optimize check_finite_and_unscale_op by fused kernel, test=develop (#31954)

paddle/fluid/operators/amp/check_finite_and_unscale_op.cu

@@ -26,18 +26,48 @@ __global__ void InverseAndMemset(const T* s, T* o, bool* found_inf) {
 }
 
 template <typename T, typename MT>
-__global__ void CheckFiniteAndUnscale(const T* in, const MT* scale, int num,
-                                      bool* found_inf, T* out) {
-  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
-
-  if (idx < num) {
-    MT val = static_cast<MT>(in[idx]) * (*scale);
+__global__ void CheckFiniteAndUnscale(const T** xs, const MT* scale,
+                                      int64_t size, int64_t* starts,
+                                      bool* found_inf, T** outs) {
+  const int64_t tid = threadIdx.x + blockIdx.x * blockDim.x;
+
+  // copy starts array from global memory to shared memory
+  extern __shared__ int64_t s_starts[];
+  for (int i = threadIdx.x; i <= size; i += blockDim.x) {
+    s_starts[i] = starts[i];
+  }
+  __syncthreads();
+
+  const int64_t num = s_starts[size];
+  int pre_xs_index = 0;
+  bool t_found_inf = false;
+  const MT t_scale = *scale;
+  for (int64_t idx = tid; idx < num; idx += gridDim.x * blockDim.x) {
+    // get the xs's index of thread
+    int xs_index = pre_xs_index;
+    while (idx < s_starts[xs_index]) xs_index++;
+    // avoid some tensor's numel is zero
+    while (idx >= s_starts[xs_index]) xs_index++;
+    pre_xs_index = xs_index - 1;
+
+    // get in data and out data
+    const T* in = xs[pre_xs_index];
+    T* out = outs[pre_xs_index];
+    int64_t in_idx = idx - s_starts[pre_xs_index];
+
+    // Unscale
+    MT val = static_cast<MT>(in[in_idx]) * t_scale;
     T narrow_val = static_cast<T>(val);
-    out[idx] = narrow_val;
+    out[in_idx] = narrow_val;
+
+    // CheckFinite
     if (!isfinite(narrow_val)) {
-      *found_inf = true;
+      t_found_inf = true;
     }
   }
+  if (t_found_inf) {
+    *found_inf = true;
+  }
 }
 
 template <typename T>
@@ -63,21 +93,54 @@ class CheckFiniteAndUnscaleGpuKernel : public framework::OpKernel<T> {
     InverseAndMemset<MPDType><<<1, 1, 0, dev_ctx.stream()>>>(
         scale_data, inverse_scale_v, found_inf_data);
 
-    for (size_t i = 0; i < xs.size(); ++i) {
-      const auto* x = xs[i];
-      auto* out = outs[i];
-      const T* x_data = x->data<T>();
-      T* out_data = out->mutable_data<T>(dev_ctx.GetPlace());
+    size_t xs_size = xs.size();
+    // calculate each tensor's start index and copy to device
+    auto h_starts_tensor =
+        memory::Alloc(platform::CPUPlace(), (xs_size + 1) * sizeof(int64_t));
+    int64_t* h_starts = reinterpret_cast<int64_t*>(h_starts_tensor->ptr());
+
+    auto d_starts_tensor =
+        memory::Alloc(dev_ctx, (xs_size + 1) * sizeof(int64_t));
+    int64_t* d_starts = reinterpret_cast<int64_t*>(d_starts_tensor->ptr());
+
+    h_starts[0] = 0;
+    for (int i = 1; i <= xs_size; i++) {
+      // the start index value of each tensor is
+      // the sum of previous tensor's size
+      h_starts[i] = h_starts[i - 1] + xs[i - 1]->numel();
+    }
+    int64_t total_num = h_starts[xs_size];
+    memory::Copy(BOOST_GET_CONST(platform::CUDAPlace, dev_ctx.GetPlace()),
+                 d_starts, platform::CPUPlace(), h_starts,
+                 (xs_size + 1) * sizeof(int64_t), dev_ctx.stream());
+
+    // copy each tensor's data address to device
+    auto h_mem = memory::Alloc(platform::CPUPlace(), 2 * xs_size * sizeof(T*));
+    const T** h_xs = reinterpret_cast<const T**>(h_mem->ptr());
+    T** h_outs = reinterpret_cast<T**>(h_mem->ptr()) + xs_size;
+
+    auto d_mem = memory::Alloc(dev_ctx, 2 * xs_size * sizeof(T*));
+    const T** d_xs = reinterpret_cast<const T**>(d_mem->ptr());
+    T** d_outs = reinterpret_cast<T**>(d_mem->ptr()) + xs_size;
+
+    for (size_t i = 0; i < xs_size; ++i) {
+      h_xs[i] = xs[i]->data<T>();
+      h_outs[i] = outs[i]->mutable_data<T>(dev_ctx.GetPlace());
+    }
+    memory::Copy(BOOST_GET_CONST(platform::CUDAPlace, dev_ctx.GetPlace()), d_xs,
+                 platform::CPUPlace(), h_xs, 2 * xs_size * sizeof(T*),
+                 dev_ctx.stream());
 
-      int num = x->numel();
+    // Launch Kernel
     int block = 1024;
-      int grid = (num + block - 1) / block;
+    int block_num = block * 20;  // each thread deal with 20 number
+    int grid = (total_num + block_num - 1) / block_num;
     VLOG(3) << "launch kernel";
-      CheckFiniteAndUnscale<T, MPDType><<<grid, block, 0, dev_ctx.stream()>>>(
-          x_data, inverse_scale_v, num, found_inf_data, out_data);
+    CheckFiniteAndUnscale<T, MPDType><<<
+        grid, block, (xs_size + 1) * sizeof(int64_t), dev_ctx.stream()>>>(
+        d_xs, inverse_scale_v, xs_size, d_starts, found_inf_data, d_outs);
     VLOG(3) << "finish kernel";
   }
-  }
 };
 }  // namespace operators
 }  // namespace paddle