update), vec_size);
+  vec_size = std::min(phi::GetVectorizedSize(p_output), vec_size);
+  while (vec_size > 1 && slice_size % vec_size != 0) {
+    vec_size /= 2;
+  }
 
-  ScatterNdCUDAKernel<T, IndexT>
-      <<<grid, block, 0, ctx.stream()>>>(p_update,
-                                         p_index,
-                                         p_output,
-                                         g_output_dims,
-                                         remain_numel,
-                                         slice_size,
-                                         end_size);
+  constexpr int loop_count = 4;
+  auto config = phi::backends::gpu::GetGpuLaunchConfig1D(
+      ctx, remain_numel * slice_size, vec_size * loop_count);
+
+  DispatchScatterNdKernel<T, IndexT, 4>(ctx,
+                                        p_update,
+                                        p_index,
+                                        p_output,
+                                        g_output_dims,
+                                        remain_numel,
+                                        slice_size,
+                                        end_size,
+                                        vec_size,
+                                        config);
 }
 
 }  // namespace funcs
